Part of the Qwen3.8-27B mixed-precision EXL3 (measured) collection.

Four EXL3/Trellis quants of Qwen/Qwen3.8-27B, each optimizing a different constraint. Every quality claim is measured against a BF16 teacher via KL-divergence — no unverified accuracy assertions.

Model	Download	Resident	v5 mean KLD	Context (32 GB)	TL;DR
K5K6-hydrated	21.61 GB	20.31 GiB	0.002760	~180k	Best fidelity, smallest download, fastest cold start
K5K6	30.60 GB	20.32 GiB	0.003210	~180k	Attention width is a launch-time knob (K6/K5/K4)
K5K6-context	20.70 GB	18.41 GiB	0.003509	262,144	Native 262k context, int8 embeddings, hardware-qualified on RTX 5090
K4	28.31 GB	17.89 GiB	0.010604	262,144	Smallest footprint, native context, no overlay needed

All four are VLM (image-text-to-text) and require the Gilded Gnosis vLLM runtime — they do not load in stock vLLM, SGLang, or transformers.

Qwen3.8-27B EXL3 K5/K6 — lower divergence than official FP8 at 71 % of its memory

Requires a custom runtime. This checkpoint does not load in upstream vLLM, SGLang, TensorRT-LLM, llama.cpp, transformers, or stock exllamav3. It needs the Gilded Gnosis vLLM fork (public image below), explicit --quantization exl3, an exact ignore list, and — for CUDA graphs — an as-yet-unmerged patch (PR #314). Treat it as an experimental, runtime-specific research artifact.

Mixed-precision EXL3 quant of Qwen/Qwen3.8-27B @ 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.

Headline evidence: on a 10,480,640-position held-out suite (5,120 contexts, 842 source clusters, calibration-overlapping documents excluded before selection), this build's mean KL divergence from BF16 is 0.003210 [0.002982, 0.003480] against official FP8's 0.005294, with 97.52 % top-1 agreement, and it diverges less on 5,105 of 5,120 paired contexts — while holding 8.2 GiB less weight. Its hydrated sibling is better still, by 0.000450 on 4,922 of 5,120 contexts. Details and limits.

quantization_manifest.json and build-receipt.json in this repo are authoritative for composition. SHA256SUMS covers the immutable payload only — the four shards, the index, the configs, the tokenizer and preprocessor files: 17 files, 30,597,223,337 bytes. Documentation that changes with card edits (README.md, .gitattributes, assets/**) is listed separately in DOCS-SHA256SUMS, because a hash map that claims to describe a build must not move when prose does. An earlier revision mixed the two, listed a crc32.txt that does not exist, and carried stale hashes for two files. The legacy config.json → quantization_config block keeps a single bits/codebook pair for loader compatibility and cannot describe this mixed checkpoint.

A rebuild of this recipe is a sibling, not this checkpoint. The published bytes are the artifact: the recipe fixes composition, widths and byte budget exactly, and it does not fix the weights. That was measured on the hydrated sibling's recipe, which comes out of the same converter. A fresh conversion returned 13 of 16 pinned payload files identical — every config, the tokenizer, the index and both quantization descriptors — with byte-identical shard headers, the same tensor names, dtypes, shapes and offsets, and the same per-role byte totals and assigned widths, while 399 of the 409 quantized modules (97.6 %) differed inside their .trellis payloads, at 41-92 % of the bytes each (mean 82 %). No scale, norm, embedding, vision or BF16 companion tensor moved. The converter is nondeterministic and that was measured, not inferred: two runs of one conversion, minutes apart, agreed on every width and global scale and disagreed on the converter's own proxy_err (receipts/converter-determinism.json). So a rebuild is a different valid artifact of the same recipe, not a broken one — and every fidelity number on this card, which measures the published bytes a downloader receives, is unaffected. And "rebuild this and you get these numbers" is no longer an untested expectation. A third conversion of that hydrated recipe — a sibling again, differing in 399 .trellis payloads and nothing else — was captured and replayed on the identical v5 shard-0 protocol (512 contexts, 1,048,064 scored positions, same shared head, same comparator): paired against the published checkpoint the difference is −3.755e-06, 95 % source-cluster bootstrap interval [−2.854e-05, +2.062e-05], which brackets zero, on 257 contexts to 255 with no ties. Two controls make that attributable to the sibling's weights rather than to the harness — replaying the published checkpoint against the same reference capture returned its mean and all 512 per-context rows bitwise, and a fresh recapture reproduces that mean exactly — so the comparison's floor is zero, not a tolerance. 97.6 % of the quantized modules come back with different bytes and the fidelity is the same to within our resolution, which is what turns "the recipe is the reproducible thing, the bytes are the artifact" into a measurement rather than a hedge. It is one sibling, one recipe, one shard, at this protocol's resolution: it bounds the converter's fidelity variance, it does not estimate it, and it is not a finding that converter nondeterminism is fidelity-neutral in general (receipts/sibling-rebuild-fidelity.json). Practically: check the digest of the published tree against SHA256SUMS rather than against your own conversion, and read a byte diff below this floor as the converter rather than as tampering, corruption or a changed recipe. The same run also showed the recorded build environment is incomplete — the pinned image has no marisa_trie, which the conversion imports on an unconditional path, so that image provably could not have finished the job; the conversion-capable image is gg-r34-convert. None of this touches the runtime's weight-reconstruction path (reconstruct+hgemm, reconstruct_fp8_slice, reconstructed prefill), which is a different sense of the word: that one is about turning stored trellis bytes back into weights at load, and it is untouched here.

Which of the four builds

Same architecture and tokenizer. The headline KLD column is the v5 held-out suite (5,120 contexts / 10,480,640 scored positions); the older overlap-corrected 127-context v3 subset is kept beside it because the rest of this card's fidelity history is stated in it. The two columns are different suites and are not comparable to each other. Capacity uses each card's documented profile: hydrated, online and K4 are real RTX 5090 MTP-3 tests; context is MTP-3 with an 8.4 MP cap, qualified on a physical RTX 5090 at utilisation 0.955. These profiles are not interchangeable (collection).

The figure plots the legacy v3 corrected means. The v5 ordering below is identical.

Byte and memory conventions for this table. The download column is whole-tree bytes — every published file of the artifact as its release evidence counted it (receipts/collection-index.json, serialized_bytes.whole_tree_bytes: hydrated 21,610,933,884 B, this build 30,597,231,933 B, context 20,696,053,306 B) — and they are serialized bytes on disk, never resident memory. *The K4 release evidence records no tree count, so that one row is the sum of its safetensors shards, 28,313,841,196 B, read from the published repository. The context edition's resident weight is measured twice: 18.41 GiB as run on the rental RTX PRO 6000 engine-budget proof and 18.19 GiB on the physical RTX 5090 at the qualified 0.955 profile. This table prints the larger figure deliberately, because receipts/vram-class-verdict.json elects 18.41 GiB for every class prediction; the 0.22 GiB gap is the rental-versus-5090 delta, not a change in the checkpoint.

Official Qwen/Qwen3.8-27B-FP8 is 28.51 GiB resident at 0.005294 on the v5 suite (0.012798 on the v3 subset) and runs on stock vLLM, which none of these do.

Measured results

30.60 GB download → 20.32 GiB (21.82 GB) resident weights, measured from the engine's own allocation log, versus 28.51 GiB for official FP8 and 21.34 GiB for Unsloth NVFP4 under identical flags.

Fidelity

The headline evidence for this build is the v5 held-out run: 5,120 contexts x 2,047 scored positions = 10,480,640 full-vocabulary positions, with five candidates scored against the same BF16 reference. It supersedes the 136-context v3 development suite and the 36-context v4 qualification as the primary result. Both are preserved below with their numbers untouched — this section adds evidence, it does not revise theirs.

The widest single view of the evidence: A is the only panel where every family appears together (v5 shard 0, 512 contexts, 1,048,064 positions, ten candidates — this build is the online circle at 0.003141), B is the same suite's 1M → 10M ladder (five vLLM builds, 5,120 contexts at 10M), C1/C2 are the superseded corrected v3 (127 contexts, 259,969 positions) and the source-disjoint v4 (36 contexts, 73,692 positions) on separate y-axes behind a barrier, and D is turboderp's own published protocol (OpenWebText, 65,536 positions), which we have never run. Two rules travel with the figure: the cross-engine floor belongs to the llama.cpp rows only, and no ratio across panels means anything. Generated by tools/make_master_kld_chart.py, which reads every one of our values from receipts/ at runtime.

These numbers are re-derivable, not merely re-runnable. 5,240,320 scored positions — five candidates x 512 contexts x 2,047 positions — reproduce bit-for-bit across independent runs with separate model loads and different harness generations: every measured field identical, including the complete per-context arrays and the whole bootstrap block, with only the capture directory paths and the additively-added tail histogram differing (receipts/capture-determinism.json), and a third harness generation's unwindowed --score-from 0 control returns the same shard-0 means to the last digit (receipts/scored-window-offset.json). The scope is part of the claim and travels with it: one GPU, one driver, one pinned rootfs, enforce_eager=True, max_num_seqs=1, one context per forward, 512 MiB bf16 KV. It is not a claim that vLLM is bitwise deterministic in general — nothing here covers CUDA graphs, max_num_seqs > 1, chunked prefill with more than one chunk per context, other GPUs or drivers, or anything downstream of the logits.

Suite identity

receipts/kld5-suite-manifest.json, schema qwen38-distribution-fidelity/6, suite token digest 510541f6861b589d44932db253ec25d96d6daaeeee4ea2ab9b65329209482b88. 5,120 contexts, 2,047 scored positions each, 842 source clusters. The corpus is 941 documents / 70,348,971 bytes fetched by tools/fetch_corpus_v5.py, logged in receipts/kld5-corpus-fetch-log.json.

Exclusion policy — contamination is zero by construction, not by audit. Every discovered document was scanned at every position for exact normalized 12-token overlap against the exllamav3 calibration data before context selection, and any document with even one hit was dropped whole: 44 of 941 documents excluded (43 code, 1 encyclopedic), 897 eligible. There is therefore nothing to correct after the fact and no "overlap-corrected subset" on this suite — unlike the v3 numbers below, which needed one.

The suite is token-disjoint from the v4 qualification suite (0 of its 160 prior context hashes is reachable), and its windows are exact-advance and non-overlapping: independently verified as 5,120/5,120 unique context token hashes with 0 overlapping windows.

Cumulative means at 10,480,640 positions

Body-only. Both operands replay through one shared BF16 LM head, so no candidate's head quantization is counted and the five rows are comparable to each other.

How closely these absolute numbers may be read. Each mean is a body-only replay value: both operands are projected through the one shared BF16 head, and the replay path is not the engine's own logit path. Replaying the unquantized model against its own live logits measures KL(live ‖ replayed) = 5.83e-04 — 32 v5 shard-0 contexts, 65,504 scored positions, context-bootstrap 95 % CI [5.15e-04, 6.64e-04], top-1 99.10 %, on the same suite, reference capture and shared BF16 head as the means above (receipts/replay-live-floor-v5.json), superseding the six-context v3 derivation of 6.54e-04 (receipts/v3-qualification-bf16.json), which its interval contains — and moving hidden-state storage from BF16 to fp32 moves a candidate's KLD by 5.6 % (docs/24). Absolute values are therefore within-suite numbers: they carry a ~6e-4 implementation offset plus a ~5 % storage systematic, and absolute differences below about 1e-3 are not resolvable. Both offsets are common-mode — every candidate replays through the identical path — so paired differences and orderings are the resolvable quantity: hydrated − online K5/K6 is −4.50e-04 [−4.69e-04, −4.33e-04] on 4,922 of 5,120 contexts (receipts/kld5-10M-paired.json), smaller than the replay floor and resolved because the floor cancels in the pairing. The floor is now derived inside the suite rather than on six out-of-suite v3 contexts, and the rule it licenses is unchanged. What it does not license: it is not a claim that candidate KLDs are 11 % smaller, and it does not let any single absolute mean be read more finely — the 5.83e-04 figure is a mean over 32 contexts whose own means span 3.09e-04 to 1.63e-03 with a worst single position of 0.2534. It is also not the cross-engine floor (0.000507), which is a different control. Method of record: docs/42-kld-method.md.

This build sits 39 % below official FP8 on mean KL divergence and above it on top-1. Its maximum single-position divergence, 22.241, is the worst of the five — a tail property that the mean does not show; the distribution tail below measures the rest of that tail on one shard, and every quantile of it sits below official FP8's.

Per-candidate receipts are receipts/kld5-10M-{hyd,k5k6,ctx,fp8,k4}.json (this build), schema qwen38-kld-ladder-cumulative/2, built by tools/kld_aggregate.py from ten verified per-shard reports produced by tools/kld_ladder.sh: capture six models over 512 contexts, replay five candidates, verify, delete 64 GB of hidden states, next shard.

Paired per-context differences

Bootstrap over 10,000 resamples, seed 1, 842 source clusters, receipt receipts/kld5-10M-paired.json. Negative means the first term diverges less.

Versus official FP8 this is settled: lower divergence on 5,105 of 5,120 contexts, with a paired CI far from zero, while holding 8.2 GiB less weight.

Versus the hydrated sibling it is also settled, and this build loses. Same K5/K6 recipe, attention encoded ahead of time instead of at load, 20.31 against 20.32 GiB resident: hydrated is lower by 0.000450 on average [-0.000469, -0.000433] and lower on 4,922 of 5,120 contexts (96.1 %). The margin is small — about 14 % of this build's mean, and below the ~6.5e-04 live-vs-replayed qualification floor quoted for the older harness, so it is a robust ordering of the replayed captures rather than a promise of a perceptible difference in serving. But it is consistent, one-directional and no longer in question. Choose this build for the launch-time attention width knob and the ~206k demonstrated context that comes with K5 attention; choose hydrated for the last fraction of fidelity and a 9 GB smaller download. Nothing else separates them.

Ladder stability

The run was aggregated at 1M / 2M / 5M / 10M scored positions. For the hydrated candidate, carried through all four checkpoints, the cumulative mean reads 0.002700 / 0.002759 / 0.002699 / 0.002760 — a spread of 6.1e-05 across a tenfold increase in scored positions. The headline numbers are not an artifact of where the run stopped.

Distribution tail

A mean and a top-1 rate say nothing about the worst positions, and one exact maximum is not a tail either. This is the whole right tail, measured on shard 0 of the same suite — 512 contexts, 1,048,064 scored positions — the identical contexts for all five candidates. Receipts receipts/kld5-1M-tail-{hyd,k5k6,ctx,fp8,k4}.json, schema qwen38-kld-ladder-cumulative/2, built by tools/kld_aggregate.py; this build's row is receipts/kld5-1M-tail-k5k6.json.

Method, in one sentence: every qwen38-fidelity-report/2 replay accumulates a 560-bin log-spaced histogram of per-position KLD (KLD_HIST_LOG10_LOW=-12.0, KLD_HIST_LOG10_HIGH=2.0, KLD_HIST_BINS_PER_DECADE=40 in tools/fidelity.py) whose bin counts add across shards, which is what makes cumulative quantiles possible at all.

What it says for this build. The ordering at p50, p95, p99, p99.9 and p99.99 is the same as the ordering of the means, so the mean is not hiding a worse tail. This build's tail is below official FP8's at every measured quantile — 0.0321 against 0.0531 at p99, 0.1446 against 0.2438 at p99.9, 0.498 against 0.812 at p99.99 — and it has the smallest share of positions above 1.0 of the five, 21 positions in 1,048,064 against FP8's 62. It stays behind the hydrated sibling everywhere in the same ordering the means report. The exact maximum is the one place it does not lead: 5.507 on this shard, just above FP8's 5.296, and 22.241 over the full ten-shard run — a single position, with 0.00200 % of positions above 1.0 behind it.

Scope, stated exactly:

• This is one 1,048,064-position shard, not the full 10,480,640-position run. The ten-shard run predates the histogram, so it could not be recomputed without re-running it.
• The quantiles are bin-bounded, not exact: each receipt carries lower / upper / estimate per quantile, with a relative bin width of about 5.6 %. The maxima and the exceedance counts are exact.
• The 10M receipts remain the source for the full-run means, intervals and paired results; nothing here replaces them.

What the v5 numbers do not say

• Absolute KLD is suite-specific. v5 values are not comparable to the v3 values in the next section: the corpus mix differs, and K4 reads 0.029679 there against 0.010604 here. Only within-suite ordering and paired differences transfer between the two.
• Cumulative percentiles come from one shard, not from all ten. The ten shard reports of the 10M run carry no token-level KLD histogram, so median/p95/p99/p999 could not be recombined across them. The tail table above closes that gap on shard 0 (receipts/kld5-1M-tail-*.json); across all 10,480,640 positions only the means, the intervals, the paired results and the exact global maximum exist.
• Captures: the reference survived, the candidates did not. The five candidates' hidden states and the BF16 references for shards 1-9 were deleted shard by shard to fit 135 GB of scratch, so unlike the v3 dataset this run is recomputable from the pinned corpus fetch log and suite manifest. The shard-0 BF16 reference was kept and is published, with the suite, all ten shard views and 79 per-shard reports, so a new candidate can be scored against the identical contexts without recapturing the reference — see Reproduce this.

Against GGUF, measured on our suite

The standing objection to this card's headline is that official FP8 is a throughput format whose quality is Q4-to-Q5 class, so beating it is a weak claim, and that llama.cpp's Q8_0 and Q6_K are the honest bar. That is now measured rather than argued.

Three GGUFs from unsloth/Qwen3.8-27B-GGUF@f1bfb127c64f7072bdd2cad55f258b9c8b2910fe were captured under llama.cpp pinned at commit ece963f41b0b02d7a0d61436ae365762c073a4c8 with tools/gguf_capture.cpp, which reads the post-final-norm state — the same mathematical point the vLLM hook takes, with bf16 rounding verified bit-identical to torch on 2,012,449 probe values — and scored against the same BF16 teacher through the same shared BF16 head, on shard 0 of the v5 suite: the same 512 contexts and the same 1,048,064 scored positions every row below saw. Manifests come from tools/gguf_manifest.py and each one carries the GGUF blob digest and the llama.cpp identity; the build script is tools/build_llamacpp.sh. Receipt receipts/cross-engine-comparator.json, per-candidate reports receipts/gguf-report-{q8_0,q6_k,q5_k_xl}.json.

The engine control, measured and not assumed. The unquantized BF16 GGUF captured in llama.cpp against the vLLM BF16 reference on identical tokens, the shared head and the same 512 contexts measures 0.000507 mean, 99.07 % top-1 and p99.9 0.0113 (receipts/gguf-report-engine-floor.json). It proves engine confounding. KL is neither additive nor a metric, so the control is not subtracted and supplies no quantization-only bound.

The p99.9 column, and why it differs from the tail table above. These p99.9 values are each report's exact shard-0 p99.9 as the comparator receipt read them; the tail table above quotes the bin-bounded cumulative estimate from the 560-bin histogram, whose bins are about 5.6 % wide — this build reads 0.1446 there and 0.1447 here, and the exact value lies inside the bin the estimate names. The two differ by construction, not by measurement.

Why several serialized cells read —, and they are not the same reason. This download ships attention in BF16 for the runtime to encode at load (30.60 GB on disk, 20.32 GiB resident), so its disk bytes are not a like-for-like payload against a GGUF file and are not presented as one; the two NVFP4 cells are empty because we publish no serialized-byte receipt of our own for a third party's checkpoint — unsloth's 21.34 GiB and gittensor's 18.77 GiB are measured resident weights, a different quantity. The two payload figures that are comparable are immutable_payload_bytes from receipts/collection-index.json (hydrated 21,610,916,123 B = 20.127 GiB, context edition 20,696,033,532 B = 19.275 GiB; the table truncates both to two decimals); they are serialized bytes, never VRAM. The FP8 figure is resident weights and is labelled as such. The K5/K6 row above is this build at its default attention width, K6; the K5 and K4 launch widths trade fidelity for KV room (knob) and were not measured against these GGUFs.

NVFP4 on the identical shard, and it carries no cross-engine term. unsloth/Qwen3.8-27B-NVFP4 at revision 9c73e2da is served by the same vLLM build as our rows, so unlike the GGUF rows there is no engine term to subtract or estimate and it is directly comparable to this build. On the same 512 contexts and the same 1,048,064 positions, through the same shared BF16 head, it measures 0.030115 mean KLD, 95 % CI [0.027637, 0.032965], 93.16 % top-1, median 0.009584, p95 0.10051, p99 0.33546, p99.9 1.6228, exact worst position 10.6285 and mean JSD 0.010104 bits (receipts/kld5-1M-nvfp4.json, with the run's own account in receipts/nvfp4-v5-measurement.json). That is 9.6x this build's 0.003141, 2.9x K4 at the same 4-bit weight class, 5.8x official FP8, 8.8x the context edition, 11.2x the hydrated sibling and 27.7x Q8_0 as measured; its p99.9 of 1.6228 is 11.2x this build's 0.1447.

Paired per context, which is a stronger statement than any ratio of means: NVFP4 loses every one of 512 contexts, against both comparators it was paired against. +0.026706 against the context edition (95 % CI [+0.024465, +0.029285], 0 wins to 512) and +0.024918 against official FP8 ([+0.022756, +0.027424], 0 wins to 512) — not one context anywhere in the shard where it is the better of the pair (receipts/kld5-1M-paired-nvfp4.json). It was not paired against this build, so its distance from this row stays a ratio of means and is not presented as a win count.

gittensor's "RTX5090" NVFP4, measured on the same shard because its card claims the 32 GB / 262,144-token axis by name. gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090 at revision 69274a0d (ModelOpt W4A4 body, MTP and vision left BF16, FP8 KV cache baked into its config) is served by the same vLLM build as our rows, so it carries no cross-engine term. It measures 0.062163 mean KLD, 95 % CI [0.058491, 0.066360], 89.85 % top-1, p99.9 2.5911 — the weakest row on this table, at 2.1x unsloth's NVFP4 and 19.8x this build. Paired per context it loses every one of 512 contexts to this build (+0.059022 in this build's favour, 95 % CI [+0.055548, +0.063005], 512 wins to 0), every one of 512 to official FP8, and 511 of 512 to unsloth's NVFP4 (+0.032048, [+0.030711, +0.033583]) — the same weight format at 2.57 GiB less measured resident weight (18.77 vs 21.34 GiB, identical flags, engine-reported), which prices that memory saving honestly: roughly double the KLD. A bf16-KV control capture moves its mean by only +0.000365 [+0.000058, +0.000679], so its baked FP8 KV cache explains about 1 % of the gap to unsloth — the rest is the weight conversion itself. Its card's serving numbers (18.8 GB weights in VRAM, 275,941-token FP8 KV pool, 80.6 tok/s decode, native 262,144 on one 5090) are its own claims, which we did not run; our measured 18.77 GiB resident is consistent with the first of them, and consistency is not verification. Its only published fidelity evidence is a 20-item smoke that its own card says not to treat as scores (receipts/kld5-1M-gt5090.json, receipts/kld5-1M-paired-gt5090.json, full account with the checkpoint's composition, digests and mirror in receipts/gittensor-nvfp4-rtx5090.json; archival mirror malaiwah/Qwen3.8-27B-NVFP4-RTX5090-archival-69274a0d).

Where this build sits, without format over-attribution. Under vLLM it measures 0.003141, below same-engine official FP8 at 0.005197 and above same-engine hydrated at 0.002700. Cross-engine, llama.cpp Q6_K measures 0.002035 and UD-Q5_K_XL 0.004444. Those are complete-pipeline observations, not quantization-format rankings. The p99.9 ordering is 0.0794 (Q6_K), 0.1447 (this build), 0.2144 (UD-Q5_K_XL) and 0.2440 (FP8); only same-engine differences isolate this runtime.

Correction, 2026-08-16 — the byte axis in the table above is not one axis, and every mixed comparison flattered us. A GGUF row is the whole file of a text-only artifact; our row is tensor payload of a multimodal tree that also carries an MTP draft. Read from each artifact's own tensor table, without downloading any payload (cross-candidate-byte-accounting.json):

All figures GiB. The embedding and head widths are not uniform across GGUF tiers, and the vision encoder is absent from every GGUF text file - it ships separately as mmproj-BF16.gguf, which no earlier comparison of ours counted. What that does to the four published claims, two against us and two for us:

1. 6 bits, Q6_K against hydrated: our sentence understated their byte spend roughly threefold. "+1.186 GiB more file" is +1.198 on tensors and +3.634 GiB of transformer body (23.1 % more than ours). Their fidelity win at 6 bits stands exactly as published - it is the price we mis-stated, in our own favour.
2. 6 bits, deployed: a multimodal Q6_K deployment is 22.18 GiB against our 20.13 GiB whole tree, so ours is 2.053 GiB smaller and needs no second file.
3. 5 bits, UD-Q5_K_XL against the context edition: this claim was wrong against us. We do not pay "0.445 GiB more" for the win - on transformer body they carry 2.148 GiB more (+14.4 %), and our deployed multimodal artifact is 0.422 GiB smaller.
4. 8 bits, Q8_0 against online K5/K6: the bodies agree to 1.7 % (24.526 against 24.119), the most format-comparable pair on the table, and they win it cleanly. That row is the reason the others are worth reading: this is not a table where every axis favours the author.

Rule from here on, and it is printed rather than footnoted: "at equal file bytes", "at equal tensor bytes", "at equal transformer body" and "as a deployed multimodal artifact" are four different claims, and whichever one a sentence means is written into the sentence. No fidelity number changes.

Update 2026-08-16 — a near-equal-file-byte K6 build. The hydrated recipe with gate_proj and up_proj promoted K5 → K6 measures 0.001634 mean KLD [0.001541, 0.001742], top-1 98.248 % (k6-parity-kld.json). It beats hydrated by 0.001066 [0.000974, 0.001173] on 511 of 512 contexts for +1.348 GiB, and its complete vLLM pipeline measures 0.000401 below the llama.cpp Q6_K pipeline on 493 of 512. The cross-engine control cannot turn that into format parity or prove a byte-gap mechanism. It carries 2.306 GiB (13.5 %) less transformer body than Q6_K.

Two disciplines this exercise was run under, both worth more than the result: the payload was predicted at 23,035,310,148 B and measured at 23,035,310,148 B, zero error; and the registered 95 % interval [0.001175, 0.001601] missed — the measurement is 0.0000328 (2.0 %) above its upper bound, so all three estimators bracketed on the pessimistic side, with the role-share bound worst and the byte-law-at-Q6_K's-surplus form best, the opposite of what the surrogate's calibration implied. The published point prediction of ~0.0016 was within 2.1 % of measurement, the most accurate advance prediction this project has made — and the miss is printed beside it rather than dropped.

The complete-pipeline observations worth stating plainly:

1. At the nominal 6-bit point, llama.cpp Q6_K measures 0.002035 and vLLM hydrated 0.002700.
2. At the nominal 5-bit point, the vLLM context edition measures 0.003409 and llama.cpp UD-Q5_K_XL 0.004444.

The engine mismatch prevents format attribution. Q8_0 has the lowest measured complete-pipeline KL at 0.001087 for 27.05 GiB, and every GGUF point at or above 5 bits measures lower than official FP8. Those cross-engine observations make this card's "lower divergence than official FP8" headline a weaker achievement than it sounds. Same-engine K4 and Unsloth NVFP4 comparisons remain directly attributable to the tested vLLM pipeline.

What this comparison does not settle. It is text-only teacher-forced fidelity on one shard of ten. It says nothing about serving 262,144 tokens with vision and MTP on a 32 GB card, which is where these artifacts actually differ, and llama.cpp KV-quant behaviour, prefill and decode speed are separate axes that were not measured here. The GGUF rows are a shard-0 ranking, not a paired per-context bootstrap against the ten-shard rows in Fidelity, because those were welded from a different position count. Shard 0 is one tenth of the suite, and it is close to it: over all 10,480,640 positions the five vLLM means read 0.002760 / 0.003210 / 0.003509 / 0.005294 / 0.010604 — 1.9-2.9 % above these shard-0 values, ordering unchanged (receipts/kld5-10M-{hyd,k5k6,ctx,fp8,k4}.json). The GGUFs have no ten-shard equivalent; extending them is unrun.

One protocol objection, bounded rather than argued. llama-perplexity scores only the second half of each window, so every position it scores has at least 256 tokens of left context, while our suite scores from position 0. Re-scoring our own captures under that restriction lowers every candidate's mean by 1.3-2.1 % at a 256-token floor and 3.9-4.9 % second-half-only, uniformly enough to change no ordering — this build reads 0.003100 and 0.003020 respectively (receipts/scored-window-offset.json). The external protocol's scoring floor therefore explains at most about 5 % of any cross-protocol gap, and nothing in the ordering above.

Cross-citation: the same three GGUFs under llama.cpp's own protocol

The rows above are those GGUFs on our axis. They have also been measured on theirs, run exactly as its authors run it, so the two can be cited side by side without either being converted into the other: llama-perplexity --kl-divergence on WikiText-2 raw test, n_ctx 512, 147,900 scored positions, KL(BF16 GGUF ‖ candidate) with both operands inside llama.cpp and each candidate's own output head inside the measured path, base Mean PPL 6.950230 ± 0.044933 (receipts/wikitext-kld-run-a.json; full protocol, delta by delta, in docs/35-external-protocol-comparability.md).

The ordering is identical, but no ratio or conversion is valid across the protocols. They differ in corpus, scoring window, head and engine; our GGUF rows are additionally cross-engine. The 0.000507 BF16 control proves confounding and cannot be subtracted. Correction, 2026-08-16: This paragraph used to call the output head "the large one". It is now measured on our own corpus and it is not: replaying each candidate through its own head over all 512 shard-0 contexts and 1,048,064 positions raises its mean by at most 5.28 % (hydrated 5.01 % of head-inclusive divergence, context 4.06 %, K4 1.17 %, unsloth NVFP4 2.64 %, and exactly 0 % for the official FP8 export, whose head is byte-identical to the shared one — the internal control), every interval excluding zero (receipts/head-attribution-v5.json). Scoring geometry is worth ≤4.9 % by the same kind of control. So the two protocol terms we have quantified are together far too small to explain a 1.1-1.6x level difference: the level gap is not decomposed, the leading unmeasured candidates being their 512-token English-encyclopedic windows against our 2,048-token five-strata ones, and the width of a GGUF's own output.weight, which is a different tensor from any head measured above.

Their harness's own floor, measured on our hardware instead of assumed. The Minimum KLD column is negative for all three — −0.000080, −0.000056, −0.000077, i.e. 5.6e-5 to 8.0e-5 — the uint16 16-nat log-probability encoding showing through rather than a candidate beating its own reference. The same term appears in the perplexity: 6.9525 in the capture log against 6.950230 in the scoring runs, identical weights on identical tokens, differing only by that stored round trip.

Tokenization is not part of the difference, and that is a measured null result. llama.cpp's GGUF BPE and our Hugging Face tokenizer produce bit-identical 297,194-token streams over this corpus — same int32 digest, no first divergence index — and the 296,960-token prefix that llama-perplexity actually scores is identical too (receipts/wikitext-kld-token-identity.json).

And one finding worth its own line: perplexity does not reproduce the KLD ordering. Q6_K has the smallest PPL delta of the three, +0.00079 against the 6.950230 base, while Q8_0 — the better quant by every divergence statistic, including a mean 2.5x lower and 0.9 points more top-1 agreement — is +0.00467. A quantization that shifts the distribution can shift it in the direction that happens to flatter a corpus mean, which is an argument for the metric this whole section is built on and against ranking quants by perplexity delta.

What this cross-citation cannot do is put this build on their axis: llama-perplexity cannot read an EXL3 checkpoint. The table at the top of this section, where every candidate is scored by one harness on one suite, stays the primary comparison.

Prior receipt: v3 development suite (136 contexts, 278,392 positions)

Superseded as the headline by the v5 run above; kept because every earlier claim on this card, including the as-served head cost and the attention-width ladder, is stated on this suite.

The figure's four points and four top-1 values are the overlap-corrected 127-context subset; the table below is the original full 136-context receipt, which is why 96.95 / 96.18 / 94.48 / 90.49 % there reads 96.97 / 96.22 / 94.50 / 90.53 % here.

KL(BF16 reference ‖ candidate), two passes, no top-k, float32 within vocabulary chunks accumulated in float64 across chunks, one shared BF16 LM head for both operands, source-cluster bootstrap. Corpus is separately sourced Gutenberg / arXiv / Wikipedia / CPython. The builder requested nine Wikipedia languages but tolerated under-filled strata, so the frozen suite is English, German and Russian only (the multilingual stratum is 6 German and 1 Russian context). Its fixed-stride 160-character scan originally reported zero calibration hits; an offset-independent 12-token scan later found exact overlap in 2/41 source documents. The suite, shared BF16 head, sentinels, comparator captures and this checkpoint's captures are published as a dataset. The complete original 136-row report is under reports-k5k6/report-k5k6-online-k6-analysis.json; its capture is under candidate-hidden/k5k6-online-k6/.

Overlap-corrected subset: conservatively removing all nine analysis contexts from either affected source document gives this quant 0.007945, FP8 0.012798, K4 0.029679, and NVFP4 0.092727 over 127 contexts. The 38 % FP8 advantage and ordering survive. The table above remains the original full-suite receipt, not a claim of zero lexical overlap.

The v3 NVFP4 number is not our current one, and the two suites must never be mixed. NVFP4 reads 0.092727 on the corrected subset above and 0.030115 on v5 shard 0 — same checkpoint, same revision, same flags, same shared-head protocol — and its v5 row is in the shard-0 table above. That gap is suite hardness, measured for all six candidates rather than argued: v3-corrected ÷ v5 shard 0 is 2.4625x official FP8, 2.5293x this build, 2.6564x hydrated, 2.7505x the context edition, 2.8688x K4 and 3.0791x NVFP4 — a band spanning 1.2504x end to end, with the ordering identical in both suites (receipts/nvfp4-v5-measurement.json, block suite_comparability_v3_vs_v5). A band of factors and not one factor is why no conversion between the suites exists: the ordering carries across, an absolute value never does, and a v3 number must never appear in the same sentence as a v5 number.

Paired over the same contexts: -0.004969 versus official FP8 (95 % CI [-0.00643, -0.00371], 136/136 contexts), i.e. 38 % lower mean KL divergence than FP8 while holding 8.8 GB less weight; and -0.022579 versus the previous K4 release (136/136 contexts).

Body-only versus as-served, measured on this v3 suite. Every row above — and every row in the v5 section — replays both operands through one shared BF16 head, so no candidate's head quantization is counted; that is what makes them comparable. The only measurement of what this build's own K6 head costs when served was made here, on the v3 suite, and has not been repeated on v5. On the original 136-context receipt the K6 head adds +0.000127 (95 % CI [+0.000105, +0.000148]) for 0.008284 as served. On the overlap-corrected subset, body-only is 0.007945 and the measured as-served result is 0.008078 (+0.000132, 95 % CI [+0.000114, +0.000151], 7/127 contexts favour the quantized head). Promoting the head to BF16 would cost +1.589 GB, so this checkpoint keeps K6. That increment is a v3-suite quantity: it must not be added to the v5 means above, which are a different suite.

Controls published with the dataset: runtime-repeat noise floor 0.000000 (three captures of the same runtime) and harness self-check 0.000000. A third control, "CUDA-graph parity 0.000000", was withdrawn: it captured a prefill forward, and FULL_DECODE_ONLY captures no prefill graph, so it could not have measured the decode path. Re-measured properly on real decode steps, graph and eager agree on 24/32 greedy 32-token sequences with mean |Δ logprob| 0.0118 on the chosen token; unquantised BF16 on this same build drifts identically (24/32, 0.0128), so this is a property of CUDA graphs here and not of the quantisation. Weakest control: live-vs-replayed logit qualification is 5.83e-04, re-derived inside the v5 suite (32 contexts, 65,504 positions, [5.15e-04, 6.64e-04], superseding the six-context v3 figure of 6.54e-04), so differences below ~1e-3 are not resolvable with these artifacts. The FP8 gap is 7.6x that floor and the K4 gap 34.5x it; the K6-head increment (0.000127) and the K5-vs-FP8 gap (0.000991) are at or below it and are reported as unresolved point estimates, not as established differences. The KLD magnitudes here are only comparable within this suite — thresholds from other models, corpora or tokenizers do not transfer.

Attention width is a runtime knob, and native context needs the K4 build

Attention weights ship in BF16 and are encoded to EXL3 Trellis at load, so the width is a launch-time choice rather than a property of the download: VLLM_EXL3_ONLINE_TRELLIS_BITS accepts 3-8. One checkpoint, several operating points.

The width ladder was measured on the v3 suite only. The v5 run scored this build at its default K6 attention (0.003210); the K5 and K4 attention widths have not been rerun on v5, so the three rows below are v3-suite quantities and belong beside 0.007945, not beside 0.003210.

On the same overlap-corrected 127 contexts, K5 costs +0.003856 versus K6 and K4 costs +0.018673. K5's mean is 0.000997 below official FP8's corrected 0.012798 — at this harness's ~1e-3 replay-resolution floor, so treat it as an unresolved point estimate, not an advantage. The original 136-context width reports remain in the dataset.

Measured on a real RTX 5090, by an independent tester

Native 262,144 context does not fit this checkpoint on a 32 GB card. An earlier revision of this card claimed it did, from a memory simulation on a 96 GB card. That was wrong twice: the simulation omitted MTP's KV (with num_speculative_tokens: 3 the engine needs 9.13 GiB for 262,144 tokens, not 8.18) and it assumed 31.84 GiB usable where a 5090 reports 31.39. Corrected, with hardware numbers (TP1, FP8 E4M3 KV, MTP-3, decode-only CUDA graphs, vision enabled):

So on a 32 GB Blackwell card with MTP-3: use K5 attention for ~206k with retrieval verified, keep utilisation at 0.95 if you serve images (0.98 leaves no vision headroom), and if you need native 262,144 take the context edition, which is hardware-qualified on a physical RTX 5090: 265,122 KV tokens at 262,144 with MTP-3 and the full 8.4 MP ceiling, 1.01x concurrency at native length, measured at --gpu-memory-utilization 0.955 with PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True, all seven gates passing (receipts/qualification-5090-context.json). That 0.955 is a per-card measurement, not a constant. It is the value that qualified on one board, GPU-506a575d (32,607 MiB, 458 MiB of it held by the driver); a second physical RTX 5090 needed 0.956, missing at 0.955 by about 0.01 GiB (receipts/second-5090-datapoint.json). Two nominally identical boards differ in exactly two quantities no configuration can move — the driver's framebuffer reserve and the CUDA context size — and a 68 MiB perturbation in either was measured to be enough to flip a gate (receipts/qualification-24gib-capped.json → residual_risk_versus_a_physical_board), while one thousandth of utilisation is only about 32 MiB. So if a card refuses to start or OOMs at startup, raise utilisation by 0.001 at a time rather than dropping the window — and do not lower max_pixels to make room, because at fixed utilisation that enlarges the KV pool and makes the large-image case fail sooner.

malaiwah/Qwen3.8-27B-K4 also holds native length at the physical limit, at 3.8x the divergence and with no overlay at all. Closing this online-K6 build's last 0.83 GiB through utilisation alone would need ~0.997, with no runtime headroom. On 48 GB and larger, native context fits at K6 with the best fidelity.

The 0.96 and 0.97 rows above are startup probes, not serving recommendations. No profile of this build has been through the seven-gate qualification at any utilisation — the tester's 0.95 arms are the only ones with a served long-context result behind them, and the 0.98 arm already OOMed on a small image. The context edition's 5090 run is the measured warning: at 0.97 it started and served text fine, then killed the vision tower on a combined long-text-plus-7 MP request wanting 62.00 MiB — with 26.50 MiB free under PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True (bounded_negative_results arm B2) and 34.56 MiB free with no allocator configuration at all (arm B) — and lowering max_pixels instead of utilisation made it strictly worse because the engine spends every freed byte on KV. Until the same gates are run here, treat any utilisation above 0.95 on this build as unmeasured for image serving.

KV costs 34,816 B/token with MTP-3 (16 full-attention layers, 4 KV heads, head_dim 256; the other 48 layers are Gated DeltaNet and hold per-sequence state instead) and 32,932 B/token without it — turning MTP off is worth about 11 % more context if you would rather have length than 2x decode. Per-token is only half the model, and the missing half matters: what one request needs is affine in the window, a·L + M, with a fixed per-request term M of 0.63 GiB under MTP-3 and 0.14 GiB without, both measured by provoking startup refusals at two windows (docs/34-vram-class-profiles.md §4.1). Two things follow that a per-token figure alone hides. Dividing a KV budget by a per-token rate does not predict a context length. And because M is charged per request, raising --max-num-seqs pays it again for every slot — at MTP-3 that is 0.63 GiB per concurrent stream before a single token is stored. The figure this card published previously, 37.4 KB/token, was pool ÷ reported tokens at one 262,144 window: a ratio that silently folded M in and overstated the coefficient by about 8 %.

4-bit KV is not available on this architecture: the runtime's generic NVFP4 KV path requires SM100 trtllm-gen and is rejected on SM120, and GLM-5.2's nvfp4_ds_mla cache is MLA-specific, which Qwen3.8 is not.

Thanks to the tester who ran this on real hardware and caught the overclaim.

Prior receipt: v4 post-selection qualification (36 contexts)

The v3 numbers come from the suite that guided recipe selection. This was the first test that did not: 160 new contexts from 100 documents with zero intersection with the development suite (context token hashes 0/160, document names 0/100, content hashes 0/100), partitioned by whole source cluster, run once, with no recipe changed afterwards. The v5 suite above is token-disjoint from this one too, is 32x its size (5,120 contexts / 10,480,640 scored positions against 160 / 327,520) with the same conservative rule applied at document-scan time instead of afterwards, and supersedes it as the headline held-out evidence; the ordering it found is unchanged there.

The original 42-context table used a fixed-stride character overlap scan. A later, offset-independent scan found exact 12-token calibration overlap in four qualification source documents. Applying the same conservative rule to every candidate — exclude every context from any source document with even one hit — leaves 36 contexts / 24 clusters:

The correction changes no ordering or paired win: the three EXL3 builds remain 47 / 41 / 32 % below FP8. The original 42-context figures and the candidate-independent correction are both preserved in docs/31. Absolute magnitudes remain suite-specific, and these 36 contexts (73,692 scored positions) are not comparable to the v5 suite's 5,120 — where this build reads 0.003210 against FP8's 0.005294 and wins 5,105/5,120 paired contexts, on 142x as many scored positions.

Public capability — MMLU-Pro, item-paired against BF16

70 MMLU-Pro questions, 14 official categories, 5 per category, pinned TIGER-Lab/MMLU-Pro@b189ec765aa7ed75c8acfea42df31fdae71f97be, official five-shot category prefixes, greedy, thinking at low reasoning effort, 5,120-token completion cap. The BF16 control ran first and the acceptance rule was frozen in receipts/public-capability-plan.json before any candidate result was seen. All six models answered the same 70 items in the same order through the same extractor, so every candidate row is paired item-by-item against that control.

The pre-registered bar, and this build's verdict

The frozen plan accepts a candidate when BF16-pass retention has a Wilson 95 % lower bound at or above 0.90 and no category loses more than two BF16 passes. The category clause is met by all five candidates — the worst case is two passes in philosophy, for this build and for the hydrated build — so the retention lower bound is the only clause that ever fails.

Only the context edition clears the bar, at 90.7 %. K4 and official Qwen/Qwen3.8-27B-FP8 read 88.1 %. This build reads 85.6 % (54/57) and does not clear it, as does the hydrated build. Three of BF16's 57 passes flipped to failures here and one BF16 failure flipped to a pass, for 55/70 absolute. That is a measured shortfall, published exactly as measured, with nothing retuned afterwards.

What the numbers do not say. 55/70 is the lowest absolute count in the matrix, and that is not a ranking: every interval in the table overlaps every other interval, including the BF16 control's and official FP8's. This build is not shown to be worse than official FP8, K4, the hydrated build or the context edition on knowledge-and-reasoning tasks, and none of them is shown to be worse than it. The KLD advantage this card reports over official FP8 is a distribution-fidelity result on 10,480,640 scored positions; it makes no capability claim, and this 70-item suite neither confirms nor contradicts it.

Why a 70-item suite cannot certify this bar

With 57 BF16 passes as the paired denominator, 56/57 is the smallest count whose Wilson 95 % lower bound clears 0.90 (56/57 → 90.7 %; 55/57 → 88.1 %; 54/57 → 85.6 %). A single paired regression is therefore the entire budget, and no result that gives up two can pass, however sound the build. The suite has too few items to certify the bar it pre-registered, and at this size it separates nothing — the point applies to official FP8 exactly as it applies to the EXL3 builds. Read it as a power limitation of a 70-item draw, not as evidence that any of these checkpoints is broken.

Two protocol facts that bound the reading

• Exact-answer agreement is 0/70 for every EXL3 candidate, and 1/70 for official FP8 (one math item, a 113-token answer both models pass). Long chains of thought differ token-wise on essentially every item, so pass/fail outcome is the only meaningful pairing unit; nothing here is a generated-text match claim.
• Four BF16 items end at the 5,120-token completion cap with no letter emitted and are scored as failures under the plan's frozen addendum, so the control itself is depressed by the cap; per-model counts are in the table (this build: 4). The earlier 2,048-cap control, where BF16 lost 7/70 to truncation, is retained unchanged at receipts/public-capability-bf16-superseded-cap2048.json.

Status of this evidence

This is a first public, licence-compatible, item-paired benchmark, not a leaderboard claim. The honest next step is more items, which is the plan's own P1: HumanEval+/MBPP-style executable cases, IFEval-style constraint following, tool schemas, and a larger MMLU-Pro draw. No capability claim on this card graduates before that.

Harness tools/public_capability.py, sweep runner tools/run_public_capability.sh, suite receipts/public-capability-suite-mmlupro-70.json. Every run receipt carries the per-item raw request, raw response, extracted letter, gold letter and digests.

Downstream task retention — 40-task smoke suite (prior, narrower evidence)

This ran before the MMLU-Pro suite above and is kept unchanged. It is the narrower evidence: self-generated tasks with contract checks, not a public benchmark.

On 40 deterministic generated tasks (10 each arithmetic, executable builtins-only code, exact-list instruction following and tool-call schema), BF16 and every comparator scored 40/40. This build had zero regressions and matched BF16's exact final-answer text on 34/40; all six differing answers still passed their contracts. Wilson 95 % lower bound is 91.2 %. This is a transparent smoke suite, not a public leaderboard; full responses are in the run receipt; its extracted-value agreement field is superseded by the strict rescore.

Throughput

Median of 3 runs, --max-num-seqs 8, greedy, 256 output tokens; prefill measured with exact token-count prompts.

* MTP figure predates the prefill patch; the two are independent and compose.

Best decode throughput of every candidate measured, and 2.3-2.5x the single-stream rate of FP8/NVFP4 with speculative decoding on. Speculative decoding uses this checkpoint's quantized draft head: 58.2 % of drafted tokens accepted, 1.745 accepted draft tokens per step, so 2.745 output tokens per speculative iteration once the verifier's own token is counted (acceptance 77.5 / 57.2 / 39.8 % by draft position). The comparators are measured without speculative decoding, so the 113.8 figure is this checkpoint against itself, not a like-for-like format comparison.

Prefill improved 2.1x in PR #316 plus PR #318: #316 adds the reconstruct+hgemm path for rows >= 128, while #318 routes native K6/MCG B12X shards to that path instead of the decode-shaped kernel. The change is slower below m=64 and 4.1-5.2x faster at m=2048; decode is untouched. It is not bit-exact — fp16 summation order changes, costing +0.43 % measured divergence with top-1 unchanged to four decimals. VLLM_EXL3_PREFILL_RECONSTRUCT_M=0 restores the exact path.

Prefill remains the weak axis, and it is now attributed rather than suspected. A 2x2 over attention representation and MLP kernel isolates it: the MLP kernel is worth 2.13-2.26x, the online attention overlay only 1.05-1.11x — so the overlay is not the bottleneck, which refutes what this card previously said. ext.hgemm measures at cuBLAS parity (0.92-1.06x) and a larger prefill chunk changes nothing, so the tuning levers are spent. The residual is the GEMM's dtype: at these shapes an FP8 matmul runs 1.85-2.02x faster than fp16 on this card, which is almost exactly official FP8's prefill lead. Closing it needs dequant emitted into an FP8 GEMM, not another dispatch tweak — the work is tracked in docs/26.

Concurrent serving: speculative depth is a concurrency-dependent choice

Measured on the user's own physical GeForce RTX 5090 (32,607 MiB, driver 610.57.04) on the immutable production image localhost/vllm:gg-r34-patched (sha256:6eca4c69…) with no source bind mounts, and on the context edition rather than this checkpoint — 11 configurations, identical frozen token-id prompts, three warmed repeats (receipts/perf-sweep-5090.json, decision record docs/36-performance-levers-5090.md). Per that receipt's own rule, none of these figures may be differenced against the table above or any other measurement on this card; they are carried here because the finding is a serving recommendation for the whole family.

The decision metric is accepted tokens per step ÷ step time, never acceptance rate. The reference row, MTP depth 3 at --max-num-seqs 8, reads 82.94 / 263.12 / 313.28 tok/s aggregate at 1 / 4 / 8 concurrent streams with step times 25.72 / 30.64 / 49.52 ms. At one stream, MTP depth 3 wins: 2.1429 accepted tokens per step over a 25.72 ms step = 83.31 per-request tok/s against depth 1's 1.6558 over 22.29 ms = 74.28. At eight streams it reverses decisively: num_speculative_tokens=1 costs 22.98 % of accepted tokens per step (2.1753 → 1.6754) but takes 37.25 % off step time (49.52 → 31.07 ms), so aggregate throughput rises 30.67 %, 313.28 → 409.35 tok/s (+22.75 % per request), and it holds 10,911 more KV tokens (283,481 against 272,570) — the only row whose needle ran at the full 261,794 tokens, retrieved exactly, with the 30-case image suite unchanged at 24/30. Depth 2 is dominated at both ends, and at concurrency 4 the choice does not matter (±2.6 %). So keep depth 3 for interactive single-stream use and set num_speculative_tokens to 1 when the deployment really serves concurrent streams. That 8-stream matrix ran at --gpu-memory-utilization 0.97 at 262,144 tokens, which is a text-only profile — a large image OOMs in the vision tower there — so a vision-capable deployment keeps one sequence at the qualified 0.955. The crossover exists because each drafted token costs a fixed slice of step time whose acceptance does not improve with batch size, while a wider batch already fills the step: past some concurrency the cheaper step buys more than the deeper draft.

The concurrent-serving variant, then, is the recipe-B command below with one value changed: --speculative-config '{"method":"mtp","num_speculative_tokens":1}' instead of 3, alongside the --max-num-seqs 8 it already carries. Nothing else moves.

Closed avenues, so nobody re-runs them. --attention-backend FLASHINFER is a no-op: the engine already auto-selects FlashInfer on SM120 with fp8 KV and head_size 256. Forcing TRITON_ATTN is up to 5.5 % worse on step time at eight streams, and its apparent +5.5 % gain at temperature 0 becomes −7.3 % at temperature 0.6, so it is acceptance noise rather than throughput. custom_ops:["all"] is 2.2-5.2 % worse on step time and is not bit-exact. Both dynamic speculative-decoding knobs are structurally unusable on this build: either one downgrades cudagraph_mode from FULL_DECODE_ONLY to PIECEWISE, which Exl3Config refuses, so the server does not start — and forced eager, the only form that runs, loses 48 % of decode. Dynamic speculative depth is no longer closed — it was a fixable bug, now fixed and measured (2026-08-16). The cudaErrorIllegalAddress that made VLLM_USE_V2_MODEL_RUNNER=1 unusable was a FlashInfer gate that admitted persistent (CUDA-graph) decode wrappers only for q_len == 1 + num_speculative_tokens, while the speculator's draft steps run q_len == 1; the draft-decode graph was therefore captured and replayed on the dynamic wrapper, whose plan buffers move on every call. Keying wrappers by the shape capture actually planned fixes it (local-inference-lab/vllm#398, closes #396; root cause proven with an instrumented capture-versus-replay address log plus a control that changes nothing except pinning the capture-time buffers alive). With the fix, one server running the per-batch-size schedule (depth 3 at batch 1-2, depth 1 at batch 3-8) measures 87.4 tok/s at C1 and 416.3 tok/s at C8, +38.0 % aggregate decode at C8 against an MRV1 baseline measured in the same window (301.6 tok/s) with C1 held — depth-1 throughput at eight streams without giving up depth-3 latency at one. Greedy outputs repeat token-for-token within an engine process in all four arms, and acceptance does not drop, so the speed is not bought with looser verification. The constraint that decides whether you want it: at --gpu-memory-utilization 0.97 the V2 runner leaves 58.56 MiB free and the EXL3 prefill reconstruct OOMs on the first 2,048-token prefill (both arms, schedule off and on), so the win was measured at 131,072 context and 0.95 utilisation — 234,256 KV tokens against 272,570 on this card's published profile, −14.1 %. Serving the published 262,144/0.97 profile under the V2 runner on a 32 GB card is not demonstrated. Cross-restart bit-exactness remains unclaimable on this stack (exl3_gemm autotune is per process). The static depths, 3 single-stream and 1 at eight streams, remain the published recipe until that KV concession is either accepted or removed (receipts/v2-runner-depth-schedule.json). Prefill did not move on any lever (3,374.4 tok/s at 2,048 and 3,255.4 at 6,144 prompt tokens for the reference row, no graph-decode arm more than 4.0 % away), which is the same conclusion the attribution work above reached from the other direction: the prefill deficit is structural, not untuned.

One reconciliation, because both numbers are published: that 82.94 tok/s at one stream sits below the context edition's qualification median of 107.56 tok/s purely because of acceptance, not speed — 2.14 accepted tokens per step here against 2.69 there, since these frozen prompts are literary prose — while step time agrees to 2.6 % (25.72 ms against 25.05 implied). The two measurements are consistent.

Serving

Two recipes, because the published image predates the patches. The pinned digest below was built on 2026-08-10 from vLLM e2666d9a; the maintained patch stack remains unmerged, so that image cannot contain it. Recipe A is what the image runs unmodified. Recipe B replaces one module before launch. The exact module used for the headline table is preserved below; a later superseding module adds K6/MCG prefill routing. Anything claiming graph decode or reconstructed prefill on recipe A is wrong.

Recipe A — unmodified image, eager only

docker run --rm --gpus '"device=0"' --ipc host -p 127.0.0.1:8000:8000 \
  -v /models:/models:ro -v /cache:/cache \
  -e VLLM_EXL3_ONLINE_TRELLIS_BITS=6 \
  -e VLLM_EXL3_ONLINE_CACHE_DIR=/cache/exl3-online \
  --entrypoint /opt/venv/bin/vllm \
  voipmonitor/vllm@sha256:820181fbbc975cd5291c411cda9771d58fecee1636d916f508f47230df20592b \
  serve /models/Qwen3.8-27B-EXL3-K5K6 \
    --served-model-name qwen38 --quantization exl3 --enforce-eager \
    --quantization-config '{"linear":{"weight":"mxfp8"},"ignore":["re:.*visual\\..*","re:.*in_proj_a$","re:.*in_proj_b$","re:.*in_proj_ba$","re:.*mtp\\..*","lm_head"]}' \
    --mm-processor-kwargs '{"truncation":false}' \
    --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder \
    --max-model-len 8192 --gpu-memory-utilization 0.85 --max-num-seqs 8 \
    --host 0.0.0.0 --port 8000

Measured on this path: 28.8 tok/s decode at concurrency 1 and 2.4k tok/s prefill — that is the honest floor without the patches.

Recipe B — headline configuration (graphs + prefill dispatch)

Replace one module inside the container, then launch. The exact headline-table module is vllm-exl3-prefill-dispatch.py at 21c2b6d, sha256:cb9e60024057e8097237a5518e6469b15f73e4139cc37f1f67e9c1485b44aedd. It incorporates PR #314 (7917c928) and PR #316 (8451183e). The current recommended main module, sha256:2df9d0799fd323798cead1edb773cab556c94798eec263ee03ded35408c6e4ee, also incorporates PR #318 (5da0bcda) and is what the later native-context receipt tested. The 8,192-token three-run table was not relabelled as a measurement of that later file.

set -euo pipefail
PATCH=$PWD/vllm-exl3-prefill-dispatch.py
SCHED=$PWD/vllm-mamba-align-scheduler.py
printf '%s  %s\n' \
  2df9d0799fd323798cead1edb773cab556c94798eec263ee03ded35408c6e4ee "$PATCH" \
  b431c1066dfee3ed56bfa7e71cc8606f9afadc300f22d7fc542c43835d1b22bf "$SCHED" |
  sha256sum -c -

docker run --rm --gpus '"device=0"' --ipc host -p 127.0.0.1:8000:8000 \
  -v /models:/models:ro -v /cache:/cache \
  -v "$PATCH:/opt/venv/lib/python3.12/site-packages/vllm/model_executor/layers/quantization/exl3.py:ro" \
  -v "$SCHED:/opt/venv/lib/python3.12/site-packages/vllm/v1/core/sched/scheduler.py:ro" \
  -e VLLM_EXL3_ONLINE_TRELLIS_BITS=6 \
  -e VLLM_EXL3_ONLINE_CACHE_DIR=/cache/exl3-online \
  -e VLLM_EXL3_GRAPH_DECODE=1 \
  -e VLLM_EXL3_PREFILL_RECONSTRUCT_M=128 \
  --entrypoint /opt/venv/bin/vllm \
  voipmonitor/vllm@sha256:820181fbbc975cd5291c411cda9771d58fecee1636d916f508f47230df20592b \
  serve /models/Qwen3.8-27B-EXL3-K5K6 \
    --served-model-name qwen38 \
    --quantization exl3 \
    --quantization-config '{"linear":{"weight":"mxfp8"},"ignore":["re:.*visual\\..*","re:.*in_proj_a$","re:.*in_proj_b$","re:.*in_proj_ba$","re:.*mtp\\..*","lm_head"]}' \
    --compilation-config '{"cudagraph_mode":"FULL_DECODE_ONLY"}' \
    --speculative-config '{"method":"mtp","num_speculative_tokens":3}' \
    --enable-prefix-caching --mamba-cache-mode align \
    --mm-processor-kwargs '{"truncation":false}' \
    --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder \
    --max-model-len 8192 --gpu-memory-utilization 0.85 --max-num-seqs 8 \
    --host 0.0.0.0 --port 8000

No published image digest contains this patch. Until one does, recipe B is a content-verified source mount, not an immutable runtime artifact. The container listens on all interfaces internally, but Docker publishes the port to host loopback only. For remote clients, keep that binding and put an authenticated TLS proxy in front; do not expose this unauthenticated generation endpoint directly.

Load-bearing details:

1. --quantization exl3 is mandatory — auto-detection only fires for GLM-5.2 metadata.

2. Both performance features need unmerged patches, which is why recipe A and recipe B are separate. VLLM_EXL3_GRAPH_DECODE=1 + cudagraph_mode: FULL_DECODE_ONLY need PR #314 (without it the loader refuses non-eager execution and decode loses ~46-50 %). Reconstructed generic EXL3 prefill needs PR #316. The maintained module additionally uses PR #318 to route native K6/MCG B12X shards into that path at prefill row counts. Current upstream heads have no CI result: pre-run jobs are blocked by repository policy labels, not by a demonstrated failure.

3. The ignore list is mandatory and its anchoring is subtle. Prefixes carry no leading model., so re:.*visual\..* matches while re:.*\.visual\..* silently does not — and the wrong pattern crashes startup (#311, fixed by PR #312). The tested config also ignores mtp.* from the generic BF16 online overlay. EXL3 ownership is checked first, so the serialized quantized draft still uses the EXL3 loader; the ignore prevents accidental re-encoding only when a draft projection is not EXL3-owned.

4. --mm-processor-kwargs '{"truncation":false}' is required for images whose expanded token sequence exceeds 2,048, or requests fail with HTTP 400 (#313). Verified here: a 2044×1622 image answers correctly with the flag.

5. First load encodes 208 attention projections to K6 — measured 957 s cold against 173 s with a warm cache at an 8,192 window; the hydrated sibling, with nothing to encode, starts in 178 s (receipts/startup-times.json). Point VLLM_EXL3_ONLINE_CACHE_DIR at persistent storage.

6. Prefix caching is on in the recipe below, and it is on here and not everywhere. At an 8,192-token window the KV pool is roughly thirty times the window, so a single request comes nowhere near the pool ceiling and the failures that stopped the native-context profile cannot occur. That is not an argument, it is the reason this recipe was measured separately: it starts healthy on the promoted image with --enable-prefix-caching --mamba-cache-mode align, answers a text and an image request exactly, and reports enable_prefix_caching: True in the engine banner (receipts/production-image.json). The context edition's native 262,144-token recipe does not enable it, and its card explains why (receipts/qualification-5090-apc.json).

   The release unit moved on 2026-08-16, and #51113 is why. Upstream vLLM #51113 (mamba align prefill-chunk splitting: a chunk that ends mid-block leaves its slot holding a short state, which a later chunk then publishes anyway — wrong tokens, HTTP 200, no crash) merged 2026-08-06, after the pinned public image was built, and is still absent from it and from fork head fa033bd4e. Cherry-picks were requested upstream on 2026-08-16 (issue #392, PR #393). Until they land it is carried as tools/vllm-mamba-align-scheduler.py (sha256 b431c106…), and the image this project serves is now the four-module localhost/vllm:gg-r34-patched-apc, manifest sha256:16a936b877b90f…, promoted from the three-module localhost/vllm:gg-r34-patched (sha256:6eca4c693f01b6…) (receipts/production-image.json). With prefix caching off the two images are not merely similar but behaviourally identical, because the added module's changed function is unreachable unless mamba_cache_mode is align — so the earlier hardware qualification carries over unchanged. One warning if you inspect the image yourself: its build-time label io.malaiwah.image.qualified still reads false and is superseded by the receipts named here — it was written before the image could possibly have been qualified, and correcting it would add a layer and change the very digest that was measured. That digest is local to the build host, so the recipes here reproduce its content with sha256-verified read-only mounts over the pullable public base.

   What prefix caching buys. On disjoint documents, so the cold case is genuinely cold: a 32,842-token prefix went 12.07 s cold → 1.04 s warm (11.6×, 2,442 of 32,842 prompt tokens recomputed, 92.6 % hit rate) and a 131,146-token prefix went 67.60 s → 2.31 s (29.3×, 3,146 of 131,146 recomputed, 97.6 %); a 38-request schedule ran 84.0 s with the cache against 144.4 s without (receipts/apc-poison-repro.json). And what we actually know about its safety. Correctness was probed adversarially before any of this shipped: seven freshly started servers, 38 requests each, 266 scored requests, nested token prefixes so later requests hit blocks published by earlier ones, and no prompt length a multiple of the measured 1,600-token mamba block, so prefill chunks end mid-block by construction — zero corrupted responses, zero wrong answers, zero acceptance collapses, on the unpatched image as well as the patched one. Thresholds were committed before the first server started; the worst repeated block was 15 characters against an 80-character threshold, with no U+FFFD anywhere. Greedy chosen-logprob drift with the cache on is 0.1063 mean absolute against a measured run-to-run floor of 0.0823 — drift, never an answer change. So the module is carried as insurance backed by upstream's own regression file — 14 failed / 6 passed against the vendored scheduler, 20 passed against this one — and not by a reproduction of our own: we tried hard to reproduce the reported corruption and could not (receipts/mamba-align-defect.json).

   LMCache is unmeasured by us. It is not part of any recipe on this card, this project has never run it, and it is the outstanding suspect in the one user report of prefix-cache corruption we have. Nothing here says LMCache is safe; the evidence above covers vLLM's own prefix cache and nothing else.

   #51812 is now recommended for this recipe, and it stays an overlay. Upstream #51812 (Qwen GDN speculative gate ordering: the vendored code gathers the speculative Q/K/V rows but hands the recurrent update the ungathered a/b gate tensors, so gate row i can belong to a different token than Q/K/V row i) merged 2026-08-11 and is absent from the promoted image. Whether that path is ever entered is no longer an argument: a CPU-only counter mounted over the engine's GDN metadata builder measured it, and in the flag set Recipe B ships — window 8,192, --gpu-memory-utilization 0.85, --enable-prefix-caching --mamba-cache-mode align, MTP-3, at eight concurrent streams (the arm additionally pinned fp8 KV and --max-num-batched-tokens 2048) — it was entered: three events in 5,825 metadata builds, 0.515 per thousand, over 468 requests with zero errors and a prefix-cache hit rate reaching 50.1 % (receipts/gdn-gate-concurrency.json). That arm served the context edition's weights: the counter reads only the scheduler's host-side arrays, so what it measures is a function of flags and traffic, with the checkpoint entering only through the size of the KV pool. The mechanism, in one sentence you can act on: at eight streams with prefix caching on, a short non-speculative request can land between speculative ones in the same batch, and the unpatched gather then misaligns the gates. All three firings had one shape — six speculative decodes plus one non-speculative request, that request beginning at token index 20 and displacing one four-token speculative decode, whose four gate rows were read from the wrong tokens. The same instrument saw zero events in 3,329 builds at eight streams with prefix caching off (below 0.90 per thousand, 95 % upper bound) and zero in 8,065 builds at a quarter of the token budget, so it is the cache path that opens this, not concurrency alone.

   So mount it if you serve Recipe B with more than one sequence in flight: tools/vllm-qwen-gdn-spec-gates.py (sha256 7cd3f5fe763b621048af4817951a841d99c8b700d9a56ded27ccaca5a56ccbe0) read-only over /opt/venv/lib/python3.12/site-packages/vllm/model_executor/layers/mamba/gdn/qwen_gdn_linear_attn.py. It is diff-identical to upstream's eight changed lines and py_compile-clean under the image's Python 3.12.3. It is an overlay and deliberately not part of the qualified digest: sha256:16a936b877b90f… is what was qualified, and a reachability count is not evidence that would survive a re-qualification, so it is mounted over the vendored file rather than promoted into the image. At --max-num-seqs 1 no mixed batch can form at all and none of this applies.

   The effect on answers was not measured, and measuring it was declined on resolution grounds. Three events in 5,825 builds cannot move a statistic whose run-to-run floor is 0.0823 mean absolute chosen-logprob error against a per-event effect of 0.002755 — a noise floor about thirty times the size of one event — so an A/B would have returned its own noise and was deliberately not run. Nothing here claims the overlay changes an answer, or that it does not. The recommendation follows a rule fixed before the GPU window opened instead: a nonzero rate in a regime we ship gets the free fix, because a silently miscomputed forward pass gives the operator no signal at all. Traffic in that arm was adversarial by design — eight speculative streams plus injected short prompts and full-cache-hit repeats — so 0.515 per thousand is an upper bound on a shipped regime, not a forecast for your workload (receipts/gdn-spec-gate-defect.json is the source-level defect analysis).

   Scope on this build. Recipe B above is smoke-proven with prefix caching on — it starts and answers a text and an image request exactly on the promoted image (receipts/production-image.json) — but the nine-gate qualification was run on the native-context profile, not on this 8,192-token one. Recipe A stays prefix-caching-off and must: it is the unmodified public image, which predates #51113.

KV-cache dtype: fp8 is the family's measured default

The recipes on this card leave --kv-cache-dtype unset; the family's qualified long-context profiles pin fp8, and that default is now measured rather than assumed. A five-arm sweep on the physical RTX 5090 served the context edition — same engine, same flag surface — at its native-window profile with the KV dtype the only deliberate flag change: no arm beat fp8 on native-or-beyond context on 32 GB with retrieval intact (receipts/kv-dtype-sweep-5090.json, decision record docs/38-kv-dtype-sweep.md). The engine derives the attention backend from the KV dtype, so each arm is measured as it actually serves:

The per-token-head family is measured, not assumed: int8 and fp8 per-token-head each dominate fp8 on both capacity and closeness to the bfloat16-KV reference, but each pays 3.0× prefill because TRITON_ATTN is the only backend on this fork that accepts per-token-head scales — and the capacity edge is TRITON_ATTN's smaller CUDA-graph pool (0.06 against 0.45 GiB), not cheaper bytes: those arms cost more per token than fp8 (35,360 against 34,816 B/token). int4_per_token_head is the real capacity lever — 502,667 tokens, 1.92× concurrency — at two named prices: 3.6× fp8's distributional error and 2.78× its prefill. nvfp4 and nvfp4_ds_mla do not start: no attention backend on this fork advertises nvfp4 for a non-MLA decoder — all five candidates answer kv_cache_dtype not supported — and the GLM-5.2-serves-nvfp4 precedent is the owner's claim about a different model, unverified here. The fidelity column is a bfloat16-KV-reference probe at a 98,304-token context — truncated top-20 KL over 70–173 paired greedy positions, a lower bound; it is not the v5 KLD and must never be differenced against any published KLD figure. Retrieval was 44/44 exact across the five arms, 4-bit included — retrieval is not fidelity, which is exactly why the KL column exists.

Reconstruct-scratch arena: +17,874 KV tokens, and it stays an overlay

A 2-hunk fork patch to exl3.py buys 17,874 more KV tokens, measured, and it is an opt-in overlay rather than part of any qualified digest. The pinned r34 image keeps one persistent fp16 prefill-reconstruct scratch per weight geometry — 790 MiB across the eight geometries that allocate at the qualified long-context profile (the head's 5120×32768 chunk needs ≥128 sampled logit rows and never triggered). The patch (overlay tools/vllm-exl3-scratch-arena.py, sha256 9aba06ebf60ca7665c0513752387c349240ab85e1ebc44d6ce8137ef157b6c15; fork PR local-inference-lab/vllm#397) shares one grow-to-max arena per device instead, sized by the largest live geometry (170 MiB), because each reconstruct is written and consumed inside one eager call on one stream. The kernels see identical operands — same shapes, strides and dtypes.

Measured on the physical RTX 5090 A/B: engine-reported KV pool 265,122 → 282,996 tokens (+17,874, +6.7 %, 9.28 → 9.88 GiB ≈ +0.60 GiB), reproduced identically across two server starts per arm, the arena's own growth log ending at exactly the predicted 170 MiB (receipts/scratch-arena.json). That A/B served the context edition at its qualified 262,144-token profile — same engine, same flag surface — and the scratch geometries are shape-derived, so the mechanism and the 170 MiB arena size are identical on this build while the pool figures above are the context edition's. Fidelity was gated rather than assumed: the 30-case deterministic vision suite returned byte-identical answers on both arms (24/30 each, equal to the rank-1 qualification reference), a full-window needle (258,925 tokens, depth 0.5) retrieved exactly, and decode did not regress (109.2–109.7 against 108.5–108.9 tok/s over three warmed C1 runs). Read the byte-identity claim narrowly — it covers that deterministic probe set, because the control shows two restarts of the unpatched baseline differ on 7 of 8 long greedy continuations (exl3_gemm autotunes kernel configs by measured time per process), so this stack is not restart-deterministic on long greedy text with or without the patch, and every cross-restart pairing is 7 DIFF / 1 MATCH either way.

Like #51812, it is an overlay deliberately not part of the qualified digest: the pinned digest is what was qualified, a larger KV pool is not evidence that would survive a re-qualification, so it is mounted read-only over the vendored file rather than promoted into the image — -v tools/vllm-exl3-scratch-arena.py:/opt/venv/lib/python3.12/site-packages/vllm/model_executor/layers/quantization/exl3.py:ro. The static prediction had been +620 MiB / +18.7k tokens; the measured gain is 95.7 % of it, and the measured number is the one to quote. On the 24 GB class the same bytes put the published 24,576-token window at 42,450 raw token headroom, supporting 40,960 at the next window step — arithmetic only, pending a 24 GB-class boot (docs/34-vram-class-profiles.md §10.2).

Client contract, inherited unchanged from upstream

Chat template, tokenizer, preprocessors, generation defaults and vocabulary (248,320, untied head) are byte-identical to upstream. generation_config.json: temperature 1.0, top_p 0.95, top_k 20. Qwen's recommendation for non-thinking mode is temperature 0.7, top_p 0.8, top_k 20, presence_penalty 1.5.

Thinking control is chat_template_kwargs: {"enable_thinking": false}, or {"reasoning_effort": "..."} where the only valid values are xhigh (default), medium and low — the upstream template raises on high.

Context: 262,144 native; verified here only to 8,192. Upstream's 1M procedure is static YaRN, not a bare max_position_embeddings bump: it needs nested rope_parameters with rope_type: yarn, factor: 4.0, original_max_position_embeddings: 262144, VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 and --max-model-len 1000000, and Qwen warns it costs short-context quality. Untested on this runtime.

Optional: narrow the input embedding table

VLLM_EXL3_EMBED_BITS=8 converts the 248,320 x 5,120 input table to per-row int8 after load (2.543 GB -> 1.272 GB, docs/32). At this family's published 8,192-token profile it moves about 1.2 GiB of weights into KV, measured on the physical card: hydrated 66.45 -> 67.62 GiB of KV = 668,852 -> 680,899 tokens (+1.80 %, concurrency 81.65x -> 83.12x); online K5/K6 345,324 -> 352,571 tokens (+2.10 %, 42.15x -> 43.04x). Correctness is unaffected on the frozen probe schedule - 22/22 planted answers and zero corruption detectors on both arms - but output is not identical: with greedy decoding 13/22 (hydrated) and 10/22 (online K5/K6) continuations diverge from the BF16-table arm tens of tokens in, which is the served consequence of the +0.000065 mean KLD already published in docs/32.

Recommended only where KV is the binding constraint - long windows, smaller cards, the context edition it was built for. At 8,192 tokens, where the KV pool is already 42x-82x the window, the headroom it buys is not worth changing the text a prompt returns (receipts/embed-overlay-8k.json).

Chat template

This repo ships chat_template.jinja byte-identical to Qwen/Qwen3.8-27B (sha256 c3cf9e34abf4f9e36c2d72165aa9c132d3e2a725b6c2586aaa3a8af9d7a81041), and the same bytes again in the chat_template key of tokenizer_config.json. Do not replace one without the other: under transformers 5.15.0 the .jinja file takes priority, so editing only tokenizer_config.json is a no-op. To override, pass --chat-template <file>.

Three upstream-template restrictions to code against. All three are Qwen's, unchanged by us, and all three surface as HTTP 400:

• reasoning_effort accepts only xhigh (default), medium and low - measured against a live endpoint, correcting an earlier revision of this card that also listed none. The server answers HTTP 400 Unexpected reasoning effort none. Supported types are xhigh (default), medium, and low. for none, exactly as it does for high, minimal and max - all of which vLLM's OpenAI surface advertises and this template rejects. If your client hard-codes high, serve with --default-chat-template-kwargs.reasoning_effort=xhigh.

• What the ladder costs, measured rather than described. Same prompt, temperature 0, 32,768-token budget so no arm is truncated, two repeats each, on the qualified 1x endpoint (receipts/reasoning-effort-1x.json):

  reasoning_effort	completion tokens	reasoning chars	latency	finish	correct algorithm
  (unset)	18,085	56,955	168.9 s	stop	2/2
  xhigh (default)	18,085	56,955	168.8 s	stop	2/2
  medium	3,436	4,767	26.2 s	stop	2/2
  low	3,517	4,606	27.0 s	stop	2/2

  Three things follow. Unset is byte-identical to xhigh, which is how the default is now measured rather than assumed. xhigh costs 5.3x the output tokens and 6.4x the wall clock of medium for the same correct answer on this task - the single largest cost knob on the client side of this card, and the one to reach for first in any agent loop bounded by a per-task timeout rather than by quality. And low is not cheaper than medium (3,517 against 3,436 tokens): the bottom of the ladder is flat, so medium is the useful floor. Read narrowly: one prompt, one shape, an exact-match check on whether the right algorithm is named - it establishes token cost, not a quality ordering.

• Exactly one system message, first. Two system messages 400 with System message must be at the beginning. (vllm#41114, open; fix PR vllm#44505 closed unmerged). Merge them client-side. A developer role is safe - vLLM folds and consolidates it (vllm#43590).

• Content blocks must be text, image, image_url or video. Anything else, including an Anthropic tool_result carrying a tool_reference item on /v1/messages, 400s with Unexpected item type in content. (vllm#52489, open).

• Only the leading system message is accepted. A second or mid-conversation system role raises. Clients must merge system instructions into the first system message before rendering history.

• Tool-call argument values are an unescaped XML boundary. Before replaying assistant history, reject values containing literal </parameter> or <parameter= delimiters. Treat replayed calls and tool output as untrusted data; the template cannot escape those sequences.

Echoing reasoning_content back on assistant history turns is what buys a full prefix-cache hit: measured 100 % prefix reuse when the client returns it, 94.7 % at ten turns when it does not. Community "fixed" Qwen templates are not recommended here: measured against --tool-call-parser qwen3_coder, qwen3.8-froggeric-v22 renders a tool call whose arguments the parser recovers as {}. Details in docs/39-chat-template-audit.md and receipts/chat-template-audit.json.

Verification performed, and not performed

Done: structural audit (1,199 logical tensors reconstructed, matching upstream), serving under the pinned image, greedy text, 96×96 and 2044×1622 image answers, MTP acceptance from server counters, 3-run throughput with <1 % dispersion, prefill at exact token counts, eager-vs-graph decode parity on real decode steps (24/32 exact sequences, with a BF16 control showing the same 24/32), the three fidelity suites above (v5 held-out at 10,480,640 scored positions, v4 qualification, v3 development), and 40/40 deterministic task-retention smoke with zero BF16 regressions.

Context length is the sharpest gap between what is claimed and what is tested. Fidelity and functional tests ran at --max-model-len 8192. The 262,144 rows above are engine allocation and startup at that length, not generation, retrieval or accuracy at length. No native-262K generation, no YaRN-1M run, and no long-context retrieval measurement exists yet.

Measured but not established: the paired MMLU-Pro run above has now covered this build and all five comparators, and this build is a measured shortfall against the pre-registered bar — 54/57 BF16-pass retention, Wilson 95 % lower bound 85.6 % against a required 90 %, which only the context edition clears. At 70 items the suite cannot certify that bar for any result that gives up more than one paired pass, and its intervals separate no two candidates, so there is still no wider public task evidence for this build: no executable-code, constraint-following, tool-schema or larger-draw result (the plan's own P1), and no GPQA/HumanEval-style run. Also not done: real OCR/chart/video evaluation, long-context retrieval or perplexity for this build, native-262K or YaRN-1M generation, multi-GPU/TP>1, non-SM120 hardware, and quant-specific safety testing. KLD and the small generated smoke do not establish broad task capability.

Safety and intended use

Upstream does not disclose training-corpus composition, knowledge cutoff, safety evaluation, or detailed intended-use limits, and this quant adds none. Quantization can alter refusal behaviour and calibration even when average divergence is low, and no quant-specific safety regression testing has been performed. Intended for research and local inference evaluation. Inherits Apache-2.0 from upstream.

Reproduce this

This section is about the numbers, not the bytes: a fresh conversion of the recipe produces a sibling rather than this checkpoint, as recorded beside the composition table at the top of this card.

Everything the v5 numbers on this card were computed from is published as a dataset: malaiwah/qwen38-27b-fidelity-suite-v5 — 5,835 files, 10,826,796,868 B (10.83 GB / 10.083 GiB), verified at revision 08bde6cc. It contains the 5,120 token-id files that are the authoritative evaluation input (retokenizing the source text does not reproduce them), the parent suite manifest whose sha256 equals the ladder pin, the ladder pin itself, all ten 512-context shard views with their capture and replay command lines, the corpus fetch log, the shard-0 unquantized BF16 hidden-state reference (512 captures plus manifest, 10.73 GB) and 79 per-shard reports — 50 ladder, 10 tail, 15 scored-window, 4 cross-engine. Receipt receipts/preserved-artifacts.json; 5,326 of the 5,835 files were re-downloaded and re-hashed end to end, and the 10 GB hidden-state tree was checked against the Hub's own LFS digests plus a three-file CDN spot check.

What it costs to replay. Because the suite and the shard-0 BF16 reference are both published, a third party can score a new candidate against the identical contexts without recapturing the reference: one candidate capture plus one replay, about 6 minutes of GPU for shard 0 on a single RTX PRO 6000 Blackwell, instead of two model loads. That is exactly how the NVFP4 row in Against GGUF was produced. Re-running the whole ten-shard ladder is a different bill — about 5 hours of GPU for the fifty candidate captures, plus about 54 minutes for the nine BF16 shard references that were deleted once their reports verified.

Seven archival mirrors keep the third-party citations resolvable — the three explained below, plus the four quant-landscape shortlist mirrors cited elsewhere on this card, each pinned to the exact upstream revision measured and each now carrying a provenance banner that names its upstream author. malaiwah/Qwen3.8-27B-NVFP4-archival-9c73e2da is a recovery mirror: upstream super-squashed its history on 2026-08-15 and the Hub now answers Invalid rev id for 9c73e2da…, the revision every NVFP4 number on this card was measured against, so the reviewed revision is otherwise unreachable. malaiwah/Qwen3.8-27B-GGUF-archival-f1bfb127 is precautionary: the five files the cross-engine table cites, at a revision that still resolves upstream. malaiwah/Qwen3.8-27B-NVFP4-RTX5090-archival-69274a0d is likewise precautionary: the gittensor checkpoint measured above, at a revision that still resolves upstream, deep-verified after upload — its 19.2 GB of weights cost essentially zero transfer because the Hub already held every chunk. Said plainly, a mirror preserves the citation — a resolvable repo id, revision and digest table that survive an upstream squash or delete — and is not independent byte-level redundancy. Hub storage is content-addressed, so our copy and upstream's plausibly reference the same underlying chunks; nobody should assume physical copies we do not hold. The measured cost of the first two mirrors was 2.34 GB of transfer for 149.3 GB of content, about 1.6 %, which is that content-addressing showing through.

Prior art and credits

• turboderp / turboderp-org/exllamav3 — the EXL3 Trellis format, LDLQ calibration, MCG codebook, and the conversion pipeline.
• Martin Vit / voipmonitor / festr2 (local-inference-lab) — the Gilded Gnosis vLLM fork, B12X Trellis kernel integration, the voipmonitor/vllm Docker images, the rtx6kpro serving wiki, and the exl3-b6 online-encoding overlay whose distribution-fidelity protocol this evaluation follows.
• Luke Alonso / lukealonso — B12X kernel co-author (PR #31, NF3/NVFP4) and the NVFP4 quantization methodology whose protection patterns informed the recipe design.
• Brandon Music / brandonmusic — the ShapleyMCG expert-bit allocation method, the EXL3 checkpoint patterns (TR3 schema, online K6, shared-expert handling) that this recipe family is modelled on, and the exllamav3 SM120 retile fork.
• Josh Cartu / jcartu — the MTP78 recipe and rank-sliced MTP runtime that informed malaiwah/GLM-5.2-EXL3-TR3-MTP78: the graft/encode/serve procedure for the MTP draft layer, plus the vLLM normalize_rank_sliced_weight_name() loader patch.
• Qwen — the base model (Qwen/Qwen3.8-27B, Apache-2.0) and its official FP8 derivative (Qwen/Qwen3.8-27B-FP8).
• nvidia/Qwen3.6-27B-NVFP4 — the role-aware recipe this design is modelled on (4-bit MLP, 8-bit attention, protected embeddings/vision/MTP).
• unsloth/Qwen3.8-27B-NVFP4 — independent confirmation of the same protection pattern.
• malaiwah / progressive-tensors — research, recipe, harness and receipts. Per-expert EXL3 provenance work, the per-bit error ladder, the GLM-5.2-EXL3-TR3-MTP78 checkpoint, and the RTX 5090 context-capacity qualification that caught a 262K overclaim.

Companion repository

Recipes, receipts, the fidelity harness, the response to an independent review, and the open items: https://github.com/malaiwah/qwen38-27b-exl3.

Upstream contributions from this work: #311 / PR #312 (overlay fallback crash), PR #314 (CUDA-graph decode), #313 (Qwen3.8 vision truncation).