Motif 3 — Mixed-Quant GGUF

A structurally intact 314.8B-parameter Motif-3 MoE compressed to 94.16 GB (87.70 GiB). Native H200 residency, latent-KV execution, OpenAI serving, tools, continuous batching, and retrieval through 128K are validated. The ds4-dfm path also passes a strict 262,080-token prompt plus decode gate on one DGX Spark.

Support my work

I work on making large language models practical on hardware they were never really designed to fit on — through mixed quantization, inference optimization, custom kernels, and serving experiments.

While much of the development happens on local hardware, calibration, profiling, and large-scale validation often require expensive on-demand GPUs.

Contributions help pay for that compute, storage, and testing infrastructure, so I can keep experimenting and publishing the results openly.

Full-topology mixed-quant conversion of Motif-Technologies/Motif-3, designed as the resident-weight baseline for one 128 GB-class DGX Spark. This is an independent conversion, not an official Motif Technologies release.

Nothing is pruned, merged, expert-dropped, or layer-dropped. The artifact retains all 53 target-model layers, the first two dense layers, all 51 sparse layers, all 384 routed experts per sparse layer with top-8 routing, the shared expert, Grouped Differential Latent Attention (GDLA), Expert-Specific PolyNorm, modified mHC, and the complete one-layer MTP predictor.

ds4-dfm

This model is served by ds4-dfm, a common ds4 release line for DFM (독자 파운데이션 모델, 독파모) model families. It was developed to serve full-scale models on a single NVIDIA DGX Spark with 128 GB of unified memory through explicit model-family C/CUDA paths and device kernels.

The release line is the dfm branch of Baekpica/ds4. The integration tag is v0.6.2-dfm (Entrpi v0.6.2 absorbed). The 32K/256K HTTP gates below remain the earlier Spark evidence at 593d251; the original 8K ds4-bench point uses cc2f277; the persistent multi-bank measurements use cf605e0. The 8K/32K remesure uses a09ff4f on dfm. The 256K remesure uses 6500480 (kernels through a09ff4f). The integration tag is not moved. One ds4-server command and one HTTP contract cover the integrated DeepSeek, Solar Open2, K-EXAONE, Motif-3, and dots3-note families; only the GGUF path and its matching weight-owner manifest change. The server provides OpenAI Chat Completions, OpenAI Completions, OpenAI Responses, and Anthropic Messages.

The production GGUF in this repository was loaded through the common release binary on DGX Spark. Context and throughput claims are limited to the exact conditions below; model quality has not been evaluated by these serving tests.

Artifact

The canonical merged/unsharded byte stream used by native ds4 is 94,162,541,472 bytes (94.16 GB / 87.6957 GiB) with SHA-256 15755a735753bc1396e5ffa539e65a779a4fd769e8833360a4d743c4c60c2f25.

Start with Motif-3-MQ87-88-FIT-00001-of-00011.gguf; compatible split-aware runtimes discover the remaining shards automatically. Exact per-shard hashes are published in MQ87-88-FIT-SHA256SUMS.

The weight files are fixed at Hub revision efd6044e25e7f8e3b459a737d021091e2e69b6c6. Later model-card/report commits do not change shard bytes or hashes.

hf download Baekpica/Motif-3-Mixed-Quant-GGUF \
  --revision efd6044e25e7f8e3b459a737d021091e2e69b6c6 \
  --include 'Motif-3-MQ87-88-FIT-*.gguf' \
  --include MQ87-88-FIT-SHA256SUMS \
  --local-dir ./Motif-3-MQ87-88-FIT

The name is a nominal recipe label, not the artifact's decimal size. The actual payload is 94.16 GB (87.70 GiB), and the exact byte counts above are authoritative. Admission depends on whether its measured aligned/repacked resident form fits the target Spark together with latent KV, workspace, and server state.

Precision recipe

Precision is assigned by module role. Always-active and control paths stay conservative; the 51×384 routed-expert stack carries the compression burden.

The completed GGUF contains 2,287 tensors: 1,273 F32, 318 BF16, 543 Q8_0, 102 IQ2_XXS, and 51 Q2_K. The 102 IQ2 tensors are separately stored gate and up matrices. They originate from 51 fused checkpoint gate/up tensors; the split is lossless before quantization and retains every expert value.

No Q4 edge-layer promotion is included in this release. MQ95/MQ97 are outside the scope of this capacity-first artifact.

Q8_0 activation calibration

Calibration was collected only after producing and freezing the separate full-topology Q8_0 reference in Baekpica/Motif-3-GGUF.

The corpus reuses the checksum-pinned normalization and target shares audited for the preceding Solar Open 2 mixed-quant release, but every record is rendered and counted again with Motif-3's official final tokenizer and chat template. It contains 1,539 documents, 4,079,555 official-tokenizer tokens, and 17,011,959 bytes.

The H200 calibration runtime directly dequantized the Q8_0 GGUF and executed the official Motif-3 equations with a layer-major schedule: mHC, expanded historical K/V GDLA, interleaved 128-token SWA/full attention, YaRN, the differential signal/noise heads, elementwise attention output gate, sigmoid top-8 routing with correction-bias selection and route renormalization, shared experts, and per-expert PolyNorm.

Gate/up importance observes the exact FFN-normalized input. Down importance observes the Expert-Specific PolyNorm output after route weighting, matching the ds4 routed-MoE execution order. No uniform, random, proxy-model, or source-BF16 activation matrix was substituted.

The machine-readable activation report is published here as Q8_0-IMATRIX-REPORT.json. Corpus construction is included in the public reproduction materials; the exact rendered corpus, final imatrix, and rank-local accumulators are preserved in the private Spark handoff.

Provenance

Only the official final Motif-3 checkpoint was used. Motif-3-Beta was not used as a source, calibration input, implementation oracle, or fallback.

Model structure and context memory

Layers whose index is divisible by four use full attention (14 layers); the other 39 layers use a bounded 128-token sliding window. The production target does not retain expanded historical K/V. Its persistent context state is latent KV plus the decoupled RoPE key for full-attention layers, bounded SWA ring state, and cache identity/position state.

On H200, creating a native 262,144-token session while the model remained resident allocated 4,236,751,872 bytes (3.946 GiB) of cache tensor payload and produced a 4,334,813,184-byte (4.037 GiB) CUDA free-memory delta including allocator overhead. This is a physical H200 measurement, not a Spark claim. Physical unified-memory residency and OS headroom on the target GB10 remain authoritative.

The higher of two native-sm_90 H200 model/runtime initialization repeats plus the 256K-session delta was 102,326,337,536 bytes (95.298828125 GiB). This supports the capacity design; it does not predict GB10 driver, allocator, or OS overhead. The conservative H200 component measurements place model/runtime at 91.262 GiB and the native 256K session at 4.037 GiB, inside the design preferences of 98 GiB and 5 GiB respectively; only the target GB10 can establish unified- memory admission and final MemAvailable.

Runtime compatibility

Motif 3 is not a Llama-family graph. A compatible runtime must implement its 384E sigmoid router, route normalization/scale, shared expert, Expert-Specific PolyNorm, modified mHC, GDLA/differential heads and output gate, interleaved SWA/full attention, YaRN, latent KV semantics, and MTP.

The target runtime is Baekpica/ds4 at the exact ds4-dfm Spark commit pinned above. The earlier feature/motif-3-model-loader revision is retained as H200 bring-up provenance. The private Spark handoff also carries an offline source snapshot and commit metadata. Stock GGUF runtimes should not be assumed to execute this architecture merely because they can parse the container.

The current native branch has an explicit Motif tensor binder and CUDA graph, production latent-KV/SWA-ring sessions, strict device-resident model loading, the official tokenizer/chat/reasoning/tool protocol, and an OpenAI-compatible ds4-server path. Motif sessions refuse streaming/offloaded weights and never fall through to the generic DeepSeek graph.

A clean build at the pinned Spark revision completed the CPU and cuda-spark targets, linked the five runtime programs, and produced only sm_121a CUDA code objects. The same revision passed the Motif loader, tokenizer, CUDA primitive, real-weight parity, resident-memory, and strict OpenAI gates below.

The public files use standard GGUF splitting. The current ds4 development loader consumes one merged GGUF, so merge from the first shard with llama-gguf-split --merge before launch. This is a one-time artifact assembly, not SSD weight streaming; production admission still requires the merged or repacked weights to be resident and forbids simultaneous physical residency of duplicate raw/repacked mappings.

Validation status

The unsharded artifact and the complete 11-file split set independently passed strict source-map validation. The verifier checked the pinned revision, architecture metadata, native 262,144-token context, all 53 target layers, all routed gate/up/down tensors for sparse layers 2–52, all 384 experts, the shared experts, and the complete MTP block. It found exactly 2,287 unique tensors with no missing, duplicate, unexpected, mistyped, misshaped, or out-of-bounds payloads. The split set totals 94,162,542,816 bytes (94.16 GB / 87.6957 GiB); every shard also has a published SHA-256 digest.

A separate 57-row numerical comparison sampled the embedding and LM head, GDLA/control paths, dense/shared paths, routed experts at layers 2, 26, and 52, both halves of the checkpoint's fused gate/up weights, and MTP. Protected F32/BF16 rows were source-exact. Minimum sampled cosine was 0.9999740 for Q8_0, 0.9417932 for IQ2_XXS, and 0.9580462 for Q2_K. The native ds4 Motif-3 binder also accepted the completed mixed artifact as the official-final 53-layer, 14-full/39-SWA, 384E top-8 topology with MTP present.

Two final explicitly rebuilt sm_90 runtime repeats copied the full 94.16 GB (87.70 GiB) image into one H200 in 9.560–12.070 seconds without SSD streaming or CPU weight offload. The measured CUDA free-memory delta for model and runtime initialization was 97,438,334,976–97,991,524,352 bytes; capacity accounting uses the higher result. Strict residency fails startup instead of silently using host-mapped weights.

Once optional CUDA preparation finishes, ds4 discards the raw GGUF tensor pages while retaining only metadata/tokenizer mapping. Measured GGUF mapping RSS fell from 91,955,608 kB to 9,416 kB and remained low through inference, so the raw file is not kept as a second steady physical weight image beside the CUDA-owned model copy.

The final all-sm_90 full-question 256K H200 attempt also measured 9,416 kB of GGUF mapping RSS during partial prefill with VmSwap: 0, after allocating the complete 262,144-token production latent cache.

The automated resident gate caps this mapping at 262,144 kB both after copy and after native graph/cache execution. Its final native-sm_90 H200 run measured 9,416 kB and 29,512–29,640 kB respectively.

The native expanded-path oracle and production latent path selected the same first token and all top-8 logits on the short fixture; full-logit cosine was 0.99490164. Direct/chunked cache replay produced cosine 1.0. The real mixed sparse-layer diagnostic measured Q2 down cosine 0.9996071 and final sparse-output cosine 0.9998363. A real-weight MTP diagnostic evaluated 19 teacher-forced rows with finite logits.

DGX Spark ds4-dfm evidence

These measurements use one DGX Spark GB10 with 128 GB unified memory, NVIDIA driver 610.43.02, CUDA 13.3, Linux 6.17, and the production 11-shard artifact merged to its canonical 94,162,541,472-byte stream. Generation was greedy, single-request, non-streaming, and used no speculative decoding or thinking. The default prefill chunk was 4,096 tokens.

The 256K request completed in 1,509.460 seconds with 262,123 total tokens. Because the request was non-streaming, it does not provide an independently measured network-visible TTFM.

The resident gate measured a 4,422,546,432-byte (4.119 GiB) latent-cache payload and a 10,418,716,672-byte (9.703 GiB) default-graph-plus-cache physical delta. Source GGUF RSS settled at 29,632 KiB; owner and worker both remained at VmSwap: 0. The weight owner held 7.00 GiB of raw tensors plus 86.07 GiB in 590 derived artifacts, with no duplicate model image or CPU weight offload. About 12 GiB remained available during the final gate.

Nsight Compute confirmed that the expanded attention kernel fell from 55.79 ms to 28.83 ms (48.3%), and the Motif group-5 QK-absorb kernel fell from 38.91 ms to 10.97 ms (71.8%). Loaded SM-clock samples were 2,398–2,411 MHz; the prior 611 MHz pinned-clock condition did not recur.

The table above contains single-request results. Concurrent 256K serving, other GB10 software configurations, other accelerators, and model quality are not claimed. The Motif ds4-dfm path uses plain decoding with --no-spec; MTP and DSpark support models remain DeepSeek-only.

DGX Spark v0.6.2-dfm remesure

Same GB10 host (driver 610.43.02, CUDA 13.3, sm_121a), same MQ87-88 artifact, aligned-Q8 VMM owner (--reserve-gb 24), greedy, thinking disabled. Engine: Baekpica/ds4 dfm @ 6500480 (kernels through a09ff4f). The integration tag remains v0.6.2-dfm; that tag is not moved. The historical table above stays the v0.5.6.3-dfm evidence.

Versus the merge-day v0.6.2-dfm baseline on this host (519.90 / 12.62 at 8K, 445.03 / 9.68 at 32K): 8K prefill +21%, 8K decode +19%, 32K prefill +23%, 32K decode +34%. Versus the v0.5.6.3-dfm published 256K row (175.61 / 2.52): prefill +35.9%, decode +137%. Concurrent 256K banks are not claimed. The 256K worker held 10,429 MiB with 4.119 GiB of latent KV; owner and worker VmSwap stayed 0; available memory stayed 11–12 GiB; SM clocks sampled 2,411–2,496 MHz.

Partial prefix reuse

Runtime Baekpica/ds4@cf817c7 extends the shared partial-prefix checkpoint pool (introduced for Solar in 042bcea) to the Motif-3 continuous lane. Motif-3 has no recurrent state, so a checkpoint stores only each sliding-window layer's 128-row window (39 layers, 5.48 MiB per slot, 32 demand-mapped slots); full-attention latent rows below the checkpoint copy directly from the retained source bank. Request boundaries are checkpoints, long prefills and decode retain roughly 24 stride-aligned ones across the configured context, and a divergent prompt restores the nearest checkpoint at or below the shared prefix and replays only the gap.

One controlled A/B run used the same GB10, this artifact, the VMM owner (--reserve-gb 16), a four-bank worker, -c 65536, 4,096-token prefill chunks, and greedy no-think streaming Chat requests. The control set DS4_SERVER_FORK_PARTIAL=0, which also skips checkpoint reservation; the treatment used the default.

All three responses were byte-identical between control and treatment. Treatment reported two partial-fork admits, 16,384 cached tokens, zero failed requests, and zero admission rejects. Source checkpoint capture added 65.7 ms (0.23%) in this single run. The engine fixture (test_motif3_batch --partial-only) additionally pinned request-boundary and periodic checkpoints and a two-segment ring-wrapping window restore, each matching its cold oracle's greedy token.

This is a bounded worker-local cache, not an unbounded radix tree. Token history and full-attention rows remain anchored to a retained bank, the pool is not serialized into disk-KV payloads, and the EXAONE bank lane keeps exact-frontier reuse only. Details: docs/motif3-partial-reuse-2026-08-22.md.

DGX Spark 196K persistent multi-bank evidence

The current runtime was launched with a 196,608-token context, three resident banks, an 8,192-token prefill chunk, and no speculative decoding. It completed all four common endpoints--OpenAI Chat Completions, OpenAI Completions, OpenAI Responses, and Anthropic Messages--with their native HTTP 200 response shapes.

Settled /v1/stats showed zero failed or serial requests, zero continuous-batch failures, three total and zero live banks, and zero speculative drafts. The VMM weight owner used 90,119 MiB, the worker used 22,283 MiB after the 8K checks, and about 6.5 GiB of system memory remained available without an OOM event. Loaded SM clock remained 2,411 MHz instead of the earlier faulty 611 MHz pin.

H200 end-to-end evidence

[!CAUTION] These H200 figures are historical correctness-first bring-up measurements. The optimized Spark results above are separate and do not retroactively establish an H200 256K pass.

The isolated legacy-trim 256K row predates the final explicit sm_90 rebuild and executes through the CUDA toolkit-compatible default code object, with MMQ already at sm_90. Its rate is partial bring-up data, not a native-sm_90 performance claim. The final overlay has every CUDA code object verified as sm_90, passes the full resident graph/cache regression, and passes the native-sm_90 32K API plus native 32K/64K/128K rows shown above. The full-question attempt used that all-sm_90 binary. At the user's direction, both 256K attempts were stopped before decode and the remaining execution and optimization were transferred to Spark. Precision, topology, and context were not reduced to improve these figures.

The isolated 256K process also uses a legacy decode-reservation constant: it keeps the final 20 tokens and omits the five-token prefix QUESTION: Return only a, while retaining the full JSON/order instruction. Final ds4 revision d878ea1 corrects the tail to 25 tokens. The handoff's separately hash-pinned 262,080-token OpenAI fixture removes only filler and preserves the complete question; that corrected server gate remains mandatory on GB10.

The OpenAI server also completed a structured get_weather tool-call/result loop. Its no-thinking continuation reused the full 165-token live prefix and evaluated only the 51-token tool-result/new-assistant suffix. Two simultaneous deterministic requests on two resident sessions both returned the expected output, exercising continuous batching and session isolation.

Machine-readable structural and numerical reports are included as MQ87-88-FIT-VERIFY.json and MQ87-88-FIT-SAMPLE-VERIFY.json. The human-readable artifact and host records are included as MIXED-QUANT.md and H200-DEVELOPMENT.md.

H200 development has validated the pinned source inventory, official router, PolyNorm, mHC, tokenizer/chat/tool handling, expanded and latent GDLA paths, the full Q8_0 GGUF, the 302,080-token Q8 activation-collection pass, strict residency, short/long native generation, and the OpenAI server path above.

This card claims the completed strict single-DGX-Spark 262,080-token-prompt plus decode gate and the separate three-bank 196K-context serving gate only under the exact conditions above. It does not claim concurrent 256K serving, equivalent performance on other hardware or software configurations, or downstream task quality.

Limitations

• This is a hardware-oriented, very-low-bit routed-expert quantization. Quality must be evaluated for the intended languages, reasoning, code, tools, and long-context workloads.
• The public GGUF is a weight artifact, not a guarantee that an unrelated runtime implements Motif-3 correctly.
• The 262,144-token metadata comes from the source architecture; the measured validation in this card applies only to the pinned ds4-dfm configuration.
• DGX Spark uses coherent unified memory; conventional host-RAM plus discrete VRAM accounting describes a different deployment.

Acknowledgements

• Motif Technologies — the Motif-3 model and official implementation.
• antirez/ds4 — the original engine, loader, server, session machinery, and routed-MoE foundation.
• Entrpi/ds4-on-spark — DGX Spark CUDA and unified-memory groundwork used by downstream ds4 development.
• ggml-org/llama.cpp — GGUF and the quantization formats used by this artifact.

Errors in this conversion, runtime port, calibration, or measurements are ours, not theirs.

License and attribution

The source model identifies its license as MIT. See the official Motif-Technologies/Motif-3 model card for intended use, evaluation, citation, and license context.