README
L0xRE-27b-Low-MTP
The self-drafting variant of L0xRE-27b-Low: the same 27B hybrid model and quantization, plus one built-in nextn/MTP prediction block (+830 MB). With this file the model can speculative-decode on its own — no separate drafter file needed.
Use this file if you want a single-file deployment with self-accelerated
decoding (--spec-type draft-mtp).
Use the standard file instead if you plan to attach a DFlash2 drafter — with a
drafter present the MTP block is idle, and the standard build is 830 MB smaller
with identical speed and quality.
Quickstart (SM89, self-drafting)
llama-server -m L0xRE-27b-Low-MTP.gguf \
-c 32768 -b 2048 -ub 1024 -np 1 -t 8 -ngl 99 -fa on \
-ctk kvarn3 -ctv kvarn2 \
--spec-type draft-mtp --spec-draft-n-max 2 \
--jinja --reasoning on --reasoning-effort medium
Requires the L0xRE runtime (branch release/l0xre-sm89-v0.4.7 of
L0xRE-BeeLLama-Low),
not stock llama.cpp.
File identity
| File | Bytes | SHA-256 |
|---|---|---|
L0xRE-27b-Low-MTP.gguf | 9,468,579,616 | 746bd40841fb18b9c1918923e89c007df70b5e4f3de64290d1399593e70c96b0 |
Architecture: 27B-parameter hybrid on the Qwen3.8-27B base — 64 transformer blocks plus one nextn/MTP block; low-bit (≈2–3 bpw) projections, higher-precision embeddings/output, 262,144 native context, 248,320 vocab.
Measured (RTX 4090, L0xRE SM89 runtime)
Paired with DFlash2-Q4 for a like-for-like comparison against the standard build:
143.0 ± 5.7 tok/s code decode (peak 150.8), acceptance 0.72, mean draft length 4.6.
MTP-standalone (draft-mtp) qualification is in progress — the numbers above
were taken with the DFlash2 drafter, not the built-in head.
Qualification boundaries
- Qualified on the L0xRE BeeLLama runtime only — not stock llama.cpp, Ollama, LM Studio, Transformers, or SGLang.
- MTP-standalone serving, longer contexts, and multi-slot serving are not part of the qualification.
Quality
Same weights as the standard build — same internal-suite results: 124–126 / 150 pass@1 with thinking on (three passes) vs 117 / 150 native-quant baseline. Internal, non-canonical suite; directional, not leaderboard.
License
Quantized derivative of Qwen3.8-27B, distributed under the base model's research/community terms (LICENSE); L0xRE runtime tooling is Apache-2.0. Full notices: seanyourhighness/L0xRE.
Similar repos
README
L0xRE-27b-Low-MTP
The self-drafting variant of L0xRE-27b-Low: the same 27B hybrid model and quantization, plus one built-in nextn/MTP prediction block (+830 MB). With this file the model can speculative-decode on its own — no separate drafter file needed.
Use this file if you want a single-file deployment with self-accelerated
decoding (--spec-type draft-mtp).
Use the standard file instead if you plan to attach a DFlash2 drafter — with a
drafter present the MTP block is idle, and the standard build is 830 MB smaller
with identical speed and quality.
Quickstart (SM89, self-drafting)
llama-server -m L0xRE-27b-Low-MTP.gguf \
-c 32768 -b 2048 -ub 1024 -np 1 -t 8 -ngl 99 -fa on \
-ctk kvarn3 -ctv kvarn2 \
--spec-type draft-mtp --spec-draft-n-max 2 \
--jinja --reasoning on --reasoning-effort medium
Requires the L0xRE runtime (branch release/l0xre-sm89-v0.4.7 of
L0xRE-BeeLLama-Low),
not stock llama.cpp.
File identity
| File | Bytes | SHA-256 |
|---|---|---|
L0xRE-27b-Low-MTP.gguf | 9,468,579,616 | 746bd40841fb18b9c1918923e89c007df70b5e4f3de64290d1399593e70c96b0 |
Architecture: 27B-parameter hybrid on the Qwen3.8-27B base — 64 transformer blocks plus one nextn/MTP block; low-bit (≈2–3 bpw) projections, higher-precision embeddings/output, 262,144 native context, 248,320 vocab.
Measured (RTX 4090, L0xRE SM89 runtime)
Paired with DFlash2-Q4 for a like-for-like comparison against the standard build:
143.0 ± 5.7 tok/s code decode (peak 150.8), acceptance 0.72, mean draft length 4.6.
MTP-standalone (draft-mtp) qualification is in progress — the numbers above
were taken with the DFlash2 drafter, not the built-in head.
Qualification boundaries
- Qualified on the L0xRE BeeLLama runtime only — not stock llama.cpp, Ollama, LM Studio, Transformers, or SGLang.
- MTP-standalone serving, longer contexts, and multi-slot serving are not part of the qualification.
Quality
Same weights as the standard build — same internal-suite results: 124–126 / 150 pass@1 with thinking on (three passes) vs 117 / 150 native-quant baseline. Internal, non-canonical suite; directional, not leaderboard.
License
Quantized derivative of Qwen3.8-27B, distributed under the base model's research/community terms (LICENSE); L0xRE runtime tooling is Apache-2.0. Full notices: seanyourhighness/L0xRE.