Qwen3.6-27B MTPLX Optimized Speed
Run this with MTPLX
MTPLX is an MLX-native runtime for native Multi-Token-Prediction speculative decoding on Apple Silicon. Up to 2.24× faster decode at real coding temperatures (temp=0.6 / top_p=0.95 / top_k=20) using the model's own built-in MTP heads — no external drafter, no greedy hack.
brew install youssofal/mtplx/mtplx
mtplx start
Project: github.com/youssofal/MTPLX
Other MTPLX checkpoints:
- Qwen3.5-4B-MTPLX-Optimized-Speed — small 4-bit speed-test
- Qwen3.5-4B-Optimized-MTPLX — small 8-bit
This is the speed checkpoint for MTPLX v0.1.0-preview. It combines the
and embeds mtplx_runtime.json so MTPLX can run it as the optimized speed
variant.
Target repository: Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed
Runtime
mtplx pull Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed
mtplx run "hello" --model Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed
mtplx serve --model Youssofal/Qwen3.6-27B-MTPLX-Optimized-Speed --port 8000 --no-stats-footer
MTPLX v0.1.0-preview supports Qwen3-Next-MTP only. This model is verified for
the qwen3-next-mtp backend.
Sources And Attribution
| Component | Source | Revision | License |
|---|---|---|---|
| Base model | Qwen/Qwen3.6-27B | 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9 | Apache-2.0 |
| MTPLX conversion and runtime contract | youssofal/mtplx | v0.1.0-preview.1 | Apache-2.0 |
The trunk is the MLXCommunity flat 4-bit affine conversion of compressed-tensors checkpoint at the revision listed above.
Verification
mtplx_runtime.json records:
- architecture:
qwen3-next-mtp - maximum MTP depth:
3 - recommended profile:
sustained - recommended draft-only LM head:
3-bit affine, group_size=64 - recommended draft sampler: temperature
0.70, top-p0.95, top-k20 - exactness gate:
Phase 0H paged-verifier smoke - exactness max absolute diff:
0.0 - verified hardware:
Apple M5 Max, 128 GB unified memory - verified timestamp:
2026-05-03T23:07:00+0100
Target sampler used for the recorded contract: temperature 0.6,
top-p 0.95, top-k 20, with thinking mode disabled. The draft proposal
sampler is intentionally slightly hotter at temperature 0.70; target
verification and residual correction remain exact.
Performance Honesty
This is the default speed lane. On the local Apple M5 Max fanmax
sustained benchmark, this artifact reached 63.056 tok/s and
62.886 tok/s at depth 3 on the long-code 192-token prompt when using its
contract-recommended 3-bit draft-only LM head plus draft sampler temperature
0.70, with thinking mode disabled and foreground app load reduced. The
proposal change reduced the measured correction tokens from 6 to 3 and verify
calls from 50 to 49 on the recorded prompt; target sampling remains temperature
0.6 / top-p 0.95 / top-k 20. A full greedy diagnostic on the same cleaned
fanmax window measured 60.108 tok/s. Earlier contract evidence measured
61.527 and 60.900 tok/s, the earlier 3-bit draft-head contract measured 60.038
and 60.061 tok/s, and the older 4-bit draft-only LM head measured 57.668 tok/s
on the same lane.
Files
model-*.safetensors: MLXCommunity flat 4-bit trunk shards.mtp/weights.safetensors: calibrated prequantized INT4 MTP sidecar, nested so generic MLX vision loaders only glob the base model shards.mtplx_runtime.json: MTPLX verified runtime contract.MTPLX_PUBLISH_MANIFEST.json: local staging manifest with source paths, materialization method, sizes, and optional hashes.
Links
- MTPLX: github.com/youssofal/MTPLX ·
pip install mtplx - Base model: Qwen/Qwen3.6-27B