mlx-community/Qwen3.8-27B-8bit plus the model's native multi-token-prediction head, so MLX runtimes with native MTP speculative decoding (oMLX mtp_enabled, mlx-lm PR 990 lineage) decode ~2.5–3× faster. Weights are otherwise byte-identical to the mlx-community conversion.

	decode, M5 Max 128 GB, single stream
mlx-community 8-bit (no head)	plain autoregressive
this repo, oMLX mtp_enabled	35.9 tok/s

Full A/B and method: MTP.md · code: bluehawana/Qwen3.827B-SGLang-mpbm5max — Track D.

What was added

model-mtp.safetensors — the 15 language_model.mtp.* tensors (quantized 8-bit/gs64 like the body), taken from EigenLabs/Qwen3.8-27B-MTP-bf16 (bit-identical to Qwen/Qwen3.8-27B's own head; norms already in MLX +1 convention), registered in model.safetensors.index.json; text_config.mtp_num_hidden_layers = 1. The mlx-community conversions declare the head in config.json but strip its weights, which silently disables MTP.

Use with oMLX

Download into your oMLX model dir (or HF cache).
Admin UI → model settings → Lightning MTP on (API: PUT /admin/api/models/<id>/settings {"mtp_enabled": true}), load. The log shows Speculative backend selected and per-request MTP[...] accept=… lines.
Optional for Qwen3.8: set chat_template_kwargs.reasoning_effort to medium/low — the template defaults to xhigh.

Draft acceptance on coding/agent traffic: 75–99 %, ~3 tokens per verify pass (depth 3). MTP is single-stream; concurrent requests fall back to plain batched decode.

Size: ~29.3 GB. Credits: Qwen team (model, Apache-2.0), mlx-community (quantization), EigenLabs / Layr-Labs MTP challenge (head export), oMLX (runtime).

bluehawana/Qwen3.8-27B-8bit-MTP-MLX

Qwen3.8-27B 8-bit + MTP head (MLX)

What was added

Use with oMLX

bluehawana/
Qwen3.8-27B-8bit-MTP-MLX