KaedeTai/Qwen3.8-27B-MTP-zh

Qwen3.8-27B MTP head — Chinese-tuned

A Multi-Token-Prediction (MTP) draft head for Qwen3.8-27B, fine-tuned on Chinese text so that speculative decoding accepts more of its drafts on Chinese output. Only the head was trained; the 27B trunk is untouched.

Because MTP drafts are always verified by the trunk (verify-then-accept), a different head can only change speed, never the model's output.

Files

File	What it is
qwen3.8-27b-mtp-zh-4bit.safetensors	Deployed head, quantized to match a 4-bit MLX trunk. Drop-in replacement for the stock head in KaedeTai/mlx-mtp-graft (qwen3.8-27b-mtp-4bit.safetensors).
fp16/mtp_ce.safetensors, fp16/mtp_ov.safetensors	The two fp16 training outputs (out_ce, out_ov runs) before quantization.

Used locally with orcarouter/Qwen3.8-27B-Uncensored-MLX (4-bit).

How it was trained

Hard self-distillation: the frozen 4-bit trunk is run forward on Chinese text, and the head is trained to predict the trunk's own argmax (not the corpus token). The goal is to imitate this specific trunk, which is exactly what speculative-decoding acceptance measures. 512-token blocks, gradient accumulation 4, lr 1e-5.

Draft acceptance rate (accepted / drafted tokens)

Text	Stock head	Chinese-tuned head
zh_web	56.8%	62.6%
zh_news	56.5%	62.4%
english	70.0%	69.6%
code	81.5%	81.8%

Chinese acceptance rises about 6 points; English and code are essentially unchanged.

中文說明

Qwen3.8-27B 的 MTP（多 token 預測）草稿頭，用中文文本微調，讓推測解碼在中文輸出時接受率更高（約 +6 個百分點），英文與程式碼不受影響。只訓練草稿頭、主幹不動；由於草稿一律由主幹驗證後才接受，換草稿頭只影響速度，不會改變模型輸出。