Qwen3.6-27B-MTP-bf16-test

This is not a standalone model. It holds only the Multi-Token Prediction (MTP) drafter head, so it carries no token embeddings and no lm_head and it cannot generate text alone. Load it as the draft model beside a Qwen3.6-27B target, which supplies both.

This repository keeps the drafter unquantized, so it costs 0.85 GB against the 4-bit build's 0.24 GB. Use it when you want the highest acceptance rate, or when you compare quantized and unquantized drafting.

Use with mlx-vlm

mlx_vlm generate \
  --model vvsotnikov/Qwen3.6-27B-4bit-test \
  --draft-model vvsotnikov/Qwen3.6-27B-MTP-bf16-test \
  --prompt "Write a quicksort in Python." \
  --max-tokens 256 --temperature 0.6

How this was produced

python -m mlx_vlm.speculative.drafters.qwen3_5_mtp.split \
  --model Qwen/Qwen3.6-27B --output Qwen3.6-27B-MTP-bf16

The tool reads the 15 mtp. tensors out of the base checkpoint, strips that prefix, and adds 1.0 to every RMSNorm weight, because Qwen stores the norm scale minus one. It finishes in under two seconds, since it touches only the two shards that hold those tensors.

Verification

This build is bit-identical to mlx-community/Qwen3.6-27B-MTP-bf16 across all 15 tensors, and the file sizes match exactly.

Check	Result
model_type	qwen3_5_mtp
block_size	3, from mtp_num_hidden_layers + 2
Tensors	15, all bfloat16
RMSNorm shift	applied exactly once, checked against the source

Pairing rules

Use a drafter and a target that come from ONE checkpoint, because the drafter binds to the target's embeddings at runtime. Prefer this bf16 build over the 4-bit one on Mixture-of-Experts targets, where quantized MTP weights are reported to cut acceptance from about 80 percent to about 10 percent.

License and attribution

The weights derive from Qwen/Qwen3.6-27B under Apache 2.0, so the original license and its terms carry over. Refer to the upstream model card for details.