The MTP head is quantized along with the model, so --speculative-config in vLLM works out of the box and roughly doubles decode throughput (measured +98 % on temp-0 prose vs the same GPU without MTP; draft acceptance 41–67 % depending on content).

Why another int4?

At the time of upload, the available 4-bit options either dropped the MTP tensors (no speculative decoding) or kept hundreds of tensors in bf16 (28 GB — no longer fits a 24–32 GB GPU next to a useful KV cache). This one follows the recipe that served Qwen3.6-27B in production for months:

W4A16, group size 128, symmetric, auto_round:auto_gptq packing (AutoRound 0.14.2, default tuning)
Quantized blocks: model.language_model.layers and mtp.layers
The 96 GDN linear-attention in_proj_a / in_proj_b tensors are kept bf16 (quantizing them hurts the hybrid-attention path far more than it saves)
Vision tower and embeddings untouched (bf16) — multimodal input still works

Measured (vLLM 0.22/0.23)

Setup	Result
Weights in VRAM	~18.0 GiB (TP1) / 9.1 GiB per rank (TP2)
RTX 5090 @ 400 W cap, TP1, fp8 KV, 215K ctx	prefill ~2,580 tok/s (74K cold doc), decode 124 tok/s with MTP n=3 / 59 without
2× RTX 3090 @ 200 W cap each, TP2, fp8 KV, 262K ctx	2.22× concurrency at full 262,144 ctx; decode ~73 tok/s with MTP
MTP draft acceptance	41–46 % (prose) · ~67 % (mixed) · higher on code/predictable text
Tool calling	works with `--tool-call-parser qwen3_coder` + `--enable-auto-tool-choice`

Serving example

bash

vllm serve Frozenlock/Qwen3.8-27B-int4-AutoRound \
  --quantization auto_round \
  --max-model-len 215000 \
  --kv-cache-dtype fp8 \
  --enable-prefix-caching \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

Context is 262,144 native (fits on 2×24 GB at TP2 with fp8 KV; on a single 32 GB card ~215K leaves KV headroom). The base model supports YaRN extension to 1M — see the Qwen3.8-27B card for the rope_parameters block; budget ~35 KB of KV per token (fp8) when sizing.

reasoning_effort (none/low/medium/xhigh) works as in the base model. Note that sending chat_template_kwargs {"enable_thinking": false} makes the template silently ignore reasoning_effort — use one knob or the other.

Notes

All speeds above were measured on power-capped GPUs (5090 at 400 W of a 600 W stock limit; 3090s at 200 W of 350 W stock). Stock power limits should do the same or better — the 3090 decode figures in particular leave headroom.
Calibration: AutoRound default dataset/iters, batch 4 × grad-accum 2, on 4× RTX 3090.
On Ampere/Blackwell the weights run through the Marlin W4A16 kernels; prefill speed is unchanged vs 8-bit (compute-bound), decode gains come from smaller weight reads, and the big win vs FP8/NVFP4 is the KV-cache room the smaller weights leave.
Quantized with intel/auto-round 0.14.2.