INT4 weight-only (W4A16) quantization of Qwen/Qwen3.6-35B-A3B — a hybrid Mixture-of-Experts model (256 experts, top-8, ~3B active) with GatedDeltaNet linear-attention, a vision tower and an MTP head.

Quantization

AWQ activation-aware scale search + symmetric INT4 weights, group_size 32, MSE observer (llm-compressor).
Quantized: the MoE expert FFNs only (the bulk of the weights).
Kept bf16 (quality-sensitive): self_attn, the router (mlp.gate), shared_expert, GatedDeltaNet (linear_attn), lm_head, embeddings, vision tower, MTP head.
Format: compressed-tensors (pack-quantized). Full recipe: recipe.yaml.

Evaluation

Served as W4A16 (vLLM, thinking enabled, temperature 0.6):

Benchmark	Score
GSM8K	96.8% (242/250)
MMLU-Pro	80.2% (401/500)

Usage (vLLM)

bash

# W4A16 — int4 expert weights, fp16 activations
vllm serve Avesed/Qwen3.6-35B-A3B-INT4-W4A16 \
  --tensor-parallel-size 2 --trust-remote-code --reasoning-parser qwen3

On Ampere the same checkpoint can be served as W4A8 (int4 weights + int8 dynamic activations) with the vllm-ampere-optimized fork, which adds the INT8-activation MoE Marlin path:

bash

vllm serve Avesed/Qwen3.6-35B-A3B-INT4-W4A16 \
  --tensor-parallel-size 2 --marlin-input-dtype int8 --trust-remote-code --reasoning-parser qwen3

Quantized with vllm-ampere-optimized/quantize.