Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled - AWQ-W4A16

Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled quantized to AWQ-W4A16 (4-bit weights).

What this is

Same footprint as W4A16 but calibrates faster and usually holds up better on instruction-tuned and multilingual models. Start here.

Caveat. Asymmetric; a few older vLLM kernels prefer symmetric W4A16.

Details


Source	`Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled`
Scheme	`AWQ-W4A16` (4-bit)
Format	compressed-tensors
Parameters	9.7B
Size on disk	8.6 GB
Compression	2.24x smaller than the 19.3 GB source
Calibration	`HuggingFaceH4/ultrachat_200k`, 256 samples
Left unquantized	`lm_head`, `re:.visual.`, `re:.vision_tower.`, `re:.vision_model.`, `re:.vision.`, `re:.multi_modal_projector.`, `re:.merger.`
Quantized on	RTX A6000
Quantized by	Sohailhosseini

Usage

vllm serve Sohailhosseini/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-AWQ-W4A16 \
  --max-model-len 32768

from vllm import LLM, SamplingParams

if __name__ == "__main__":
    llm = LLM("Sohailhosseini/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-AWQ-W4A16", max_model_len=32768)
    out = llm.chat(
        [{"role": "user", "content": "What is quantization? Answer in one sentence."}],
        SamplingParams(temperature=0.6, max_tokens=512),
    )
    print(out[0].outputs[0].text)

Provenance

Produced with HF-quantized. recipe.yaml in this repo is the exact modifier stack that was applied, and the scheme, ignored layers, calibration set and hardware are in the table above.

Licence is inherited from the source model. Quantization does not change what you are permitted to do with the weights.