Qwen-3.6-27B-Sable

Qwen-3.6-27B-Sable

FP8 release of a Qwen3.6-27B-family vision-language checkpoint. The models it derives from and the changes made are documented in NOTICE.

Format

• Qwen3_5ForConditionalGeneration (64 decoder layers, hybrid attention, hidden size 5120, 27-block vision encoder, 262 144-token context).
• Language-model Linear layers quantized to FP8 with the compressed-tensors format (dynamic per-token activation scales); vision tower, linear-attention blocks, lm_head and the MTP head kept in bfloat16. Scheme in recipe.yaml.
• Two safetensors shards of about 31 GB combined, plus model-mtp.safetensors.
• Standard Qwen3.6 tokenizer, processor and chat template.

Usage

vllm serve magicworld7/Qwen-3.6-27B-Sable \
  --data-parallel-size 2 --max-model-len 81920 --gpu-memory-utilization 0.9 \
  --enable-prefix-caching

Query through the OpenAI-compatible chat API with text and image_url parts. Sampling defaults are read from generation_config.json. For reproducible results pass --revision with a specific commit hash.

Transformers users can load the repository with AutoModelForImageTextToText and AutoProcessor; compressed-tensors must be installed.

Notes

Later checkpoints under this account (Kestrel, Harrier) supersede Sable; it stays available so that existing deployments and evaluations can be reproduced.

License

Apache License 2.0 (LICENSE). NOTICE provides the change notice and upstream attribution required by the license. Qwen is a trademark of Alibaba Cloud.