Alibaba Qwen Team Releases Qwen3.6-27B: A Dense Open-Weight Model Outperforming 397B MoE on Agentic Coding Benchmarks

Alibaba Qwen launched Qwen3.6-27B, a 27B dense open-weight multimodal model under Apache 2.0.

It adds stronger agentic coding performance and a new thinking-preservation mode for longer tool use and reasoning.

The model uses a hybrid Gated DeltaNet/self-attention architecture and ships in BF16 and FP8 weights.

Qwen says it is supported by common inference stacks including Hugging Face, SGLang, vLLM, and KTransformers.