Qwen3.6 27B

Qwen3.6-27B is the first open-weight release in the Qwen3.6 family, presented as a dense 27-billion-parameter multimodal model that supports both thinking and non-thinking modes. It is distributed as a post-trained causal language model with a vision encoder, with weights published in the Hugging Face Transformers format and confirmed to be compatible with vLLM, SGLang, and KTransformers for deployment. The architecture combines linear and full-attention paths, using a 64-layer layout arranged as 16 groups of Gated DeltaNet followed by FFN, with one Gated Attention block per group finishing the cycle, and a hidden dimension of 5120 over a 248,320-token padded vocabulary.

The release focuses on practical developer productivity rather than raw scale, with the Qwen team framing it as delivering flagship-level agentic coding performance while remaining straightforward to deploy as a dense model. Headline upgrades include more fluent handling of frontend workflows and repository-level reasoning, alongside a new Thinking Preservation option that retains reasoning context from earlier messages to streamline iterative coding sessions. As a dense 27B it is positioned to surpass the previous-generation open-source flagship Qwen3.5-397B-A17B across major coding benchmarks, making it a strong fit for teams that want top-tier coding assistance without the routing complexity of a mixture-of-experts setup.