Qwen3.6 27B

Qwen3.6-27B is a 27-billion-parameter dense model designed as a versatile multimodal assistant that handles text, vision, and video inputs while producing text outputs. Its architecture pairs a causal language model with a vision encoder and uses a hybrid attention layout, alternating Gated DeltaNet linear-attention blocks with gated full-attention layers across 64 transformer layers. The model ships in Hugging Face Transformers format and is compatible with common serving stacks like vLLM, SGLang, and KTransformers, making it straightforward for developers to deploy on their own infrastructure. Compared with the prior 27B generation, it delivers notable gains in STEM reasoning, spatial understanding, object localization, video comprehension, document OCR, and visual agent behavior, giving it a broad multimodal footprint for real-world applications.

The release prioritizes practical developer workflows, particularly agentic coding across frontend and repository-level tasks, where the model now reasons over multi-file changes and UI generation with greater fluency. A new Thinking Preservation option allows long-running coding sessions to retain reasoning context from earlier messages, reducing overhead during iterative debugging and improving continuity across long agent loops. With a 262K-token context window, support for structured output and tool calling, and an open-weight license that lowers deployment friction, Qwen3.6-27B is well suited for teams building autonomous coding assistants, visual analysis pipelines, and research prototypes that need a balance of reasoning depth, multimodal understanding, and self-hosted control.