Qwen3.6 35B-A3B

Qwen3.6-35B-A3B is a sparse Mixture-of-Experts model in the Qwen family, carrying 35 billion total parameters while activating only 3 billion per token. It is the first open-weight variant of the Qwen3.6 series, released by the Qwen team to bring the line's stability and real-world coding focus to self-hosted and community workflows. The weights ship as a post-trained Hugging Face Transformers artifact compatible with major inference stacks, including vLLM, SGLang, and KTransformers, making deployment straightforward for teams that want a capable coding model without managing proprietary APIs.

The model is tuned for agentic coding, with the Qwen team highlighting stronger handling of frontend workflows and repository-level reasoning, as well as a new Thinking Preservation option that retains reasoning context from earlier messages to streamline iterative development. Despite its compact active footprint, it is reported to surpass its predecessor Qwen3.5-35B-A3B and rival considerably larger dense models such as Qwen3.5-27B and Gemma4-31B on agentic coding tasks. It continues to support multimodal thinking and non-thinking modes alongside text output, positioning it as a versatile open-source choice for developers building coding assistants, multimodal agents, and other productivity tools.