Qwen3.6 35B A3B is the first open-weight variant of the Qwen3.6 series, positioned by its publisher as a stability-and-utility-focused successor to the February Qwen3.5 release. It ships as a post-trained causal language model in Hugging Face Transformers format, with weights that are compatible with Hugging Face Transformers, vLLM, SGLang, and KTransformers, giving developers flexible deployment options. The release narrative emphasizes two upgrades: more fluent agentic coding, including frontend workflows and repository-level reasoning, and a Thinking Preservation option that retains reasoning context across messages to streamline iterative development.
The architecture follows a Mixture-of-Experts design with 35B total parameters and 3B activated, arranged across 40 layers with a hidden dimension of 2048, combining linear Gated DeltaNet layers with Gated Attention layers and MoE blocks. Third-party tracking places the model on a cost-efficiency chart at a blended price of about $0.14 per million tokens and a composite benchmark score of 32.4, with a GPQA score around 86 sitting alongside comparable entries from GLM and GPT-class systems. Together, these traits make the model a practical fit for coding assistants, repository-aware agentic workflows, and other tasks where reasoning continuity and open-weight self-hosting matter.
