35B total / 3B active (MoE); FP8 variant released
262,144 tokens (native; extensible to ~1,010,000)
Apache-2.0
Open-weight Qwen3.6 release positioned as the first open variant in the Qwen3.6 line, focused on stability and real-world coding productivity. Official release notes emphasize stronger agentic coding behavior for frontend and repository-level workflows, plus a new thinking-preservation option that retains historical reasoning context to improve iterative development. The architecture remains a sparse MoE design with 256 experts and 8 routed experts plus shared capacity, combining Gated DeltaNet and gated attention blocks in a hybrid layout tuned for coding and agent tasks. Published benchmark tables in release materials report gains over prior Qwen3.5-35B-A3B on multiple coding-agent evaluations including SWE-Bench Pro, Terminal-Bench 2.0, Claw-Eval averages, NL2Repo, and QwenWebBench, while maintaining competitive general-agent and knowledge performance. Release artifacts are provided in standard Transformers format with compatibility guidance for vLLM, SGLang, KTransformers, and related inference stacks.