Qwen3.6-35B-A3B doubles down on agents

The headline here is not raw model size; it’s execution quality. The public benchmarks and early user reports both point to the same thing: stronger coding-agent behavior, better tool use, and more complete UI-plus-logic outputs than you’d expect from a model this cheap to run.

• –Qwen’s own benchmark table shows the strongest movement in coding-agent tasks: SWE-bench Verified 73.4, Terminal-Bench 51.5, and MCP-Atlas 62.8.
• –General knowledge and reasoning look comparatively steady rather than breakout: MMLU-Pro sits at 85.2 and GPQA at 86.0, which supports the “better at doing, not magically wiser” read.
• –The model’s 262k native context and explicit tool-call support make it feel tuned for long-horizon agent loops, not just chat completion.
• –For local developers, the 3B-active MoE setup matters as much as the scores: it suggests a better quality-to-cost tradeoff for agent workloads than denser models in the same class.
• –The Reddit demo is believable because it matches the release posture: fast enough reasoning, then a working frontend without the usual half-finished feel.