Qwen3.6-35B-A3B Tops Claude Opus 4.7

The real story is not “better at pelicans,” it’s that a laptop-runnable open model is now credible enough to challenge frontier proprietary models in creative structured output. That’s a strong signal that sparse MoE + better post-training is making local models genuinely useful, not just cheap.

• –The model card says Qwen3.6-35B-A3B is Apache 2.0, with 35B total params, 3B activated, and native support for multimodal input.
• –Alibaba is positioning the release around agentic coding and “thinking preservation,” which matters more for day-to-day dev workflows than benchmark vanity.
• –The benchmark table shows competitive results across coding, tool use, reasoning, and vision-language tasks, with especially strong agentic coding scores.
• –Willison’s comparison is deliberately absurd, but it’s a useful reminder that model quality is now multi-dimensional: a model can be “worse” overall and still be better at a specific task on consumer hardware.
• –For developers, the practical takeaway is simpler: this is another strong open model to try locally before defaulting to a closed frontier API.