Qwen launches Qwen3.6-35B-A3B MoE model

Hot take: if the benchmark claims hold in real workflows, this is the kind of efficiency jump that makes large-model capabilities feel far more deployable.

• –The 3B-active/35B-total setup is the main story: it promises much lower serving cost than dense models while keeping a much larger expert pool behind the router.
• –The multimodal plus coding angle broadens the appeal beyond pure chat, especially for agentic and developer tooling use cases.
• –Apache 2.0 matters as much as the model itself for adoption, since it removes licensing friction for commercial and local deployments.
• –The risk is that “on par with 10x larger models” is launch-language; real-world agent reliability and multimodal robustness still need independent validation.