Qwen3.6-35B-A3B tops Gemini Flash at coding

Interesting signal, but not a lab-grade verdict: this is one prompt, one task, and the model still had syntax issues. Even so, the economics are getting harder to ignore when a local MoE model can produce stronger architecture than a paid API at consumer-GPU speeds.

• –Qwen3.6’s official positioning around agentic coding and repository-level reasoning lines up with the post’s claim that it made better design choices
• –A* pathfinding is a decent stress test for structure and state handling, but it is still a narrow benchmark for general coding quality
• –The biggest practical takeaway is cost and latency: local inference removes per-token spend and gives solo devs instant iteration loops
• –Syntax mistakes still matter, but they are cheaper to fix than weak system design, especially for scaffold-first coding workflows
• –If this result holds up across broader tasks, it strengthens the case for open-weight local models in everyday coding assistance