Qwen3.6-35B-A3B shines in OpenCode

This reads less like a benchmark and more like a proof of life for serious local coding. The interesting part is not that the model is flawless, but that it stayed useful across multiple planning and build cycles instead of spiraling when the task got messy. The model seems strong at error-driven iteration, which matters more than one-shot perfection for real codebases. The user’s intervention to reduce churn suggests a practical sweet spot: high-level architectural guidance plus enough code-editing discipline to adapt when scope is wrong. Running it locally on a 4090 with a 262k context and tuned cache settings shows the hardware and software stack is finally reaching a usable ceiling for large-agent workflows. OpenCode appears to be a good harness for exposing that capability, especially because it can issue parallel tool calls and keep the loop moving. The post is still anecdotal, so the real signal is developer confidence: local models are moving from toy fallback toward a credible option for some coding sessions.