Ollama Local LLM qwen3-coder Astro 6 Claude Code AI Engineering Benchmark Apple Silicon
Benchmarking local qwen3-coder on an Astro 6 build on M4 Max — 4 model failures, not Mac failures
Honest benchmark: 64 GB unified memory, 40 GPU cores, qwen3-coder:30b on Ollama, Claude Code as the harness. The model confused Astro with Svelte, ignored an explicit STOP, and stalled 30 min on a single turn. Technical diagnosis + the real fixes.
Read more