Running Coding Models Locally on a Mac: Flint, Qwen3.6, and Kapitsa

Local model competition has moved from “can it run?” to “which inference engine is best for coding agents?” Public Kapitsa benchmarks for Qwen3.6-35B-A3B on an M4 Pro show large differences between llama.cpp, Ollama, and Kapitsa. These are project benchmarks, not universal results.nnCoding agents also need first-token latency, long-context behavior, tool-call reliability, memory peaks, and recovery behavior. A faster engine that emits invalid JSON can be more expensive than a slower stable engine. Fix the model, quantization, context, and prompt before comparing results.

Sources

See the linked primary sources in the Chinese edition.