Fixed three bugs that made Qwen3.5-122B a daily driver on Mac Studio

Running Qwen 3.5 122B on an M3 Ultra Mac Studio for long-context coding, the author found every follow-up took minutes before the first token. The model wasn't the problem—three cache bugs in the serving stack were: a per-turn message ID in the system prompt, a missing assistant reply in the conversation history, and stale checkpoint data poisoning the disk cache. After fixes, follow-up latency on a 130k-token context dropped from minutes to 11 seconds. The code is open-sourced as qMLX.

Why it matters: A concrete local-inference debugging story: three cache bugs fixed, dropping follow-up latency on 130K-token context from minutes to usable. Specific bug descriptions and post-fix numbers make it genuinely useful for anyone running large models on Mac. Not scored higher becaus...