Fixed three bugs that made Qwen3.5-122B a daily driver on Mac Studio
mrzk.io
July 11, 2026
12 min read
45/100
Summary
A follow-up question on a 50,000 token conversation with an AI can take three to five minutes before the first token is generated. A three-week effort was dedicated to debugging a cache leak in order to optimize performance on the Mac Studio.
Key Takeaways
- The author experienced significant latency issues with the DS4 Flash model, taking three to five minutes for the first token to appear after a 50,000 token conversation, making it unsuitable for real-time coding tasks.
- The author switched to the Qwen 3.5 122B model, which provided better performance for long-context tasks on the M3 Mac Studio Ultra, operating fully offline with no API calls or data leaving the machine.
- The author developed qMLX, a specialized serving stack for hybrid attention, after forking the rapid-mlx project to address specific performance issues related to the model's architecture.
- The Qwen 3.5 model's architecture allowed for efficient memory usage and deep SSD-backed KV cache, enabling better long-context retention without performance degradation.
Coverage & discussions
Hacker News · mrzk.io
Community discussions
Historical count: 5 · 5 saved in sample · Historical sample
Analysis ready
Community Sentiment
Positives
- The detail in the post is appreciated, showcasing a useful addition to the AI space that encourages deeper exploration of model capabilities.
- Impressive debugging skills are on display, and the benchmarks provided are a valuable resource for understanding model performance.
Concerns
- There's a concerning over-reliance on LLMs for content generation, leading to repetitive phrasing that detracts from the writing's authenticity.
- The unique message ID bug breaking the KV cache feels predictable, raising doubts about the model's robustness in handling such changes.
Related Articles
Local Qwen isn't a worse Opus, it's a different tool
Jun 18, 2026
Running Gemma 4 26B at 5 tokens/sec on a 13-year-old Xeon with no GPU
Jul 15, 2026
Why your local LLM feels dumber than it is
Aug 22, 2026
Qwen 3.8 27B is excellent, but it defaults to overthinking things
Aug 16, 2026
LLM Neuroanatomy II: Modern LLM Hacking and Hints of a Universal Language?
Mar 24, 2026