July 11, 2026
Cache me outside, Mac how bout that
Fixed three bugs that made Qwen3.5-122B a daily driver on Mac Studio
He fixed the endless waiting on his fancy Mac — but the comments came for his writing too
TLDR: A developer fixed three bugs that made his offline AI on a Mac painfully slow, turning multi-minute waits into something usable for real work. Commenters were split between admiring the detective work, obsessing over the benchmarks, and brutally accusing the post of sounding AI-written.
A developer’s three-week mission to stop his super-expensive Mac Studio from acting like a very smug coffee break machine has turned into exactly the kind of internet drama people love. The big win: he figured out why follow-up questions on huge chats were taking three to five minutes just to start answering, then patched three bugs and open-sourced the result as qMLX. In plain English, he made his local AI setup feel usable instead of painfully frozen.
But the real show was in the replies. One camp was genuinely impressed, calling the debugging deep-dive a gift to fellow tinkerers and praising the benchmark charts, even while nitpicking the graph scale like true forum perfectionists. Another camp zeroed in on the most absurd twist: a tiny changing message ID in the prompt was enough to wreck the system’s memory and force it to re-read everything from scratch every time. That detail had commenters reacting like they’d just watched a detective reveal the killer was the butler’s nametag.
Then came the spiciest side quest: was the post itself written by AI? One reader said the write-up sounded suspiciously like chatbot prose, quoting a polished phrase as evidence. Another went full scorched earth, dismissing it as “AI slop” and demanding the author “write it yourself.” So yes, the tech story is about fixing a slow AI assistant — but the comment-section story is about people applauding the engineering, roasting the prose, and arguing over whether the author got help from the very machines he was trying to speed up
Key Points
- •The article describes replacing DS4 Flash with Qwen 3.5 122B on an M3 Mac Studio Ultra to better support long-context local coding workflows.
- •With the previous setup, follow-up prompts in conversations beyond 50,000 tokens could take three to five minutes before the first token was generated.
- •The author selected Qwen 3.5 122B because its active-parameter profile, local offline operation, reasoning and tool-calling balance, and memory fit matched the M3 Ultra and 96GB unified memory constraints.
- •The author forked rapid-mlx into qMLX to specialize serving for the model’s hybrid attention architecture.
- •The article attributes multi-minute follow-up delays to cache behavior around hybrid attention, where GatedDeltaNet-based SSM layers cannot be rewound or trimmed, causing repeated reprocessing of long contexts.