oMLX oQ rescues aging M1 Max

This is the kind of performance win that actually changes how people use local models, not just a nice benchmark bump. If the numbers hold beyond one machine, oMLX is becoming a serious Apple Silicon backend for agentic coding by attacking the two pain points that matter most: prefill latency and cache churn.

• –The key signal is prefill: Claude Code timing out usually means the server cannot absorb long contexts fast enough, which makes local inference feel unusable even when decode speed is acceptable.
• –oQ-quantized models look like the immediate practical improvement here; DFlash is promising, but the post explicitly says it has not been tested yet.
• –The 32k benchmark context matters because agent workflows live in long-context territory, where repeated recomputation hurts the most.
• –This is less about raw model quality and more about turning a marginal Mac into something steady enough for daily local coding use.