A developer posted on Hacker News that three specific code fixes in the QMLX framework turned Qwen3.5-122B into a usable daily driver on Mac Studio. The thread received 60 points and 24 comments.
The post links directly to the detailed write-up at mrzk.io.
What the Fixes Address
QMLX is a quantization and inference library optimized for Apple Silicon. The author identified and patched three bugs that previously caused excessive memory fragmentation, incorrect KV-cache sizing, and stalled token generation on unified memory systems.
After the patches, the 122B model loads and generates responses without crashing or swapping to disk on an M2 Ultra Mac Studio with 192 GB RAM.
Performance Numbers Reported
The post does not publish full benchmark tables, but the author states the model now sustains interactive use for coding and research tasks. Early comments note generation speeds sufficient for daily workflows once the memory bugs were resolved.
No official tokens-per-second figures appear in the thread.
How to Apply the Fixes
Readers can follow the steps in the linked blog post. The changes involve three small modifications to the QMLX memory allocator and cache initialization routines, followed by recompilation.
The repository and patch details are hosted at the URL above. No separate pull request has been submitted yet.
Tradeoffs Observed
- Pros: Full 122B parameter model runs locally on consumer Apple hardware; no cloud API costs; full context window preserved.
- Cons: Requires manual patching and rebuild; 192 GB unified memory is effectively mandatory; single-threaded bottlenecks remain in parts of the pipeline.
Comparison with Alternatives
| Setup | Model Size | Hardware Requirement | Daily Driver Status |
|---|---|---|---|
| QMLX (patched) | 122B | Mac Studio 192 GB | Yes |
| MLX default | 72B | Mac Studio 64 GB | Yes |
| llama.cpp (Metal) | 70B | Mac Studio 64 GB | Yes |
| Unpatched QMLX | 122B | Mac Studio 192 GB | No (crashes) |
The patched QMLX path is currently the only route that keeps the full 122B Qwen3.5 model resident without swapping.
Who Benefits
Developers who already own a high-memory Mac Studio and need the specific capabilities of Qwen3.5-122B gain immediate value. Users with 64 GB or less unified memory, or those unwilling to compile custom frameworks, should continue with 70-72B models instead.
Verdict
The three targeted patches demonstrate that framework-level memory bugs, not raw hardware limits, were the main blocker for running 122B-scale models on current Mac Studio configurations. Once merged or widely shared, the changes could expand the set of practical local models on Apple Silicon by one size class.
The work also highlights how small, hardware-specific allocator fixes can unlock entire model tiers without new silicon.
Top comments (0)