Apologies for the frequent releases. I am prioritizing performance, stability, and user-experience improvements in these patches so oMLX can deliver the best possible experience. For the major 0.4.0 upgrade notes, please see the 0.4.0 release notes.
This patch release focuses on memory-pressure stability, robust model discovery, managed server lifecycle controls, and macOS app and CLI quality-of-life improvements.
Highlights
- Improved prefill memory handling. Static memory reserves were reduced, throttle tiers were simplified, and prefill chunk floors were raised for more stable throughput.
- Idle-model eviction before prefill throttling. Loaded but idle models can now be freed before new work is rejected or throttled.
- Managed server lifecycle controls. The macOS app and admin surfaces now expose server lifecycle control, and when the macOS app is installed the server can also be controlled from the CLI with
omlx start,omlx stop, andomlx restart. - Copyable model IDs in the SwiftUI app. Model ID copy buttons were added to the Models list and per-model detail header.
Fixes
- Fixed TurboQuant KV cache selection for MLA models by excluding incompatible MLA models from TurboQuant cache paths. by @popfido (#1626)
- Fixed scheduler handling so prefill error outputs are preserved during decode. by @ken-zzzzz (#1622)
- Fixed inaccessible secondary model directories so they are tolerated instead of breaking model discovery or admin reload flows.
- Fixed LFM2 pythonic tool parser support.
- Fixed a macOS crash when deleting chat template kwargs. by @nethbotheju (#1634)
- Fixed shell integration behavior so editing shell init files requires explicit user consent. (#1633)
- Fixed app CLI wrapper symlink resolution in bundled macOS builds.
macOS App
- Added model ID copy affordances in the SwiftUI Models screen and model detail screen.
- Improved server lifecycle integration between the macOS app and local control server.
- Improved shell integration consent flow and related UI copy.
New Contributors
Thank you to everyone making their first contribution in this release: