oMLX server issues affect startup, memory, and model loading
oMLX has 7 recent tracked items. One server startup problem is tied to an invalid MCP setting, where an unused `cwd` field can stop the server from launching. Memory errors can happen during prefill even on Macs with large amounts of RAM, because the work can go past the `metal_cap` limit.
In the chat screen, action buttons can be cut off when prompts are long or the window is small. The setting that should start the server automatically when the app launches does not work as expected. The `gpt-oss-20b-tq3` model can fail to load on oMLX 0.4.4 because of an internal `` error.
Two improvement requests ask for server settings to stay reachable even when the server is stopped, and for Specfill to show a more complete list of .
Key points
- oMLX has 7 recent bug or improvement items.
- An unused `cwd` field in MCP settings can prevent the server from starting.
- Prefill can hit the `metal_cap` memory limit even on high-RAM Macs.
- The launch-time auto-start setting for the server may not work.
- `gpt-oss-20b-tq3` may fail to load on oMLX 0.4.4 because of a `` error.