What's new in v0.12.4
- chore: bump version to 0.12.4 (2f0e64c)
- Harden benchmark methodology metadata (#1455) (0be4361)
- docs(readme): the front door was still on the pre-0.12 tier map (#1454) (deee741)
- fix(app): the GB column was two metrics wearing one hat (#1453) (342c442)
- fix(deepseek): prevent reopening disabled thinking (#1452) (5fb38a9)
- fix(engine): return partial output for repetition-guard stop instead of 503 (#1450) (822cbf1)
- fix(cache): persist deepest prefix first on shutdown (#1451) (2d7bbc7)
- fix(audio): pre-resolve Kokoro's spaCy G2P model at the route gate so first TTS request can't 500 (#1254) (2883393)
- fix(cache): replay tokens across prompt boundaries (#1449) (c682eb8)
- feat(install): make the curl banner recommend what the app recommends (#1447) (bfe40ca)
- fix(deepseek): materialize cache graphs during decode (#1448) (1ce3dd0)
- fix(deepseek): accept sampled DSML tool aliases (#1446) (481ae17)
- feat(app): give 8 GB Macs a recommendation instead of a rejection (#1443) (1f05e91)
- fix(cache): auto-enable prefix reuse for dense-recurrent aliases (qwen3.5/3.6 dense, bonsai) (#1445) (18db012)
- feat(models): serve LFM2.5-2.6B, whose eight quants share one repo (#1442) (d119166)
Install: brew upgrade rapid-mlx or pip install -U rapid-mlx==0.12.4 (or just rapid-mlx upgrade).