v0.12.1
What's new in v0.12.1
- chore: bump version to 0.12.1 (#1431) (ca2056a)
- perf(suffix): make SuffixDecoding cheap on traffic it doesn't suit, and give its telemetry an exit (#1419) (5f96d6f)
- docs(guides): Gemma 4 12B serve recipe for an 18 GB Mac (#1423) (87527b8)
- fix(rapid-mac): app-owned first-run gate + cheetah tray icon (#1430) (306e369)
- fix(gemma4-26b): correct the MoE flag, replace fabricated suffix numbers, add a 32GB serve guide (#1429) (51b5680)
- fix(codex): stabilize DeepSeek engineering turns (#1428) (89370a5)
- chore(rapid-mac): app is download-only; reserve rapid-mlx name for the engine (#1427) (aa234c7)
- fix(cache): exclude transient Codex priming from reusable prefix (#1425) (d7c267e)
- fix(tests): align community bench gates with schema v3 (#1426) (85b91a7)
- chore(rapid-mac): cut 0.12.0 + make release signing-optional (#1422) (4c2fafe)
- fix(codex): make engineering priming evidence-aware (#1424) (049d68d)
- fix(packaging): make [all] include a validated audio stack (#1421) (f3e4a8d)
- fix(codex): recover from unavailable test commands (#1418) (7b337e9)
- fix(codex): bound DeepSeek engineering action priming (#1417) (402f5ea)
- feat(rapid-mac): loopback-only telemetry endpoint override for local auditing (#1416) (d39ab8c)
- test(deepseek): add progressive context soak harness (#1414) (2476ee5)
- perf(dspark): add checkpoint verify and exact rollback (#1411) (a261b49)
- chore(rapid-mac): 0.12.0 polish — drop rapid-desktop branding, Apache-2.0, fix Launch page (#1413) (88cdd18)
- fix(prefill): adapt long-context memory pressure (#1410) (105f131)
- fix(agent): terminate repetition and preserve Responses failures (#1409) (807bdd8)
- fix(gemma4): unbreak bench, unlock live KV quantization, expose the KV-projection override (#1408) (03e8062)
- feat(app): add apps/rapid-mac — Ollama-style local-LLM menu-bar app (#1406) (7046049)
- fix(deepseek): harden V4 DSML tool protocol (#1407) (2ff3fdc)
- fix(bench): load weights on the mlx-step worker so freeform bench isn't 0 tok/s (#1404) (5821da0)
- feat(bench): submit community benchmarks over HTTP instead of a GitHub PR (#1403) (a73cf8c)
- fix(deepseek): stabilize Codex loops and adapt DSpark depth (#1402) (458069c)
- fix(deps): cap mlx/mlx-lm/mlx-vlm + gate bound bumps on a coherence sweep (#1248) (#1400) (4f03c49)
- fix(auth): normalize bearer token whitespace for rate limit stability (#1291) (4d6fe3e)
- fix(release-gauntlet): pre-flight detects console-script 'rapid-mlx serve' too (#1376) (d263a27)
Community contributors
Install: brew upgrade rapid-mlx or pip install -U rapid-mlx==0.12.1 (or just rapid-mlx upgrade).