v0.8.13 v0.4.0
cheaper-inference/OmniRouteOct 9, 2026 → Oct 10, 20260 days between releases
What's New
Introduces MLX inference runtime features and parallel layout support, replacing HID device enhancements with a breaking Router behavior change.
Breaking Changes
- Router() now sends language-undecided inputs to the multilingual checkpoint by default; set Router(default="english") to revert.
New Features
- Added predict_tournament to narrow large choice sets without embeddings.
- Added scalar and per-option-count min_confidence gates for Agent and Router.
- Added support for option_layout="parallel" throughout MLX inference.
- Added custom checkpoint registration methods (register, unregister, registered, models).
- Added cold checkpoint construction improvements with resident-model requests.
Improvements
- Fixed FP16 kernel crashes on hosted macOS 26 by using FP32 intermediate computation.
- Deduplicate builds and synchronize targeted unloads per checkpoint.
- Release MLX free-buffer caches on unload and eviction.
- Keep resident-model requests and status reads available during cold checkpoint construction.
Bug Fixes
- Fixed FP16 kernel crashes on hosted macOS 26.
Upgrade Notes
To maintain the previous fallback behavior for undecided inputs, update Router initialization to Router(default="english"). Existing checkpoints remain sequential; do not enable parallel layout by editing an old checkpoint's configuration.
The Two Releases
v0.8.13
Oct 9, 2026· 13 days ago
This release enhances HID device support for the Powerplay receiver and Logitech G733, improves GUI functionality with custom actions and friendly names, and resolves various platform-specific bugs.
Full release noteslaya-mlx v0.4.0
Oct 10, 2026· 12 days ago
Selectively ports upstream Laya v0.4.1 runtime features to native MLX inference, introducing parallel layout support, confidence gates, and FP32 CPU attention fixes.
Full release notes