v1.830.0 v0.4.0
AudarAI/Audar-ASR-V1Oct 9, 2026 → Oct 10, 20260 days between releases
What's New
Selectively ports upstream Laya v0.4.1 runtime features to native MLX inference, introducing parallel layout support, new confidence gates, and improved language routing.
Breaking Changes
- Router() behavior changed: language-undecided inputs now default to the multilingual checkpoint instead of the English checkpoint. Use Router(default='english') to restore previous behavior.
New Features
- New `predict_tournament` method for narrowing choice sets without embedding models.
- Scalar and per-option-count `min_confidence` gates for Agent and Router.
- Full support for `option_layout='parallel'` checkpoints throughout MLX inference and compilation.
- Enhanced language detection for acronyms, fragments, and mixed scripts.
Improvements
- FP32 intermediate computation fix for CPU parallel-option attention on macOS 26.
- Release of MLX free-buffer caches on unload and eviction.
- Keep resident-model requests and status reads available during cold checkpoint construction.
Upgrade Notes
Set `Router(default='english')` to retain the previous fallback behavior for language-undecided inputs. Existing checkpoints remain sequential; do not enable parallel layout by editing an old checkpoint's configuration.
The Two Releases
v1.830.0
Oct 9, 2026· 12 days ago
This release introduces new token scopes for job dependencies, allows operators to compose full-code apps, and grants operators access to AI sessions. It also improves system reliability by replacing the hub sync script with an opt-in daily sync and adding retry logic for transient AI failures.
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, new confidence gates, and improved language routing.
Full release notes