v9.13.0 v0.9.3
webiny/webiny-jsAug 27, 2026 → Sep 7, 202611 days between releases
What's New
Introduces FP8 support and Mistral-specific features, replacing the previous LEMUR and MUVERA focus.
Breaking Changes
- Server error codes aligned to OpenAI standard may break clients expecting legacy codes
New Features
- First-class FP8 and concurrent DFlash decode support
- Mistral sequences capacity gauge
- Separate prefill and decode throughput logging
- NVFP4 checkpoint loading and cuTile Blackwell acceleration
- Tool chat template JSON rendering alignment
Improvements
- Significant performance improvements through first-class FP8 support
- Concurrency fixes and scheduler improvements
- Server error codes aligned to OpenAI standard
- Improved hybrid FP8 serving
Bug Fixes
- Multimodal inputs (Qwen media work) kept out of decode
- Windows build errors fixed
- OpenAI chat message selection fix
- Cached prompt tokens reported in usage
Upgrade Notes
The provided release notes describe two different repositories (webiny-js and mistral.rs) with disjoint feature sets, so a direct upgrade path cannot be determined from the text.
The Two Releases
v9.13.0
Aug 27, 2026· 2 months ago
This release introduces LEMUR learned multi-vector retrieval, expands late-interaction capabilities with configurable pooling, and converts the MUVERA algorithm to Torch, alongside various bug fixes for agents and cross-encoders.
Full release notesv0.9.3
Sep 7, 2026· 1 month ago
This release focuses on significant performance improvements through first-class FP8 support and concurrent decoding, alongside server standardization and fixes for multimodal inputs and tool templates.
Full release notes