v0.9.1 v0.9.2
medium risk
webiny/webiny-jsAug 14, 2026 → Aug 20, 20266 days between releases
What's New
Adds Qwen3.5 video processing, MTP support, and CUDA graphs to v0.9.1
New Features
- Implement better video processor for Qwen3.5
- Add MTP (Multimodal Token Processing) support
- Support CUDA graphs for MTP and batch buckets
- Add support for DFlash and DFlash 2 block-diffusion draft models
- Add in-place pooled recurrent state kernels and fused MLP
Improvements
- Improve engine metrics for scheduler status and TTFT/ITL
- Improve OpenAI server compatibility with /v1/responses
- Performance optimizations for MTP and GDN rollback
- Better support for hybrid models and auto device map
- Misc fixes for FAv3 and Qwen3.8
Bug Fixes
- Fix misc issues relating to FAv3
- Fix issues around gdn rollback for Qwen3.8
- Better support for hybrid models and auto device map
The Two Releases
v0.9.1
v0.9.1
Aug 14, 2026· 2 months ago
This release introduces significant improvements to the core scheduler, concurrent serving, and support for new models like Muse Glimmer, along with LoRA kernel improvements and GGUF support.
Full release notesv0.9.2
v0.9.2
Aug 20, 2026· 2 months ago
This release focuses on performance optimizations and new features for the Qwen3.5 model, including video processing, MTP support, and CUDA graph improvements.
Full release notes