v0.0.21 v0.27.0
anomalyco/opentuiAug 9, 2026 → Aug 10, 20261 day between releases
What's New
Introduces Kimi K3 support, PyTorch 2.13.0 upgrade, and expands Model Runner V2 to encoder-only workloads.
Breaking Changes
- PyTorch 2.13.0 upgrade is a breaking environment change for XPU and CPU backends
New Features
- Support for Kimi K3 model with full stack integration
- Model Runner V2 expands to encoder-only and pooling workloads
- New models: Qwen3.5 dense and MoE, K-EXAONE-2.0-750B-A37B, VaultGemma
- Resilient large-scale serving with fault tolerance framework
Improvements
- FlashAttention 4 integration adds FP8 KV cache and headdim-256 support
- DeepSeek-V4 performance improvements via sequence parallelism and adaptive topk
- New JIT warmup infrastructure removes first-request compilation stalls
Bug Fixes
- Mamba/hybrid fixes including ReplaySSM caching and hybrid attention
Upgrade Notes
Upgrading requires updating the PyTorch environment to 2.13.0 or higher. The version jump indicates a significant architectural shift, so verify configuration compatibility for the new Model Runner V2 and encoder-only workloads.
The Two Releases
v0.0.21 - 2026-08-09
Aug 9, 2026· 2 months ago
A minor maintenance release focused on protocol utilities and version bumping. It does not introduce new features or breaking changes, serving as a preparatory step for subsequent development.
Full release notesv0.27.0
Aug 10, 2026· 2 months ago
Introduces Kimi K3 support, PyTorch 2.13.0 upgrade, and FlashAttention 4 integration on SM100. Model Runner V2 expands to non-generative workloads like encoding and pooling.
Full release notes