v2.0.5-beta v0.29.0
high risk
MiniMax-AI/MiniMax-01Sep 7, 2026 → Sep 9, 20262 days between releases
What's New
The provided release notes describe updates for two different software projects, preventing a direct comparison.
Breaking Changes
- Ten deprecated model architectures removed
- FlexOlmo, Olmo3, and Hunyuan V1/VL migrated to Transformers backend
- PyAV video decoder backend removed
- CLI command renamed from `python -m vllm.entrypoints.openai.api_server` to `vllm serve`
New Features
- Model Runner V2 is now the default for all models
- Added new models including Hy4-preview, Qwen3.8-Flash-Next, and GraniteSWA
- RL weight sync improvements via sharded_rdt
- Mamba prefix caching improvements
Improvements
- Significant performance optimizations for Kimi K3 and DeepSeek V4
- Enhanced speculative decoding with per-request acceptance stats
- CUDA graph memory profiling for KV cache auto-sizing
- Batch-sharded sampling cuts per-step logits memory by 1/TP
Upgrade Notes
The older release notes refer to a file browser project, while the newer release notes are for a language model inference engine. Please verify the correct repository before attempting an upgrade.
The Two Releases
v2.0.5-beta
v2.0.5-beta
Sep 7, 2026· 1 month ago
This release addresses a security vulnerability in the public metadata API and introduces profile default settings for users and admins.
Full release notesv0.29.0
v0.29.0
Sep 9, 2026· 1 month ago
A major release making Model Runner V2 the default, introducing new models, and optimizing performance for Kimi K3 and DeepSeek V4.
Full release notes