v1.7.18 v0.29.0
high risk
vllm-project/vllmSep 8, 2026 → Sep 9, 20261 day between releases
What's New
Model Runner V2 becomes default with performance boosts and breaking changes
Breaking Changes
- Ten deprecated model architectures have been removed
- FlexOlmo, Olmo3, and Hunyuan V1/VL migrated to Transformers backend
- PyAV video decoder backend has been removed
- Startup command `python -m vllm.entrypoints.openai.api_server` deprecated in favor of `vllm serve`
- Several environment variables have been removed
New Features
- CUDA graph memory profiling for KV cache auto-sizing in Model Runner V2
- Batch-sharded sampling reducing per-step logits memory by 1/TP
- Support for prompt embeds and `extract_hidden_states` speculation in MRV2
- New model support: Hy4-preview, Qwen3.8-Flash-Next, GraniteSWA, GraniteMoeSWA, NemotronH_Omni_Reasoning_V3, and Kimi K3 NVFP4
- RL weight sync improvements including `sharded_rdt` P2P backend
Improvements
- Performance improvements for Kimi K3 include fused MXFP4 top-k finalization and GEMM-RS extensions
- Performance improvements for DeepSeek V4 include fused shared experts and MLA gate merging
- Mamba prefix caching delivers 9-25% TTFT improvement with new retention interval controls
- Enhanced speculative decoding features include adaptive verification extended to logprobs
- Kimi K3 performance gains include tuned Hopper low-latency GEMM dispatched on SM100
Upgrade Notes
Update your startup command from `python -m vllm.entrypoints.openai.api_server` to `vllm serve`. Be aware that the PyAV video decoder backend and specific environment variables have been removed.
The Two Releases
v1.7.18
Sep 8, 2026· 1 month ago
Adds tool call interception feature for security testing.
Full release notesv0.29.0
v0.29.0
Sep 9, 2026· 1 month ago
This release marks a major milestone by making Model Runner V2 the default for all models, accompanied by significant performance optimizations for Kimi K3 and DeepSeek V4, expanded model support, and several breaking changes.
Full release notes