v9.4.0 v0.30.0
vllm-project/vllmSep 19, 2026 → Sep 22, 20263 days between releases
What's New
Major overhaul introduces Model Runner V2, new models, and several breaking changes.
Breaking Changes
- Scale-out endpoints now require `--enable-scale-out`.
- GPTQ activation ordering (`g_idx`) has been removed.
- Items deprecated in v0.29 have been removed.
- The `all` Mamba cache mode is deprecated.
- The `python -m vllm.entrypoints.grpc_server` command is deprecated.
New Features
- Model Runner V2 with dual-batch overlap and FULL CUDA graphs.
- Fast Start feature using `--load-format ipc_cache`.
- Watermarking capabilities using Gumbel-max generation.
- HiSparse host-resident tier for sparse-MLA decode.
- Support for new models like DeepSeek-V4.1-Flash and GLM-5.3-Flash.
Improvements
- Significant performance optimizations and CUDA graph improvements.
- Enhanced large-scale serving capabilities with PCP+DCP.
- New quantization strategies introduced.
- Persistent per-GPU weight-cache daemon for Fast Start.
Upgrade Notes
Enable scale-out endpoints explicitly using `--enable-scale-out` if previously used. Remove references to deprecated commands like `python -m vllm.entrypoints.grpc_server`. Update GPTQ configuration as activation ordering `g_idx` is no longer supported.
The Two Releases
v9.4.0
Sep 19, 2026· 1 month ago
Adds on-demand chat history fetching, allows per-device webhooks to merge with global targets, and adds option to ignore status media downloads. Also includes various reliability fixes for history sync, device purge, webhook delivery, and route ordering.
Full release notesv0.30.0
Sep 22, 2026· 30 days ago
This release introduces Model Runner V2 with significant performance optimizations and CUDA graph improvements, adds support for new models like DeepSeek-V4.1-Flash and GLM-5.3-Flash, and enhances large-scale serving capabilities with PCP+DCP and advanced KV offloading. It also introduces new quantization strategies and watermarking features.
Full release notes