v3.3.4 v0.28.0
vllm-project/vllm-omniAug 22, 2026 → Aug 31, 20269 days between releases
What's New
Major architectural shift from reverse proxy to LLM serving platform with new runtime and speech stack
Breaking Changes
- Removed legacy stage-config loading path
- Rebased onto vLLM 0.28.0
New Features
- Host Weight Runtime (HWR) for managing weight artifacts
- Scheduler-managed paged KV cache for diffusion models
- Native full-duplex serving for PersonaPlex and NVIDIA Nemotron VoiceChat
- New model integrations: HiDream-O1-Image, SANA-Video 2B, LTX-2.5, and pi0 VLA
- Advanced MiniMax H3 deployment optimizations
Improvements
- Advanced MiniMax H3 deployment stack with modular pipelines and continuous batching
- Expanded full-duplex speech serving capabilities
- GPU, NPU, ROCm, and MUSA optimizations
Upgrade Notes
Migration is required due to the removal of the legacy stage-config loading path and the rebase onto vLLM 0.28.0. Review configuration files for compatibility with the new Host Weight Runtime and updated model integrations.
The Two Releases
v3.3.4
Aug 22, 2026· 2 months ago
This release prioritizes renewing the user experience and UI design, introducing a modern App Store and improving Forward Auth, TLS/ACME management, and Docker hardening.
Full release notesv0.28.0
Aug 31, 2026· 2 months ago
This release features 397 merged changes, introducing a new Host Weight Runtime for distributed layerwise offload, a scheduler-managed paged KV cache for diffusion models like HunyuanImage3, and an expanded full-duplex speech stack. It rebases onto vLLM 0.28 and adds extensive support for multimodal models including video, music, and VLA.
Full release notes