v0.16.2 v0.29.0
immich-app/immichSep 8, 2026 → Sep 9, 20261 day between releases
What's New
Model Runner V2 is now default with new models and performance optimizations.
Breaking Changes
- Ten deprecated model architectures removed.
- PyAV video decoder backend removed.
- CLI command `python -m vllm.entrypoints.openai.api_server` deprecated in favor of `vllm serve`.
- Removed environment variables `VLLM_TEST_FORCE_FP8_MARLIN` and `VLLM_ROCM_USE_AITER_FP4_ASM_GEMM`.
New Features
- Model Runner V2 rollout with CUDA graph memory profiling.
- New models: Hy4-preview, Qwen3.8-Flash-Next, GraniteSWA, GraniteMoeSWA.
- Speculative decoding acceptance stats and adaptive verification.
- RL weight sync P2P backend (sharded_rdt).
- New defaults for FlashInfer all-reduce and prefix cache hash.
Improvements
- Performance boosts for Kimi-K3 and DeepSeek V4.
- Mamba prefix caching improvements (9%-25% TTFT improvement).
- Speculative decoding metrics in OpenAI API responses.
- Batch-sharded sampling cuts per-step logits memory by 1/TP.
- Fused MXFP4 top-k finalization in K3 latent tail.
Upgrade Notes
Update your CLI commands from `python -m vllm.entrypoints.openai.api_server` to `vllm serve`. Remove the deprecated environment variables `VLLM_TEST_FORCE_FP8_MARLIN` and `VLLM_ROCM_USE_AITER_FP4_ASM_GEMM`. Note that the PyAV video decoder backend has been removed, which may affect video processing workflows.
The Two Releases
v0.16.2
Sep 8, 2026· 1 month ago
Minimal release notes: "Release created because its been a long time since the previous one and several minor changes have been made. Thank y"
Full release notesv0.29.0
Sep 9, 2026· 1 month ago
This is a major release defaulting Model Runner V2, adding new model support (Hy4-preview, Qwen3.8), and optimizing performance for Kimi-K3 and DeepSeek V4. It includes speculative decoding enhancements and Mamba prefix caching improvements.
Full release notes