What breaks when upgrading to vLLM v0.29.0?

vLLM v0.29.0 removes ten deprecated architectures, including Arctic, Chameleon and MPT, and drops the PyAV video decoder. FlexOlmo, Olmo3 and Hunyuan models now run through the Transformers backend instead of native code. The `python -m vllm.entrypoints.openai.api_server` entry point is deprecated in favour of `vllm serve`.

Which new models does vLLM v0.29.0 serve?

vLLM v0.29.0 adds Hy4-preview, Tencent's 770B/49B mixture-of-experts model, and Qwen3.8-Flash-Next with several quantization options. The release also covers GraniteSWA variants, NemotronH_Omni_Reasoning_V3 with multi-token prediction support, and Kimi K3 NVFP4 checkpoints.

How much faster is vLLM v0.29.0?

Gains in vLLM v0.29.0 are kernel-specific rather than one headline number. Kimi K3's Mamba metadata path reports a 6.6–7.6× kernel speedup, Hopper low-latency GEMM improves 12.9–25.2%, Blackwell autotuning cuts end-to-end latency by 33.6%, and Mamba prefix caching improves time-to-first-token by 9–25%.

Do I have to switch Model Runner V2 on?

No. Model Runner V2 is the default for all models in vLLM v0.29.0, so upgrading changes the execution path automatically, where earlier releases needed an opt-in. The V2 runner ships CUDA graph memory profiling and batch-sharded sampling, which reduces logits memory by 1/TP.

vLLM · 2026-09-09 · major

vLLM v0.29.0 — Model Runner V2 becomes the default for every model

vLLM v0.29.0 makes Model Runner V2 the default for all models and adds serving support for Tencent's Hy4-preview and Qwen3.8-Flash-Next. The release lands 594 commits from 277 contributors and removes ten deprecated architectures.

GitHub card for the vLLM inference and serving engine repository

Model Runner V2 is now vLLM's default execution path, closing a rollout that had been opt-in for several releases.

Quick facts

Maker	vLLM project
Version	v0.29.0
Released	September 9, 2026
License	Apache-2.0
Default runner	Model Runner V2
Contributors	277 across 594 commits
Breaking change	Ten architectures removed

What is it?

Model Runner V2 becomes the default for all models in vLLM v0.29.0, ending its opt-in phase. The release also strips out ten deprecated architectures — Arctic, Chameleon and MPT among them — and deprecates the `python -m vllm.entrypoints.openai.api_server` entry point in favour of `vllm serve`. New serving support covers Hy4-preview, Tencent's 770B/49B mixture-of-experts model, and Qwen3.8-Flash-Next.

How does it work?

The V2 runner adds CUDA graph memory profiling and batch-sharded sampling, which cuts logits memory by 1/TP. Underneath it, the release retunes kernels per hardware generation rather than applying one global optimisation: Mamba metadata handling, Hopper low-latency GEMM, Blackwell autotuning and Mamba prefix caching each got their own pass.

Why does it matter?

Teams serving open-weight models on vLLM pick up the speedups without a config change, but the removals are a real migration. Anyone still running Arctic, Chameleon or MPT, or launching through the old `api_server` module path, has to move before upgrading. FlexOlmo, Olmo3 and Hunyuan models now route through the Transformers backend, and the PyAV video decoder is gone.

Who is it for?

inference and platform engineers

Frequently asked questions

What breaks when upgrading to vLLM v0.29.0?: vLLM v0.29.0 removes ten deprecated architectures, including Arctic, Chameleon and MPT, and drops the PyAV video decoder. FlexOlmo, Olmo3 and Hunyuan models now run through the Transformers backend instead of native code. The `python -m vllm.entrypoints.openai.api_server` entry point is deprecated in favour of `vllm serve`.
Which new models does vLLM v0.29.0 serve?: vLLM v0.29.0 adds Hy4-preview, Tencent's 770B/49B mixture-of-experts model, and Qwen3.8-Flash-Next with several quantization options. The release also covers GraniteSWA variants, NemotronH_Omni_Reasoning_V3 with multi-token prediction support, and Kimi K3 NVFP4 checkpoints.
How much faster is vLLM v0.29.0?: Gains in vLLM v0.29.0 are kernel-specific rather than one headline number. Kimi K3's Mamba metadata path reports a 6.6–7.6× kernel speedup, Hopper low-latency GEMM improves 12.9–25.2%, Blackwell autotuning cuts end-to-end latency by 33.6%, and Mamba prefix caching improves time-to-first-token by 9–25%.
Do I have to switch Model Runner V2 on?: No. Model Runner V2 is the default for all models in vLLM v0.29.0, so upgrading changes the execution path automatically, where earlier releases needed an opt-in. The V2 runner ships CUDA graph memory profiling and batch-sharded sampling, which reduces logits memory by 1/TP.

Try it

pip install -U vllm && vllm serve <model>