vLLM v0.29.0 Released with Model Runner V2 Default and New Model Support | Tool Update
vLLM v0.29.0 was released on Sep 8, 2026 featuring 594 commits from 277 contributors including 91 new ones, with Model Runner V2 now default for all models.
Published on MyPrivateClaw
Sep 10, 2026, 3:19 PM UTC
Coverage date
Sep 8, 2026
Last updated
Sep 10, 2026, 3:19 PM UTC
News summary
Model Runner V2 is now the default for all models in v0.29.0, completing the rollout that began with pooling models; MRV1 remains for a few ROCm models and features MRV2 does not yet support. v0.29.0 adds native support for Hy4 preview (Tencent's 770B/49B active MoE with Gated DeepSeek Sparse Attention and native MTP), Qwen3.8 Flash Next (BF16/FP8/NVFP4 with MTP), GraniteSWA, GraniteMoeSWA, NemotronH Omni Reasoning V3 with MTP, and Kimi K3 NVFP4 checkpoints. v0.29.0 removes ten deprecated model architectures, migrates FlexOlmo Olmo3 and Hunyuan V1/VL to the Transformers modeling backend, removes the PyAV video decoder backend, deprecates python m vllm.entrypoints.openai.api server in favor of vllm serve, and removes environment variables VLLM TEST FORCE FP8 MARLIN and VLLM ROCM USE AITER FP4 ASM GEMM. v0.29.0 enables FlashInfer all reduce by default for TP CUDA groups (opt out with VLLM…