vLLM has released v0.29.0, a 594-commit upgrade that makes Model Runner V2 the default for all models. Model Runner V1 remains available for some unsupported configurations and ROCm models.
vLLM Watch analysis
What happened
The release adds support for new models including Hy4-preview, Qwen3.8-Flash-Next, Granite variants, NemotronH Omni Reasoning V3 and Kimi K3 NVFP4 checkpoints. It also expands speculative decoding, multimodal inputs, prefix caching, queue admission and distributed weight updates.
The release notes report internal Mamba prefill checkpoints delivering a 9% to 25% improvement in time to first token. They also report about a 5% end-to-end latency improvement from one Kimi-K3 optimisation. These are release-note claims, not independent WittyWires benchmarks.
Our top picks
- Model Runner V2 becomes the house default
The newer execution path now covers all models, with V1 retained where V2 still lacks required features. - A wider model door
New support covers Hy4-preview, Qwen3.8-Flash-Next, Granite variants, NemotronH Omni Reasoning V3 and Kimi K3 NVFP4 checkpoints. - Mamba serving gets a useful nudge
vLLM says internal prefill checkpoints improve time to first token by 9% to 25% in tested workloads. - Speculative decoding gets more instrumentation
Responses can expose per-request acceptance statistics for draft tokens. - Multimodal serving gets less fussy
Video embeds, modality-scoped processor settings and cache and media-processing fixes land together. - Operators get more control
New queue admission flags help cap queued requests or tokens.
Why it matters
For teams running open models in production, this release changes the ordinary execution path while widening the models and workloads vLLM can support.
The migration is not complete. Workloads relying on features still missing from Model Runner V2 can fall back to V1, and vLLM says it is targeting V1’s removal in v0.32.
Our read
This is meaningful infrastructure work, not just a long changelog. Read it if you run vLLM, especially with hybrid Mamba models, speculative decoding, multimodal inputs or newly supported architectures. Treat the performance numbers as hypotheses to test on your own hardware.
What to watch
- Model Runner V2 across your existing model and hardware mix.
- V1-only features blocking migration before the stated v0.32 target.
- Whether the reported gains survive production-shaped workloads.
- Upgrade work caused by removed architectures and deprecated entry points.
Discussion spark: Which v0.29.0 change would most affect your vLLM deployment: Model Runner V2, new model support, multimodal serving or queue controls?
Sources and evidence
- v0.29.0 (10 September 2026, 23:53 UTC)
This is an independent WittyWires tracker and is not operated by, endorsed by, or affiliated with vLLM or the vLLM Project.