vLLM v0.33.0 ships with 927 commits as v0.34.0 release candidates land
vLLM tagged v0.33.0 on 2 November 2026 with shorter Docker image names, then published two v0.34.0 release candidates within ten days.
vLLM, the community-developed inference engine hosted on GitHub, tagged v0.33.0 on 2 November 2026. Two release candidates for v0.34.0 followed on 11 and 12 November. The previous stable release, v0.32.0, arrived on 18 October 2026, so the project is tagging a new version about every two weeks.
The v0.33.0 release page counts 927 commits from 389 contributors, 122 of whom were contributing for the first time. The v0.32.0 page recorded 814 commits from 363 contributors, with 121 new. Both figures are the project’s own tallies.
The clearest packaging change in v0.33.0 is a shorter image name. Every build is now also published under vllm/vllm, with variants for EFA networking (-efa), AMD ROCm (-rocm, plus -rocm100 and -rocm72), CPU-only (-cpu) and Intel XPU (-xpu). The latest tag follows the same pattern, so latest-rocm resolves to the newest ROCm image. Pre-built artifacts are attached to the release page.
The v0.34.0 candidates carry only brief notes at the time of writing. The first candidate includes a fix to speculative decoding on XPU so that greedy argmax treats NaN values as negative infinity. The second stops the documentation and tests from setting gpu_memory_utilization in most examples.
The highlights section for v0.33.0 did not load in the material we checked, so the feature list for that release is not covered here. The release pages do not restate the project’s licence.
Why it matters
If you deploy vLLM from containers, the new vllm/vllm name is worth adopting now, and you should pin a version tag rather than latest given the two-week cadence. The documentation change in rc2 suggests the project wants users to stop hard-coding memory utilisation, which is a hint about defaults to watch when v0.34.0 is final. Be sceptical of commit and contributor counts as a quality signal. They measure activity, not stability, and a release candidate’s one-line note is not a changelog. Read the full v0.33.0 highlights before upgrading production, and test speculative decoding paths if you run on XPU.