

The vLLM 0.33.0 release candidate describes a broad update to the open-source inference server, including new model architectures and changes to the software packages used to run it. The candidate was published on November 2; a release candidate is a preview for testing, not a claim that a final stable release is available.
According to the project’s notes, the candidate adds support for models including Yandex AliceAI-Foundation-80B-A3B and NVIDIA Nemotron Parse 3.5. The notes also describe an upgrade to PyTorch 2.14 and changes to CUDA build artifacts. Support in a serving engine is separate from the licenses, memory needs and availability of each model.
Check compatibility before upgrading
The release notes say the CUDA 12.9 wheels and images are no longer published for this version in favor of CUDA 13.0 artifacts. Operators should verify their driver, container and hardware requirements against the actual release they choose. The candidate notes also list removed or deprecated components, which can affect existing deployment scripts.
Some performance figures in the notes use multi-GPU data-center systems and specific settings. They should not be assumed to apply to a desktop GPU or a different workload. We have not installed the candidate or tested its performance.
Our view: model support is useful, but the compatibility changes may matter more to someone running an existing service. Read the version-specific notes and validate on a separate environment before replacing a working inference stack.
Watch: vLLM background from PyTorch
This 2024 PyTorch conference talk by vLLM researchers explains the project; it predates the 0.33 release candidate.
EXPLORE THE SOURCE
Projects in this story
GitHub repositories
- vllm-project/vllm ↗View project details
Independent reporting based on the linked sources. Product and performance claims are attributed to their publishers unless noted.



