vLLM v0.33.0 Jumps to CUDA 13 with Faster Speculative Decoding and New OCR Models
The open-source LLM serving engine ships a major platform update with CUDA-13-only wheels, 11 new model architectures and large claimed throughput gains.
The vLLM project released version 0.33.0 on 2 November 2026, a major update combining 927 commits from 389 contributors. The serving stack moves to PyTorch 2.14 and Triton 3.8, and the release ships wheels for CUDA 13.0 only, ending the CUDA 12.9 line. Eleven new model architectures are supported, including the MiniCPM-OCR 1.0 document parser, Nemotron Parse 3.5 and the ColUltraVBERT visual document retriever. Serving additions include an OpenAI-compatible /v1/decisions endpoint, late-chunking token embeddings and concurrent worker startup for tensor parallelism above one.
The updated speculative decoding stack carries the headline numbers: a project pull-request benchmark reports DeepSeek-V4.1-Flash output rising from 1,373 to 3,390 tokens per second on an NVIDIA GB300 configuration, plus an 18 percent decode gain from quantized lm-head drafting at single-request concurrency. The release also removes the Medusa and MLP speculators and the legacy Ray executor, so teams on older pipelines face migration work before upgrading.