The vLLM project released v0.33.0 on November 2, 2026, spanning 927 commits from 389 contributors. The release is published under the tag v0.33.0rc3 with GitHub's prerelease flag off and a full v0.33.0 announcement body; this report follows the project's own presentation as the stable v0.33.0 release. Read the release notes.
What changed
- New model support: Reflection Beam-1.0 (501B MoE, 23B active, with DFlash2 drafter), Yandex AliceAI-Foundation-80B-A3B, Recursive Flint, Muse Spark Open, Step-5 Preview, NVIDIA Alpamayo 2 Super, Nemotron-Labs Dory1.0-4B, Nemotron Parse 3.5, MiniCPM-OCR 1.0, Domyn Edge and the ColUltraVBERT visual document retriever.
- Platform baseline: PyTorch 2.14.0 with torchvision 0.29.0 and Triton 3.8.0; CUDA 12.9 wheels and images are discontinued — release artifacts are CUDA 13.0 only. The default ROCm image moves to ROCm 10.0, with 7.2 still published as
-rocm72. - Speculative decoding: a new XPress refiner for block-diffusion drafters; drafters no longer reserve token-budget slots — in the cited PR benchmark, DeepSeek-V4.1-Flash with DSpark on GB300 rises from 1,373 to 3,390 output tokens per second; optional quantized draft
lm_head(+18% decode at concurrency 1, per the PR); EAGLE/MTP prefix-cache hits keep their last block via lookahead hashes, on by default. - Serving and API: an OpenAI-compatible
/v1/decisionsendpoint on by default; late-chunking token embeddings; a Rust frontend parity push; concurrent worker startup for TP > 1; a watchdog that dumps stack traces from stuck processes.
Who is affected and what to do
Self-hosting teams on CUDA 12.9 must move to the CUDA 13.0 artifacts (or pin an older release) before upgrading. Users of Medusa, the MLP speculator or the legacy Ray compiled-graph executor need to migrate — all three are removed. The NIXL connector now requires NIXL 1.5 or newer, and requests arriving while generation is paused receive HTTP 503 instead of queueing.
Vendor-reported performance
The notes cite large gains on specific hardware — for example opt-in latency-specialized decode for Kimi-K3 on GB300 (1.5–2x at TP8/TP16), FlashInfer DCP prefill with 39–48% lower mean time-to-first-token at 100k input on 8xB200, and MonoKernel decode on MI300X/MI325X (2.32x P50 interactivity at concurrency 1). These figures come from the project's pull-request benchmarks, not independent testing.
Previous coverage: v0.32.0. Benchmark caveats: our verification guide.