v2.2.21 v0.32.0

Name: vllm
Author: vllm-project

high risk

vllm-project/vllmOct 6, 2026 → Oct 18, 202612 days between releases

What's New

This release shifts from NocoBase bug fixes to vLLM's structured reasoning and low-precision KV cache, dropping Python 3.10 support.

Breaking Changes

Python 3.10 is no longer supported; the minimum is Python 3.11.
Default CORS origins changed from `*` to none.
Removed deprecated APIs `wait_for_inflight_requests` and `AsyncLLM.pause_generation()`.
Removed `torch.compile` sequence parallelism and AsyncTP passes.
Dependencies upgraded to `transformers>=5.16.1` and `tokenizers>=0.23.2`.

New Features

New `/v1/systemone` endpoint for structured reasoning.
New `hf` tool and reasoning parser for structured decisions.
LoRA support for BERT, RoBERTa, and XLM-RoBERTa embedding models.
New models: EmbeddingGemma 2, MiniCPM-V 4.7, Nemotron 3.5 ASR.
Video input support in the Transformers backend.

Improvements

Comprehensive low-precision KV cache support including NVFP4 and FP8.
Model Runner V2 becomes the default for ROCm.
Performance optimizations via shared-memory tensor arena and least-loaded API worker selection.
Expanded model support including EmbeddingGemma 2, MiniCPM-V 4.7, and video inputs.

Upgrade Notes

Ensure your environment uses Python 3.11 or later. Update CORS configuration to explicitly list allowed origins. Remove calls to deprecated APIs like `wait_for_inflight_requests`.

The Two Releases

v2.2.21

Oct 6, 2026· 16 days ago

This release primarily addresses bug fixes within the client-v2 UI editor and notification systems, alongside dependency upgrades for nodemailer, hono, and moment.

Full release notes

v0.32.0

Oct 18, 2026· 4 days ago

This release introduces structured reasoning capabilities via the `/v1/systemone` endpoint and HF parsers, significantly enhances low-precision KV caching (NVFP4, FP8) across hardware, and establishes Model Runner V2 as the default for ROCm. It also brings expanded model support, performance optimizations for startup and serving, and critical security fixes.

Full release notes