vLLM Releases Version 0.28.0

**vLLM 0.28.0 Released: New Features and Performance Enhancements**

OpenAI‑inspired large‑language‑model inference library vLLM has just shipped version 0.28.0, adding a suite of improvements that streamline deployment and boost throughput. The update, released on GitHub on 2026‑08‑29, introduces a new “dynamic batching” scheduler that automatically adjusts batch sizes based on token‑level latency, reducing average inference time by up to 15 % on typical GPU workloads. Additionally, the release adds first‑class support for the latest NVIDIA Hopper GPUs, along with updated CUDA 12.1 bindings, allowing users to tap into the full performance potential of the newest hardware. Minor bug fixes address memory‑leak issues in the KV‑cache manager and correct a rare crash that could occur when loading models with mixed‑precision weights.

The v0.28.0 changelog also expands the public API, adding a `vllm.InferenceEngine` context manager that simplifies multi‑model orchestration in production environments. Documentation has been updated to include detailed deployment guides for Kubernetes and Docker Swarm, and the test suite now covers 98 % of the codebase, improving reliability for large‑scale inference pipelines. Community feedback on the release was modest but positive, with the announcement receiving 20 up‑votes and one comment on Hacker News, indicating growing interest in the library’s evolving capabilities.

With these enhancements, vLLM continues to solidify its position as a leading open‑source solution for fast, scalable language‑model inference. Developers can now benefit from tighter integration with modern GPU architectures, more efficient batching, and a clearer deployment path, making it easier to bring state‑of‑the‑art models to production at lower cost and higher speed.