vLLM v0.28.0 Adds Disk Offloading for the KV Cache

vLLM v0.28.0 shipped on August 26, 2026 with 584 commits from 270 contributors. Tiered KV cache offloading now reaches disk, the default max_num_batched_tokens rises from 8192 to 16384, and bitsandbytes moves to an out-of-tree plugin.

vLLM inference open-source self-hosting LLM

The inference server vLLM released v0.28.0 on August 26, 2026, carrying 584 commits from 270 contributors, 76 of them new ¹.

Tiered KV cache offloading now reaches disk, and secondary tier managers can be plugged in from outside the tree via module_path. Partial secondary-tier load results, tiering metrics, and a canonical CPU layout for parallelism-agnostic offload also landed ¹
Defaults changed: max_num_batched_tokens goes from 8192 to 16384, prefix caching is enabled by default for Mamba models, and the Blackwell CUDA graph capture default rises to 1024 ¹
Listed as breaking changes: bitsandbytes support moved to an out-of-tree plugin, the Transformers dependency was bumped to 5.15.0, the deprecated calculate_kv_scales was removed, and so was override_attention_dtype ¹
DeepSeek V4’s sparse MLA now works end-to-end for plain decode, MTP, and DSpark speculative decoding ¹. A stack-wide optimization effort for Kimi-K3 landed in the same release ¹

Being able to push the KV cache all the way out to disk is a practical escape hatch for long-context work on setups short of GPU memory. The default and breaking changes are the kind that warrant a check if you run existing launch options or Docker images as-is. Note that Hy4 preview, whose weights Tencent published on August 28, points to vLLM as an official deployment path.

Sources

vLLM v0.28.0 Release Notes - Official vLLM release notes (published August 26, 2026)