vLLM 0.22, DeepSeek V4, Max KV Cache Compression

vLLM 0.22 delivers DeepSeek V4 production optimizations, massive KV cache compression, Rust frontend, and multi tier offloading.

vLLM 0.22 Stable Release is Here — Major Updates and Optimizations

I carefully reviewed the Release Notes and related technical blogs, and distilled the six most noteworthy changes to help you quickly decide whether to upgrade — and how to do it.

DeepSeek V4: From “It Runs” to “It Performs”

If you follow large model inference, DeepSeek V4 is definitely on your radar — a 1.6T total parameter, 49B active parameter MoE architecture with native support for 1 million token context.

In v0.20, vLLM’s support for V4 was still at the “it can run” stage. What v0.22 delivers is production-ready capability.

Architecture Refactoring: The model code has been reorganized from scattered locations into a dedicated vllm/models/deepseek_v4/ package. This is not just about code cleanliness — the independent package gives V4 its own fully optimized inference pipeline, free from the abstraction overhead of the generic model base class.