vLLM v0.31.0 Released with DeepSeek-V4.1-Flash Performance Enhancements

Core summary

One sentence to understand this update

vLLM v0.31.0, featuring 717 commits, introduces significant performance improvements for DeepSeek-V4.1-Flash through FlashMLA mega attention and V4.1 NVFP4 compressed KV cache.

Impact & opportunity

What this could mean

Builders working with large language models can achieve faster and more efficient inference, especially for DeepSeek-V4.1-Flash, by upgrading to vLLM v0.31.0.