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

Core summary

One sentence to understand this update

vLLM v0.31.0 is released with 717 commits, notably improving DeepSeek-V4.1-Flash performance using FlashMLA mega attention and NVFP4 compressed KV cache.

Impact & opportunity

What this could mean

Builders using vLLM can expect significant performance gains, especially with DeepSeek-V4.1-Flash, enabling more efficient and cost-effective deployment of large language models.