Chinese artificial intelligence firm DeepSeek has introduced FlashMLA, a decoding kernel designed to optimize Nvidia’s Hopper GPUs.

Engineered to leverage the architecture of the H800 GPU, FlashMLA reduces memory consumption by 40-60% compared to traditional attention mechanisms, without sacrificing positional accuracy. This results in 2.3x faster inference speeds for 175B parameter language models compared to previous state-of-the-art implementations.

DeepSeek has already deployed FlashMLA in production settings, indicating its suitability for practical applications.

FlashMLA marks the first major release of Deepseek’s Open Source Week initiative. The software is available as open-source on GitHub, allowing developers and researchers worldwide to adapt it for their projects.

Recent DeepSeek developments