On the first day of DeepSeek Open Source Week, we officially released our latest technological achievement, FlashMLA, an efficient Multi-Layer Attention decoding kernel specifically designed for NVIDIA Hopper architecture GPUs. This technology is optimized for variable-length sequence scenarios and can significantly enhance the inference performance of large models. The core technical features of FlashMLA include comprehensive support for BF16 precision and a paged key-value cache with a block size of 64.

Recently, Canva introduced a series of brand new AI features, the most notable being its image generation tool - 'Dream Lab'. This tool was launched after Canva acquired the generative AI startup Leonardo.ai and utilizes Leonardo's Phoenix model (just don't confuse it with Adobe's Firefly AI). Users can now generate images in various styles simply by describing them, such as

Alibaba Damo Academy has released the industry's first remote sensing AI large model, capable of identifying farmland, crops, buildings, and more. The model is now available for use on the AI Earth geo-science cloud platform. The Remote Sensing AI Interpretation General Segmentation Model (AIE-SEG) standardizes the image segmentation task. The model can recognize nearly a hundred categories of remote sensing land cover, with an accuracy improvement of 25% to 30%.