On February 21, DeepSeek revealed on social media platform X that it will be open-sourcing five code repositories over the next week, with new content unlocked daily. The company emphasized its commitment to sharing “small but sincere progress” as part of its mission to accelerate tech innovation. The company’s online services have been tested and are now ready for deployment in production. The company, which calls itself a “small team,” highlighted that each shared piece of code builds momentum, contributing to the collective growth of the community. “No ivory towers – just pure garage-energy and community-driven innovation,” DeepSeek stated.
As part of this initiative, DeepSeek introduced its first model today, FlashMLA, a decoding kernel optimized for Hopper GPUs. The model is designed for variable-length sequences and is already in production. [Deepseek on X]
Related
