Kimi K3 Open-Sourced: Anthropic Finally Shows Its Hand

Moonshot AI open-sourced the Kimi K3 model, whose benchmark performance is comparable to Anthropic's Claude Sonnet 4. The open-source release includes the model weights and a technical report. The report reveals that Anthropic has adopted a scaled-up version of the Chinchilla optimal ratio, supporting a larger batch size and higher learning rate for more efficient training. Kimi K3 is based on a Mixture of Experts (MoE) architecture, with a total of 48 experts and 2 active experts per token, with 37 billion activated parameters out of 264 billion total parameters. The model is designed for scaling and efficiency, with a 128K context length and FP8 mixed-precision training. The open-sourcing of Kimi K3 is seen as a move to democratize access to advanced AI and challenge established players like Anthropic and OpenAI.