Moonshot AI releases Kimi K3, the largest open-weight model yet

Moonshot AI released the full weights of Kimi K3 on Hugging Face on July 27, a 2.8-trillion-parameter mixture-of-experts model — the biggest open-weight model published to date, now at about 11,700 likes. Only 104B parameters fire per token (16 of 896 experts), and it carries native text, image, and video understanding with a 1-million-token context window. The architecture is new: Kimi Delta Attention plus Attention Residuals, with quantization-aware training in MXFP4 weights and MXFP8 activations from the start. Its own benchmark tables claim 88.3 on Terminal-Bench 2.1, 93.5 on GPQA Diamond, 67.5 on DeepSWE, and 90.0 on Video-MME — all provider-reported. Two caveats: the custom Kimi K3 License isn't a standard open license, and running it yourself is not casual — the download alone is roughly 1.4TB.

Summarized by the Neuricon wire. Every claim links back to its primary source where available.