Moonshot AI is set to publish the complete weight files for Kimi K3 on July 27, making it the largest open-weight AI model ever released. The model, which launched via API on July 16, packs 2.8 trillion total parameters in a mixture-of-experts design, though only 16 of its 896 experts activate per token — roughly 50 billion active parameters per step.
The release comes with a catch: at MXFP4 four-bit quantization, the weights total about 1.4 terabytes, requiring roughly eighteen 80-gigabyte accelerators or equivalent hardware just to hold the model in memory before processing any input. Kimi K3 also introduces a proprietary attention mechanism called Kimi Delta Attention, which Moonshot says enables decoding up to 6.3 times faster than standard approaches, alongside a 1-million-token context window.
On independent benchmarks, Kimi K3 ranks 4th on the Artificial Analysis Intelligence Index among 189 tested models, and 1st on the Frontend Code Arena leaderboard for blind developer testing. Moonshot has not yet published the license that will accompany the weights, so commercial usability remains unconfirmed. The release caps a dense week of open-weight launches, following DeepSeek V4's stable release on July 24.