Kimi K3 Open Weights: 2.8T Parameters, License & Rankings

896 experts, 16 active, a one-million-token window

Moonshot released Kimi K3 on July 16, 2026, and its Kimi Code changelog records it that day as released, open-sourced and available in Kimi Code. The downloadable weights came later: writing on July 20, Nathan Lambert reported that they were scheduled for July 27, the date of this entry. [4][3]

Kimi K3 is a mixture-of-experts model with 2.8 trillion total parameters and 104 billion active per token, drawn from 896 experts with 16 activated per token across 93 layers, 69 using Moonshot's Kimi Delta Attention and 24 using gated multi-head latent attention. The model card lists a one-million-token context window and quantization-aware training, storing weights in MXFP4 and activations in MXFP8. [1]

Moonshot released the weights under its own Kimi K3 License rather than a standard open-source license, and its model card describes the release as built on a "Stable LatentMoE framework" that the company credits with roughly a 2.5x improvement in scaling efficiency over Kimi K2, its prior open-weight model. Moonshot's own card also calls K3 the "world's first open 3T-class model," a claim about scale rather than an independently verified capability ranking. [1]

Atlas interpretation: The active-parameter count matters more than the headline total for anyone weighing whether to run the model themselves: 104 billion active parameters is closer to the footprint of a single large dense model a lab could serve directly, while the 2.8 trillion total is what drives the multi-terabyte download the summary above describes. Sparse activation is the mechanism that lets the Kimi line keep growing in total size without a proportional increase in serving cost. [1]

A narrowing gap to the closed frontier

Independent tracking placed Kimi K3 second on the Vals AI index and third on Artificial Analysis's Intelligence Index at release, behind Anthropic's Claude Fable and OpenAI's GPT-5.6 Sol Max, and first on the Frontend Code Arena benchmark. Writing on Interconnects, Nathan Lambert argued the release showed the gap between frontier closed models and the best open ones compressing from roughly six to nine months to something closer to three to five months. [3]

Atlas interpretation: Lambert's broader argument is that a release at this capability level is, in his words, economically decelerationist for closed labs: a usable open substitute weakens the pricing power and the fundraising narrative that justify further frontier training runs, even as he calls that outcome a net good for the field. He separately credits Chinese labs, Moonshot among them, with reaching this scale on far less capital than their American counterparts have raised. Both points are arguments about incentives and capital efficiency rather than measurements of K3 itself, and they come from a single commentator's analysis rather than a benchmark. [3]