Kimi K3

What it is

Kimi K3 is a multimodal model released by Moonshot AI on 16 July 2026 ^[1][2]. It has 2.8 trillion total parameters, with 104 billion activated per token. Wikipedia describes it as the largest open-weights model family ^[3][1]. It understands text, images and video ^[3].

Origins

Moonshot reported a private long-horizon Elo of 1547 for K3, which is 732 points above K2.6 ^[2][1].

The weights were released on 27 July 2026 under the Kimi K3 License. Under that license, hosts earning over US$20 million a year from K3 must negotiate with Moonshot, and Wikipedia reports revenue sharing of up to 30 percent. Large products built on the model must attribute Kimi K3 ^[2][1][3].

Main features

The architecture has 93 layers, one of them dense, with 896 experts. Each token is routed to 16 experts, plus 2 shared experts that are always active ^[3].

The attention design mixes 69 Kimi Delta Attention (KDA) layers, a linear-attention type, with 24 Gated MLA layers. Linear attention keeps the cost of processing long inputs lower than standard attention does. The context window is 1,048,576 tokens, which is about one million tokens ^[3].

The model's vision component is a 401-million-parameter MoonViT-V2 encoder. The weights use MXFP4 with MXFP8 activations, a low-precision number format, through quantization-aware training ^[3]. Quantization-aware training means the model is trained with the precision reduction already in place, so it keeps more of its quality when run at low precision.

Moonshot's self-reported scores at max effort are 93.5 on GPQA Diamond, 43.5 on Humanity's Last Exam (HLE) without tools and 56.0 with tools, 88.3 on Terminal-Bench 2.1, 91.2 on BrowseComp and 84.8 on OSWorld-Verified ^[3]. The scores are the company's own figures.

Agent Swarm was expanded to up to 300 sub-agents ^[1]. The model debuted at No. 3 on the Artificial Analysis leaderboard and topped Arena.ai's front-end web development benchmark ^[2][1].

Its place in development

Kimi K3 received attention from outside the company. White House OSTP director Michael Kratsios accused Moonshot of distilling Anthropic's Claude models ^[2]. Distillation means training one model on the outputs of another, and the accusation concerns that practice. In August 2026 the legal startup Harvey launched a legal model based on Kimi K3 ^[2].

The line continued with Kimi K3.1 Pro in October 2026 ^[1][2].