Moonshot AI has released Kimi K3, a 2.8-trillion-parameter mixture-of-experts model that the company describes as the world’s first open 3T-class system and the largest open-weight AI model ever released.
The model, which went live on July 16 through the Kimi app, Playground, and API, activates 16 of 896 experts per token using Stable LatentMoE, with an architecture built around the company’s proprietary Kimi Delta Attention mechanism, a hybrid linear-attention design that Moonshot says delivers up to 6.3 times faster decoding in million-token contexts compared to standard attention.
Kimi K3 supports a 1,048,576-token (1 million) context window and accepts text, image, and video input with native vision capabilities. Moonshot says full model weights will be released by July 27, 2026, making it the most significant open-weight challenge yet to Western frontier models.
On benchmarks, Moonshot claims K3 reaches 93.5 percent on GPQA Diamond, the strongest open-weight result on that graduate-level science benchmark at publication, and 91.2 percent on BrowseComp for web-based information retrieval. On Frontend Code Arena, the model jumped from rank 18 to rank 1, ahead of Anthropic’s Claude Fable 5. Independent evaluations show K3 broadly competitive with Claude Fable 5 across 35 shared benchmarks, winning on long-horizon coding tasks including Terminal-Bench 2.1 (88.3 percent) and SWE-Marathon, while Fable 5 leads on reasoning, knowledge, and vision benchmarks.
The pricing is aggressive: US$3 per million input tokens and US$15 per million output tokens, compared to Fable 5’s US$10 and US$50, roughly 70 percent less per token. A cached input rate of US$0.30 per million tokens makes it economical for high-volume agentic workloads.
The release positions Moonshot as perhaps the most credible open-weight challenger to Western labs, following similar moves by DeepSeek, Z.ai, and others in China’s rapidly maturing AI ecosystem. Moonshot was founded in 2023 and has raised significant funding from Chinese backers including Alibaba.
Reuters reported that the model outperforms leading US models on several evaluations, though independent verification of the full benchmark suite is still pending.
Sources: Tom’s Hardware; Reuters; LLM Stats