Moonshot AI introduced Kimi K3 on July 16, 2026, calling it “our most capable model” in its Kimi K3 tech blog. The model is a 2.8-trillion-parameter mixture-of-experts system with a 1-million-token context window and native vision, and its weights are published on Hugging Face under the Kimi K3 License.
On this page
Architecture and inputs
According to the blog, Kimi K3 is built on two pieces of Moonshot’s own research, Kimi Delta Attention and Attention Residuals, and activates 16 out of 896 experts per token. The Hugging Face model card lists 2.8T total parameters and 104B activated parameters, a context length of 1,048,576 tokens, and text, image and video as input modalities.
The blog said the full model weights “will be released by July 27, 2026.” The model card now carries the weights, and states that “both the code repository and the model weights are released under the Kimi K3 License.”
Where it runs
Moonshot lists four ways to use the model: Kimi.ai, Kimi Work, Kimi Code and the Kimi API. The model card adds that the API is reached on platform.kimi.ai by selecting kimi-k3, and that Moonshot provides OpenAI- and Anthropic-compatible endpoints.
Benchmark results reported by Moonshot
All figures below are Moonshot’s own. The model card reports its Kimi K3 results with reasoning effort set to “max” and temperature 1.0.
| Benchmark | Kimi K3 | Source |
|---|---|---|
| GPQA Diamond | 93.5 | Model card |
| Terminal-Bench 2.1 | 88.3 | Model card |
| BrowseComp | 91.2 | Model card |
| Video-MME (with subtitles) | 90.0 | Model card |
| DeepSWE v1.1 (mini-SWE-agent harness) | 67.3 | Tech blog |
The DeepSWE figure is the one Moonshot gives in its tech blog, where it reports the score on DeepSWE v1.1 tasks with the mini-SWE-agent harness.
Moonshot does not claim the top spot overall. Its blog says that “while its overall performance still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol, Kimi K3 demonstrated frontier-level performance across our evaluation suite, consistently outperforming other tested models.”
What it changes
Kimi K3 extends Moonshot’s K-series, which its blog index traces back through Kimi K2.6 in April 2026 and Kimi K2 in July 2025. With the weights posted, developers can run the model outside Moonshot’s own apps and API, subject to the terms of the Kimi K3 License. The release is listed with the other July 2026 launches in our AI model release timeline.
