Kimi K3

What is Kimi K3?

Kimi K3 is Moonshot AI's open-weight model family with a 1M-token context window and native vision, built for long-horizon coding, research and documents. The current flagship is Kimi K3.1 Pro, a post-trained K3, with a faster Highspeed variant. Use it in the Kimi app, Kimi Work, Kimi Code or the pay-as-you-go Kimi API.

What can you do with Kimi K3?

Moonshot AI introduced Kimi K3 in July 2026 as a 2.8-trillion-parameter Mixture-of-Experts model that activates 16 of its 896 experts, with a 1M-token context window and native understanding of text, images and video. The company positions it for long-horizon coding, knowledge work such as slides and interactive research reports, and reasoning, and its showcases include playable 3D games and video edits built from source clips.

Kimi K3.1 Pro, the first model of the K3.1 series, is a post-trained version of K3 with much more reinforcement-learning compute. Moonshot highlights coding across real codebases and game engines such as Unity, Unreal Engine and Godot, 3D and CAD modeling, financial research and fewer factual errors. K3.1 Pro Highspeed is the same generation tuned for 2 to 3 times the speed.

Try it on a small task

Suggested workflow: in Kimi Code, select Kimi K3.1 Pro with the /model command and give it one bounded task in a test repository, such as explaining a module and adding tests for it. Review the diff and run the tests yourself before giving it something larger.

For knowledge work, upload a report you know well to kimi.ai and ask for a short slide deck or a summary with sources. Check every figure against the original document, since even Moonshot's own factual-accuracy tests show errors on hard prompts.

Moonshot's benchmark results

Moonshot reports that K3.1 Pro scores 65.7% on Terminal-Bench 4.0, ahead of GPT-6 Astra (58.2%) and just behind Claude Opus 5.5 (66.4%), and 92.7% with tools on BenchCAD, below Opus 5.5 (96.2%) and GPT-6 Astra (95.9%). On Kimi Hallu Bench Hard, which Moonshot describes as an internal benchmark designed to stress-test factual accuracy, the claim-level hallucination rate fell from 23.4% for K3 to 8.3%. As a stress test, it does not show how often the model errs in everyday use. At K3's launch, Moonshot said the model still trailed the strongest proprietary models; it describes K3.1 Pro as matching or approaching them.

Kimi K3 pricing and access

K3.1 Pro and K3.1 Pro Highspeed are available on kimi.ai and the Kimi apps for iOS, Android and HarmonyOS, in the Kimi Work desktop app from version 3.1.0, in Kimi Code and through the Kimi API as kimi-k3.1-pro and kimi-k3.1-pro-highspeed. Kimi Enterprise administrators must switch K3.1 Pro on for their organization. K3 weights are on Hugging Face; K3.1 Pro weights are due by November 4, 2026.

The API is pay as you go with no subscription. Per million tokens, K3.1 Pro costs $2 for input and $10 for output, Highspeed $4 and $20, and K3 $3 and $15; cached input costs $0.20 for K3.1 models and $0.30 for K3, and cache writes are billed separately. Prices exclude tax. In the Kimi app, the free Adagio plan has limited agent tasks, projects and storage. Paid monthly plans are Plus at $19, Pro at $39, Max at $99 and Ultra at $199, or $15, $31, $79 and $159 per month billed annually. Access to K3.1 Pro Highspeed starts with Pro, and chats with up to 1M tokens of context start with Max.

View image