Kimi K3: Moonshot AI's open-weight model explained

Updated October 25, 2026. An earlier version of this article described Kimi K3 with generic claims (a 2025 release, MATH and HumanEval leadership, comparisons with o1) that no source supports. The text below has been rewritten from Moonshot AI's model card and press coverage of the launch.

Kimi K3 in short

Kimi K3 is a large language model from Moonshot AI, a Chinese start-up. Moonshot unveiled it on July 16, 2026, as reported by Fortune, and then published the weights on Hugging Face. According to the official model card, it is a Mixture-of-Experts model with 2.8 trillion parameters, of which 104 billion are activated for each token (16 experts out of 896), with a context window of about one million tokens. Fortune called it the largest open-weight model available at launch.

What the model card says

• Architecture: 93 layers combining Kimi Delta Attention and Gated MLA, plus a MoonViT-V2 vision encoder. Weights are quantized in MXFP4 with quantization-aware training.
• Inputs: text and images, according to the specification table.
• Reasoning: thinking is always enabled. The API returns a separate reasoning field that must be passed back in multi-turn and tool-calling conversations.
• Access: the API is available on platform.kimi.ai under the name kimi-k3, with OpenAI- and Anthropic-compatible endpoints. For self-hosting, Moonshot recommends vLLM, SGLang or TokenSpeed.
• Price: Fortune reports $15 per million output tokens, versus $4.40 for Z.ai's GLM-5.2, which makes K3 expensive by Chinese standards.

Performance: promising, but self-reported

The figures published by Moonshot place Kimi K3 close to the best closed models of the moment. On GPQA Diamond, it scores 93.5, against 92.6 for Claude Fable 5 and 94.1 for GPT-5.6 Sol. On BrowseComp, a web research benchmark, it leads with 91.2. On the other hand, it trails Claude Fable 5 on FrontierSWE (81.2 vs 86.6) and OfficeQA Pro (63.3 vs 69.9). These scores come from Moonshot itself, sometimes on in-house benchmarks. Fortune points out that the performance claims had not been independently verified at the time of launch.

The license: open weights, with conditions

Kimi K3 is not released under a standard open-source license but under the Kimi K3 License. Two clauses matter for companies:

• A business that offers the model as a service and whose revenue, with its affiliates, exceeds $20 million over 12 consecutive months must sign a separate agreement with Moonshot AI before commercial use.
• A product with more than 100 million monthly active users or more than $20 million in monthly revenue must display "Kimi K3" prominently in its interface.

Purely internal use, which does not make the model or its outputs available to third parties, is exempt from both clauses. Have your legal team read the license before any external deployment.

What it means for a marketing team

For most marketing teams, Kimi K3 will be used through the API or a third-party tool rather than self-hosted, given the size of the model. Its strengths on paper suggest a few uses:

• Long document analysis: the one-million-token window lets you work on a full brand platform, a set of studies or a large brief corpus in a single conversation.
• Web research and monitoring: its BrowseComp score points to good performance on multi-step research tasks, useful for competitive monitoring.
• Agentic workflows: compatibility with existing APIs makes it easy to test in tools that already use OpenAI or Anthropic models.

The limits are just as concrete. Permanent reasoning increases latency and output token consumption, which weighs on cost at $15 per million output tokens. Before sending customer data or confidential documents, check where data is processed and what the provider's terms say: our guide on cloud AI vs local AI sets out the criteria. And as with any model, generated content must be reviewed by a human before publication, within a clear AI governance framework.

Kimi K3 among open-weight models

Kimi K3 arrived in a busy summer for open-weight models. Z.ai had released GLM-5.2 in June, and a Semgrep benchmark on vulnerability detection had shown it ahead of Claude Code on one specific task. In the United States, Thinking Machines Lab released Inkling on July 15, 2026, under the Apache 2.0 license, which is less restrictive than the Kimi K3 License. This article describes Kimi K3 as published in July 2026: check which version is available before deploying it.

About Brandeploy

Language models like Kimi K3 speed up the production of text and analysis, but they do not guarantee that the visuals published across your channels respect your brand. Brandeploy is a creative automation platform that helps marketing teams produce and adapt their assets, from banners to localized versions, while keeping a consistent brand identity. Book a demo to see the Brandeploy platform in action.

FAQ

What is Kimi K3 and who developed it?

Kimi K3 is a large language model from the Chinese start-up Moonshot AI, unveiled on July 16, 2026. According to its model card, it is a Mixture-of-Experts model with 2.8 trillion parameters, 104 billion of them activated per token, and a context window of about one million tokens.

Is Kimi K3 open source?

Its weights are published on Hugging Face, but under the Kimi K3 License rather than a standard open-source license. Companies offering it as a service with more than $20 million in revenue over 12 months need a separate agreement with Moonshot AI, and very large products must display "Kimi K3" in their interface. Purely internal use is exempt from these two clauses.

How does Kimi K3 compare with Claude and GPT?

On Moonshot's own figures, Kimi K3 scores 93.5 on GPQA Diamond, against 92.6 for Claude Fable 5 and 94.1 for GPT-5.6 Sol, and leads BrowseComp with 91.2. It trails Claude Fable 5 on FrontierSWE and OfficeQA Pro. These scores are self-reported and were not independently verified at launch.

How much does Kimi K3 cost?

Fortune reported a price of $15 per million output tokens at launch, which is high compared with other Chinese models such as GLM-5.2 at $4.40. Since thinking is always enabled, output token consumption should be monitored. Check the current rates on Moonshot's platform before any project.

How can a marketing team use Kimi K3?

Mainly through the API or a third-party tool, for long document analysis, multi-step web research or agentic workflows. Before sending confidential data, check where it is processed and the provider's terms, and keep human review of all generated content before publication.