Kimi K3 open weights are out: a near-frontier AI model with its own licence

Kimi K3's open weights went public on 27 July 2026, eleven days after Chinese AI company Moonshot AI launched the model as a hosted service. K3 has 2.8 trillion parameters, which the trade publication PureAI describes as the largest open-weight model released so far, and it ships under its own licence instead of a standard open-source one. No small or medium business will run a model this size on its own equipment. The release matters for a different reason: it changes who is able to host a near-frontier AI model, and on what terms.

What happened

Moonshot introduced Kimi K3 on 16 July through its Kimi apps, its Kimi Work desktop program, the Kimi Code tool and its API (the interface that software uses to call the model). In its launch post the company promised the full weights "by July 27, 2026". It met the date. The Kimi K3 repository on Hugging Face, a public library where AI developers publish models, shows its first files on 27 July, and a technical report was published alongside it.

"Weights" are the numbers a model learns during training. Publishing them means anyone with enough hardware can run the model themselves instead of renting it from its maker. That is the difference between an open-weight model such as K3 and a closed model, which can only be used through its maker's own services.

Moonshot is direct about where K3 sits on quality. Its launch post says overall performance "still trails" the two leading closed models, Claude Fable 5 and GPT-5.6 Sol. Independent testing points the same way. PureAI reported that K3 scored 57 on the Artificial Analysis Intelligence Index, placing fourth of 189 models: behind those two, and ahead of models including Claude Opus 4.8 and GPT-5.5.

Key details

The Kimi K3 License is short and worth reading in full if you plan to build on the model. In summary:

• Anyone may use, copy, modify, fine-tune, distribute and sell the model and its code, provided the copyright and permission notice is kept and the use complies with applicable laws.
• A company that runs a "Model as a Service" business (giving third parties access to AI models, for example through an API), and whose revenue together with its affiliates exceeds US$20 million over any 12 consecutive months, must sign a separate agreement with Moonshot before using the model commercially.
• A commercial product with more than 100 million monthly active users, or more than US$20 million in monthly revenue, must display "Kimi K3" prominently in its interface.
• Internal use that does not make the model available to third parties is exempt from both conditions, as is use through Moonshot's own products or its certified inference partners.

Why it matters

The gap between open and closed models keeps narrowing. A model that Moonshot itself places just behind the two leaders can now be downloaded. That puts pressure on prices and gives software makers another serious option to build on.

Open weights mean a choice of host. A closed model runs wherever its maker decides. An open-weight model can be run by any hosting company with the hardware and the right under the licence, in whichever country that company operates. For a business that cares where its data is processed, the same model may become available from several suppliers with different locations and terms.

"Open" does not mean open source. The Kimi K3 License is generous for ordinary users, but it is a custom document with commercial conditions attached. Each open-weight release now tends to arrive with its own terms, and the terms differ.

Running it yourself is a data-centre job. A recommendation of 64 or more accelerators (the specialised chips that run AI models) rules out a server in the back office. In practice, almost every business that uses K3 will do so through someone else's hosted service.

The published scores come mostly from the vendor. PureAI noted at launch that the extra coding results in Moonshot's own launch materials had not yet been independently reproduced. Moonshot also lists limits of its own. The model's answers can become unstable if the software around it does not pass back all of its earlier reasoning, and it can show what the company calls excessive proactiveness, making unexpected decisions when instructions are ambiguous.

What this means for businesses

For most small and medium businesses nothing needs to change this week. The practical effect arrives indirectly: the AI features inside the software you already pay for may start to run on models like this one, and suppliers will have more room to move on price.

The question to keep asking is where your information goes. A hosted model receives whatever you type or upload. If your business is covered by the Privacy Act, the OAIC's guidance on Australian Privacy Principle 8 says you must take reasonable steps to ensure an overseas recipient of personal information does not breach the principles, and you generally remain accountable for what that recipient does with it. That applies to any overseas AI service, whatever country it is in.

A short checklist:

• Ask each software supplier which AI model sits behind a feature, who hosts it and in which country.
• Keep personal, client and commercially sensitive information out of trial accounts until you have read the terms.
• Judge a model on your own work, such as a real quote, a real support email or a real spreadsheet, and not on leaderboard scores.
• If you let a model take actions for you, write down what it may change and what needs a person's approval. Moonshot's own advice is to give K3 explicit constraints.
• If you are building a product on K3, keep a copy of the licence with your project records.

If you would like help comparing suppliers, reading the terms or working out where a hosted service keeps data, our IT consulting and cyber security services cover that kind of review, and cloud migration and hosting covers the question of where your own systems should run. You can ask us for a quote and we reply within one business day.

Key takeaways

• Moonshot AI published the weights of Kimi K3 on 27 July 2026, as promised at the model's 16 July launch.
• At 2.8 trillion parameters it has been described as the largest open-weight model released to date, and it ranks just behind the leading closed models in independent testing.
• The custom Kimi K3 License allows free use, with conditions only for very large hosted services and products.
• Running it needs data-centre hardware, so businesses will meet it through hosted services.
• Where a model is hosted, and under what terms, matters more to a small business than its benchmark scores.

Frequently asked questions

What is Kimi K3?

Kimi K3 is a large AI model from Moonshot AI that reads text and images. Moonshot says it is designed for long-running coding, knowledge work and reasoning tasks. Its weights are public, so other organisations can run it.

Is Kimi K3 open source?

Not in the strict sense. The weights are public and free to use, modify and sell, but under Moonshot's own Kimi K3 License, which adds conditions for large hosted services and very large commercial products.

Can a small business run Kimi K3 on its own servers?

Realistically, no. Moonshot recommends configurations with 64 or more accelerators, which is data-centre equipment. Small businesses will use it through Moonshot's service or another host.

How much does Kimi K3 cost to use?

Moonshot's launch post lists US$3.00 per million input tokens, US$0.30 for cached input and US$15.00 per million output tokens on its own API. Other hosts set their own prices.

Is it safe to put client information into Kimi K3?

That depends on who hosts the version you use and on their terms, not on the model. Check the hosting country, how long data is kept and whether it is used for training before you enter anything personal or confidential.