Chinese AI startup Moonshot AI has released the full weights for Kimi K3, the largest and most capable version yet in its increasingly prominent Kimi K family of open AI models. The release gives researchers and enterprise developers access to the model for self-hosting, but companies evaluating it must also examine the custom Kimi K3 usage license closely.

Kimi K3 previously debuted through Moonshot's hosted API earlier this month. VentureBeat covered that launch, including the model's 2.8 trillion-parameter architecture, one million-token context window and frontier benchmark performance. The latest announcement completes the rollout with the publication of the full model weights, a 47-page technical report describing training innovations and obstacles, and much of the infrastructure needed to operate the model independently.

The release includes the complete 2.8 trillion-parameter Mixture-of-Experts model, inference infrastructure, optimized attention kernels, MoE communication libraries and deployment components for researchers and enterprise developers that want to self-host Kimi K3 rather than access it only through an API. Moonshot's announcement is available at and the technical report is available at https://github.com/MoonshotAI/Kimi-K3/blob/main/k3_tech_report.pdf.

Moonshot is also providing implementation support for ecosystems including vLLM and SGLang. The accompanying technical report describes architectural innovations behind the model, including Kimi Delta Attention, Attention Residuals and Stable LatentMoE. Moonshot says those techniques support what it describes as the world's first open 3T-class model, which activates 104 billion parameters from a pool of 896 experts while supporting native multimodal reasoning and a one million-token context window.

The Kimi K3 license gives developers and enterprises broad rights to download, modify and deploy the model for commercial purposes. That may appeal to organizations seeking lower-cost, frontier-level AI that they can control and run offline. However, the license also adds obligations for larger companies and AI service providers that are not found in traditional open-source licenses such as Apache 2.0 or MIT. Those conventional software licenses generally allow commercial use, redistribution and modification without revenue-based negotiations or product attribution requirements, which is why the precise wording of AI model licenses has become a central procurement issue.

Key restrictions in the Kimi K3 license

The Kimi K3 license is published at Its core text states:

"Permission is hereby granted, free of charge, to any person (the "Licensee") obtaining a copy of this software — including the model weights, parameters, configuration files, inference and training code, and associated documentation (collectively, the "Software") — to deal in the Software without restriction. This includes, without limitation, the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software; to run, deploy, fine-tune, or otherwise modify the Software and create derivative works from it; and to permit persons to whom the Software is furnished to do so, in each case subject to the following conditions:

1. The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. Licensee's use of the Software must comply with applicable laws and regulations.

2. "Model as a Service" means giving a third party access to language model inference or fine-tuning (e.g., via API) in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data. This does not include (a) end-user products with model capabilities solely embedded within specific features or harnesses, or (b) mere relaying of requests to models hosted by others.

If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose.

1. If the Software (or any derivative works thereof) is used for any of the Licensee's commercial products or services that have more than 100 million monthly active users, or more than 20 million US dollars (or equivalent in other currencies) in monthly revenue, "Kimi K3" must be prominently displayed on the user interface of such product or service.

2. The requirements set forth in Sections 2 and 3 do not apply to: (a) internal use of the Software, defined as any use that does not make the Software, its outputs, or its underlying capabilities available to third parties; or (b) any use of the Software accessed through Moonshot AI's official products or certified inference partners."

Clause 2 is the most significant for many enterprises. It requires a separate commercial agreement from companies that earn more than $20 million in aggregate revenue over any consecutive 12-month period and operate what Moonshot calls a "Model as a Service." The license defines that term as "giving a third party access to language model inference or fine-tuning (e.g., via API) in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data." It excludes "end-user products with model capabilities solely embedded within specific features or harnesses" and "mere relaying of requests to models hosted by others."

That language appears to leave room for many non-technology-focused enterprises to use Kimi K3 without a separate commercial agreement. Banks or consumer brands using the model as a front-end chatbot or customer service agent would not necessarily be operating Kimi K3 as a "Model as a Service." By contrast, hyperscalers or startups offering model training tools may fall within the commercial license terms.

The wording is also notable because the revenue calculation is not limited to products built on Kimi K3. Instead, it refers to the aggregate revenue of the licensee and its affiliates. That could bring smaller companies owned by larger parent organizations into the requirement to obtain a commercial license from Moonshot if the parent organization generates at least $20 million per year and the software is used somewhere, in some capacity, as a "Model as a Service."

Clause 3 separately requires attribution when a commercial deployment reaches significant scale: "If the Software... is used for any of the Licensee's commercial products or services that have more than 100 million monthly active users, or more than 20 million US dollars in monthly revenue, 'Kimi K3' must be prominently displayed on the user interface."

For enterprise software vendors, AI copilots and consumer applications, that branding requirement may be as important as the revenue threshold. Companies that usually abstract away the underlying model may need to disclose it directly within their products. That is a practical product-design issue as well as a legal one, because attribution language can affect user interfaces, customer contracts and white-label deployments.

Moonshot provides a carve-out for internal enterprise use. If an organization uses Kimi K3 only as an employee tool for information retrieval, spreadsheet and document creation, or employee questions, clauses 2 and 3 may not apply. The license says those sections "do not apply to: (a) internal use of the Software, defined as any use that does not make the Software, its outputs, or its underlying capabilities available to third parties."

Developers focused quickly on the license

The licensing provisions became a major topic of discussion after the model weights were released. AI researcher Nathan Lambert, previously co-leader of the Olmo family of open models at AI startup Ai2, summarized the issue in a post on X at https://x.com/natolambert/status/2081760901020201086?s=20:

"Kimi K3 license. It's inspired by MIT but distinctly non-commercial, where any company making over $20M/yr must get a specific commercial deal (and display Kimi K3 if over 100M users or $20M/mo revenue)."

Lambert's assessment reflected what many developers observed after reviewing the newly published license. Kimi K3 provides unrestricted access to model weights for researchers, startups and many enterprises, but commercial obligations become more complex for larger organizations.

The broader community reaction to the release itself was largely positive. Developers praised Moonshot for publishing the weights along with supporting infrastructure, including attention kernels, MoE communication libraries and agent tooling, and described the release as a significant contribution to the open-weight AI ecosystem.

Others noted the pace of ecosystem support, with inference projects such as vLLM and SGLang, as well as cloud providers and infrastructure partners, moving quickly to support Kimi K3 deployments.

Discussion also centered on two practical caveats. The first was licensing: many developers argued that Kimi K3 is more accurately described as an open-weight model than a fully open-source model because of the commercial conditions attached to larger deployments. The second was operational. At roughly 1.5 TB of model weights, Kimi K3 remains aimed primarily at well-resourced organizations capable of running large-scale inference infrastructure, even as reports emerged of successful deployments on clusters of consumer RTX 5090 GPUs.

Moonshot follows a broader custom-license trend

Moonshot is not the first frontier AI developer to adopt an "open, but not completely open" licensing approach. Meta's Llama family, for example, has long been distributed under its own community license rather than a traditional open-source software license. That license requires a commercial agreement for those building with the model and exceeding 700 million monthly users. Meta's Llama license is available at

Other frontier model developers have also adopted bespoke licensing terms governing redistribution, commercial use or attribution. Kimi K3 follows the same broad pattern but uses a different mechanism. Rather than broadly restricting redistribution, Moonshot ties certain commercial rights to company scale.

Organizations operating a Model-as-a-Service business above specified revenue thresholds must negotiate a separate commercial agreement. The largest commercial deployments must also visibly attribute Kimi K3 within their products.

For enterprises, the practical distinction between "open weights" and "open source" is becoming more important. Downloading and modifying frontier models may be straightforward, but the legal conditions attached to commercial deployment can be more complicated. The distinction also affects vendor due diligence: teams comparing models now have to weigh not only capability, latency and infrastructure cost, but also whether the license permits the intended customer-facing use case without additional approval.

Enterprise considerations

For CIOs, chief AI officers and engineering leaders, the first step is to determine how the organization plans to use Kimi K3 before assessing its technical performance.

If the model will remain entirely inside the organization, such as for developers, researchers, legal teams or internal productivity workflows, the published license appears more permissive. Those deployments may qualify as internal use under Moonshot's terms and avoid the commercial licensing provisions that apply to customer-facing AI services.

Organizations planning to build products on top of Kimi K3 face a different analysis. Legal, engineering and product leaders should determine whether the planned deployment constitutes "Model as a Service" under the license, whether the company or its affiliates exceed the $20 million revenue threshold, and whether future growth could trigger the requirement to negotiate a commercial agreement with Moonshot.

Companies expecting products to exceed 100 million monthly active users or more than $20 million in monthly revenue should also evaluate the license's attribution requirement and how it fits with existing branding, contractual commitments and white-label offerings.

Kimi K3 highlights a broader shift in frontier AI. As more highly capable models become available as downloadable weights rather than only through hosted APIs, enterprises will need to evaluate licensing terms with the same rigor they apply to benchmarks, security reviews and infrastructure planning. The distinction between open and closed AI models is increasingly shaped by legal frameworks that determine who can commercialize them, under what conditions and at what scale.