The Chinese AI startup Moonshot AI today released the full model weights for its largest and most performant model yet — Kimi K3 — but enterprises evaluating the technology should read the attached custom usage license as carefully as the benchmark charts. The release completes a rollout that began earlier this month when Kimi K3 debuted through Moonshot’s hosted API, offering a 2.8 trillion-parameter architecture, a one million-token context window and frontier-level benchmark performance. Today’s announcement adds the full weights, a 47-page technical report and much of the infrastructure needed to run the model independently, including optimized attention kernels, MoE communication libraries and deployment components.
The licensing provisions, however, introduce conditions that distinguish this release from conventional open-source frameworks like Apache 2.0 or MIT. While the Kimi K3 license grants broad rights to download, modify and deploy the model for commercial purposes, it imposes additional obligations on larger companies and AI service providers — including revenue thresholds, attribution requirements and a separate commercial negotiation process for certain use cases. For enterprise leaders evaluating whether to adopt Kimi K3, understanding these terms may prove as consequential as assessing its technical performance.
What Moonshot AI Released Beyond the Weights
The full package includes the complete 2.8 trillion-parameter Mixture-of-Experts model, inference infrastructure, optimized attention kernels, MoE communication libraries and deployment components aimed at researchers and enterprise developers who want to self-host the system rather than consume it exclusively through an API. Moonshot is also releasing implementation support for ecosystems including vLLM and SGLang, alongside a technical report detailing the architectural innovations behind the model, including Kimi Delta Attention, Attention Residuals and Stable LatentMoE. Together, those techniques underpin what the company describes as the world’s first open 3T-class model, activating 104 billion parameters from a pool of 896 experts while supporting native multimodal reasoning and a one million-token context window.
At roughly 1.5 TB of model weights, Kimi K3 remains a system aimed primarily at well-resourced organizations capable of operating large-scale inference infrastructure, even as reports emerged of successful deployments on clusters of consumer RTX 5090 GPUs. The release also includes agent tooling and support for fine-tuning workflows, giving enterprise teams a substantial degree of flexibility in how they adapt the model for specific use cases.
The New Kimi K3 License: Key Restrictions Enterprises Should Understand
The full text of the Kimi K3 License, as published by Moonshot, opens with a permissive grant modeled on the MIT license: anyone obtaining a copy of the software — including model weights, parameters, configuration files, inference and training code, and associated documentation — is granted broad rights to use, copy, modify, merge, publish, distribute, sublicense, sell copies, run, deploy, fine-tune, create derivative works, and permit others to do so. But the license then introduces three conditions that depart significantly from traditional open-source frameworks.
What is the Kimi K3 license and what are its commercial restrictions? The Kimi K3 license grants broad rights to use, modify and deploy the model for commercial purposes, but requires companies operating a “Model as a Service” business with aggregate annual revenue exceeding $20 million to enter into a separate commercial agreement with Moonshot AI. Additionally, any commercial product or service using the model that reaches more than 100 million monthly active users or generates more than $20 million in monthly revenue must prominently display “Kimi K3” in its user interface. Internal use of the model — defined as any use that does not make the software, its outputs, or its underlying capabilities available to third parties — is exempt from both requirements.
The most significant clause is Section 2, which requires a separate commercial license from companies earning $20 million in annual revenue and operating 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.” The definition explicitly excludes “end-user products with model capabilities solely embedded within specific features or harnesses” and “mere relaying of requests to models hosted by others.”
This distinction matters greatly for enterprise adoption. Banks, consumer brands, healthcare organizations and other non-tech-focused enterprises that want to use Kimi K3 as a front-end chatbot or customer service agent should generally be fine under the published license, since they are not offering a “Model as a Service” to third parties. But hyperscalers, AI startups offering model training tools, cloud providers building API products, and any company that allows external users to exercise meaningful control over model inputs and training data may fall under the commercial license terms if their revenue exceeds the threshold.
The wording is further notable because it does not limit the revenue calculation to products built on Kimi K3. Instead, it references the aggregate revenue of the licensee and its affiliates, potentially bringing even smaller companies that are part of larger parent organizations into the requirement that they must commercially license the model from Moonshot — provided the parent firm generates $20 million or more per year and is using the model somewhere, in some capacity, as a “Model as a Service.”
Clause 3 adds an attribution requirement once a commercial deployment reaches significant scale: if the software is used for any commercial product or service that has more than 100 million monthly active users, or more than $20 million 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 prove just as significant as the revenue threshold. Companies that typically abstract away the underlying model — presenting it as a proprietary AI assistant or white-labeled capability — may instead need to disclose Moonshot’s branding directly within their products, a consideration that affects contractual commitments, marketing positioning and white-label agreements.
Internal Use Carve-Out Provides a Clear Path for Many Enterprises
For enterprises that intend to keep Kimi K3 inside the organization, Moonshot offers a significant exemption. Clauses 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.” This carve-out means organizations deploying Kimi K3 as an internal employee tool for information retrieval, document creation, data analysis, legal research, software development support or employee Q&A can likely proceed under the permissive MIT-style baseline without triggering commercial license obligations.
The internal use exemption also extends to organizations that access Kimi K3 through Moonshot’s official products or certified inference partners, suggesting that enterprises may have multiple paths to compliant deployment depending on their infrastructure choices and business model.
How Developers and the AI Community Reacted to the License Terms
The licensing provisions quickly became one of the dominant topics of discussion following the release of the model weights. AI researcher Nathan Lambert, previously the co-leader of the Olmo family of open models at AI startup Ai2, summarized the issue in a post on X: “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 noticed as they dug into the newly published license: while Kimi K3 offers unrestricted access to the model weights for researchers, startups and many enterprises, commercial obligations become significantly more complex for larger organizations.
The broader community reaction was largely positive toward the release itself. Developers praised Moonshot for publishing not only the weights but also supporting infrastructure, including attention kernels, MoE communication libraries and agent tooling, viewing the release as a significant contribution to the open-weight AI ecosystem. Others highlighted the rapid pace of ecosystem support, with inference projects such as vLLM and SGLang, along with cloud providers and infrastructure partners, moving quickly to support Kimi K3 deployments.
At the same time, discussion centered on two practical caveats. One was licensing: many developers argued the model is more accurately described as open weight than fully open source, given the commercial conditions attached to larger deployments. The other was operational: at roughly 1.5 TB of model weights, Kimi K3 remains a system aimed primarily at well-resourced organizations capable of operating large-scale inference infrastructure, even as reports emerged of successful deployments on clusters of consumer RTX 5090 GPUs.
Moonshot Is Not the First to Customize Open AI Licensing
Moonshot is far from the first frontier AI developer to embrace an “open, but not completely open” licensing strategy. Meta’s Llama family, for example, has long been distributed under its own community license requiring a commercial agreement for those building with the model and exceeding 700 million monthly users, rather than a traditional open-source software license. Other frontier model developers have likewise adopted bespoke licensing terms governing redistribution, commercial use or attribution.
Kimi K3 follows that broader trend, albeit with 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, while the largest commercial deployments must visibly attribute Kimi K3 within their products. The $20 million annual revenue threshold is notably lower than Meta’s 700 million monthly user trigger, meaning a wider range of commercially active companies could face licensing obligations depending on their business model.
For enterprises, the practical implication is that “open weights” and “open source” are becoming increasingly distinct concepts. Downloading and modifying frontier models may be straightforward; understanding the legal conditions attached to commercial deployment increasingly is not. The gap between technical access and commercial permission is widening, and organizations that fail to account for licensing terms when planning AI deployments may find themselves facing retroactive compliance requirements or renegotiation with model developers after they have already built significant product dependency.
What Enterprise Leaders Should Do Next When Evaluating Kimi K3
For CIOs, chief AI officers and engineering leaders evaluating Kimi K3, the first step is to determine how the organization intends to use the model before assessing its technical performance. The licensing analysis should precede the benchmark evaluation, not follow it.
If the model will remain entirely inside the organization — supporting developers, researchers, legal teams or internal productivity workflows — the published license appears substantially more permissive. Those deployments likely qualify as internal use under Moonshot’s terms, avoiding the commercial licensing provisions that apply to customer-facing AI services. The internal use carve-out covers any use that does not make the software, its outputs or its underlying capabilities available to third parties, which encompasses most employee-facing tools, internal analytics platforms, research environments and enterprise knowledge management systems.
Organizations planning to build products on top of Kimi K3 should take a different approach. 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. This analysis should account for affiliate relationships, since the revenue calculation includes the aggregate revenue of the licensee and its affiliates, potentially extending obligations to subsidiaries, parent companies or sister organizations.
Companies expecting products to reach more than 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. Application developers, enterprise SaaS providers and consumer platform companies that typically mask the underlying model infrastructure may need to redesign user interfaces, update terms of service and negotiate with customers who expect exclusive or proprietary branding.
Legal teams should also consider the jurisdictional implications of the license. The reference to “20 million US dollars (or the equivalent in other currencies)” suggests Moonshot intends global enforcement, and organizations operating across multiple jurisdictions should assess whether their local revenue in any currency triggers the commercial license requirement.
More broadly, Kimi K3 illustrates a transition taking shape across frontier AI. As the industry’s most capable models increasingly become available as downloadable weights instead of exclusively through hosted APIs, enterprises will need to evaluate licensing terms with the same rigor they apply to benchmarks, security reviews and infrastructure planning. The days when “open source AI” meant a single, universally permissive license are receding. In their place, a patchwork of bespoke commercial terms is emerging — tailored by individual developers to protect specific business interests while still enabling broad distribution and adoption.
The next competitive battleground may not simply be whether AI models are open or closed, but the increasingly nuanced legal frameworks that determine who can commercialize them, under what conditions and at what scale. For enterprises building AI strategy around frontier models, the question is no longer just “which model performs best,” but “under what terms can we actually use it, for how long, and at what point does commercial success require renegotiating the deal.”
