Kimi K3's full weights are here, but they're 'open' with a caveat: What enterprises should know
Chinese AI startup Moonshot AI, creators of the increasingly famous "Kimi K" family of powerful, open AI models, today released the full weights for its largest and most performant version yet — Kimi K3. But enterprises evaluating the model should read the attached custom Kimi K3 usage license as carefully as the benchmark charts.
VentureBeat previously covered Kimi K3 when it debuted through Moonshot's hosted API earlier this month, including its 2.8 trillion-parameter architecture, one million-token context window and frontier benchmark performance. Today's announcement completes that rollout with the release of the full model weights, a 47-page technical report documenting its training innovations and obstacles, and much of the infrastructure required to run the model independently.
That 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.
While Kimi K3's license grants broad rights to developers and even enterprises to download, modify and deploy the model for commercial purposes — a win for those seeking low-cost, frontier-level AI they can control and run offline — it also imposes some added obligations on larger companies and AI service providers not found in traditional open-source licenses like Apache 2.0 or MIT.
The key restrictions enterprises should understand
Here's the text of the new Kimi K3 License in full:
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.
3. 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.
4. 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.
The most significant clause is 2., as it requires a separate commercial license from companies earning $20 million in annual revenue and operating what Moonshot calls a "Model as a Service," which it defines 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" but not " end-user products with model capabilities solely embedded within specific features or harnesses" or "mere relaying of requests to models hosted by others."
That would seem to allow for free use of the model for most non-tech focused enterprises: banks and consumer brands seeking to use it as a front-end chatbot or customer service agent should be fine, since they're not using Kimi K3 as a "Model as a Service." But firms like they hyperscalers or startups offering model training tools might fall under the commercial license terms.
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 parents into the requirement that they must commercially license the model from Moonshot, pending the parent firm generates revenue of $20 million or more per year as is using it somewhere, in some capacity, as a "Model as a Service."
Regardless of this requirement, in clause 3, the license also requires attribution once 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 prove just as significant as the revenue threshold. Companies that typically abstract away the underlying model may instead need to disclose it directly within their products.
But, if your enterprise is sticking to using Kimi K3 internally as an employee tool for information retrieval, creating spreadsheets and documents, or answering employee questions, Moonshot offers a significant carve out, stating that 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."
Developers immediately zeroed in on the license
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.
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.
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.
What enterprise leaders should do next
For CIOs, chief AI officers and engineering leaders, the first step is to determine how the organization intends to use Kimi K3 before evaluating its technical performance.
If the model will remain entirely inside the organization—for example, supporting developers, researchers, legal teams or internal productivity workflows—the published license appears substantially more permissive. Those deployments may qualify as internal use under Moonshot's terms, avoiding the commercial licensing provisions that apply to customer-facing AI services.
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.
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.
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 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.
