July 28, 2026, (Inside AI) — Moonshot AI has released the model weights for Kimi K3, a 2.8 trillion parameter language model, under a custom licence that diverges from standard open-source terms. The Beijing-based startup published the weights on July 27 alongside a 47-page technical report detailing training techniques, obstacles, and self-hosting infrastructure.
The release marks a significant moment for enterprise AI, as Kimi K3 has been turning heads in Silicon Valley since its debut earlier this month. It reportedly matches the performance of leading US models like Anthropic's Claude Fable 5 while being cheaper to use, with API pricing at $3 per million input tokens and $15 per million output tokens.
Kimi K3 uses a Mixture-of-Experts architecture and supports a one million-token context window. The technical report includes inference infrastructure details, optimised attention kernels, MoE communication libraries, and deployment components, which could enable researchers and enterprise developers to self-host the model rather than rely solely on Moonshot's API.
"Kimi K3 is a 2.8 trillion parameter model with a one million-token context window and Mixture-of-Experts (MoE) architecture that, until now, has only been accessible via Moonshot's hosted API since making its debut earlier this month." — Moonshot AI technical report.
However, the custom 'Kimi K3' licence imposes obligations not found in permissive licences like MIT or Apache 2.0. For instance, if a licensee or its affiliates operate a 'model as a service' business with annual aggregate revenue exceeding $20 million, they must enter a separate commercial agreement with Moonshot AI before using the software or derivative works commercially. Moonshot defines 'model as a service' as providing third-party access to language model inference or fine-tuning via API, where the third party exercises meaningful control over inputs, parameters, or training data. End-user products with embedded model capabilities and relay services are exempt.
Another requirement targets large-scale commercial users: any licensee with more than 100 million monthly active users or generating over $20 million in monthly revenue must prominently display 'Kimi K3' as a label on the user interface of products or services using the model. This ensures that companies abstracting the underlying model disclose it directly to users. Internal use and usage through Moonshot's official products or certified inference partners are exempt from these obligations.
This licensing approach echoes Meta's Llama models, which also require a commercial agreement for users exceeding 700 million monthly active users. Moonshot's move reflects a growing trend among AI developers to balance openness with commercial control, especially as Chinese open-weight models gain traction globally. Recent data indicates that 30 to 46 percent of enterprise token usage on Chinese open-weight AI models comes from US businesses via platforms like OpenRouter.
Distillation Accusations Shadow Kimi K3's Rise
The model's performance has not come without controversy. Anthropic and US government officials have accused Moonshot AI of using Claude Fable 5 outputs to train Kimi K3 through distillation, a technique where a smaller model learns from a larger one. US officials argue that large-scale extraction constitutes intellectual property theft, distinguishing it from legitimate distillation. Moonshot has not publicly addressed these claims in the technical report, but the allegations could complicate enterprise adoption, particularly for US companies wary of legal and geopolitical risks.
Despite the accusations, demand for Kimi K3 has been overwhelming. Earlier this month, Moonshot AI paused new signups for its API as demand overwhelmed its GPU capacity, underscoring the model's appeal. The release of weights under a custom licence may alleviate some pressure by enabling self-hosting, but the commercial restrictions could limit uptake among large cloud providers and AI service platforms.
For enterprises, the Kimi K3 licence presents a calculated trade-off. The model offers state-of-the-art performance at a fraction of the cost of Western alternatives, but the revenue-based triggers and branding requirements may deter some. Companies using the model internally or through Moonshot's partners face fewer hurdles, making it an attractive option for research and non-commercial applications. The detailed technical documentation further lowers the barrier to self-hosting, a key advantage over fully closed models from OpenAI and Anthropic.
Moonshot's strategy highlights the evolving landscape of open-weight AI, where licensing terms are becoming as critical as model capabilities. As Chinese AI firms continue to close the gap with US leaders, the interplay between openness, regulation, and geopolitical tension will shape enterprise decisions. The full technical report is available on arXiv, offering insights into the model's architecture and training methodology.
