On 27 January 2026 the Beijing AI company Moonshot AI officially released and open-sourced its new model, Kimi K2.5. According to public information, the model is built on a trillion-parameter mixture-of-experts (MoE) architecture and natively trained on 15 trillion mixed vision-and-text tokens, making progress in agent, code-generation and vision-understanding directions.
Open-sourcing becomes a routine move in model competition
The open-sourcing of Kimi K2.5 continues the trend of dense open-sourcing among Chinese large models. For the industry, open source lowers the barrier to innovation at the application layer; for companies, it also spreads the compliance question from “model provider” to “model user”.
Judicial signals have already indicated that open source does not equal exemption from liability. See our notes on the AI open-source ecosystem consensus and the compliance ledger for the Kimi K2.6 open-source model.
Three compliance ledgers to settle
First, the licence’s commercial-use conditions. Open-source licences differ markedly on the scope of commercial use, attribution obligations and how derivative works may be licensed. Companies using open-source weights in downstream products should check the licence clause by clause, avoiding disputes over attribution or relicensing.
Second, downstream services still require filing. Using an open-source model to provide services does not exempt a company from algorithm filing and generative AI service filing. Filing targets the “service”, not the “source of the model”; see our notes on filing and registration for generative AI services and our algorithm filing service.
Third, training data and ownership of results. Multimodal models involve enormous volumes of vision and text data, and the legality of their sources directly affects the model’s usability; at the same time, the ownership of results formed by fine-tuning an open-source model must be clarified in the cooperation agreement. See our IP protection service and data compliance service.
Recommendations for application-layer companies
First, build an open-source component inventory recording the models used, licence types and obligations. Second, treat “filing” as a go-live checklist item rather than a remedy after the fact. Third, in contracts with customers or partners, clarify the model source, data use and ownership of results. See our commercial contracts service.
The spread of open-source models keeps lowering the technical barrier, but the compliance barrier has not fallen with it. Whoever can use open source “cleanly” is more likely to avoid pitfalls in scaled delivery.
This article is based on Moonshot AI’s Kimi K2.5 release information and public reporting. It is for general reference only and does not constitute legal advice.