Microsoft is evaluating Moonshot AI’s Kimi K3 model for use in Copilot as the company looks to reduce the rapidly rising cost of running its artificial intelligence products.
The US technology group is already working to make Kimi K3 available through its Azure cloud platform, according to The Information. Microsoft engineers are also assessing whether the Chinese open-weight model could support parts of Copilot, including its agentic enterprise workflows.
The evaluation comes as Microsoft shifts Copilot Cowork toward metered pricing, under which customers pay according to the number of tokens consumed. That model makes inference costs increasingly important, particularly for applications that perform long, multi-step tasks and generate large volumes of tokens.
Microsoft estimates that moving some Copilot workloads away from models developed by OpenAI and Anthropic and toward Kimi K3 could save as much as $600 million in inference costs.
Kimi K3, recently released by Beijing-based Moonshot AI, has 2.8 trillion parameters and is designed for multimodal and agentic workloads. It supports a context window of roughly 1 million tokens and has posted competitive results against several leading Western models.
The Financial Times
Its headline cost advantage, however, is not straightforward. Kimi K3 reportedly costs about $0.94 per Intelligence Index task, compared with $0.55 for GPT-5.6 Terra and $1.04 for GPT-5.6 Sol when maximum reasoning is enabled. The model can also generate longer reasoning chains, increasing total token consumption.
Its infrastructure efficiency may be more important for Microsoft. Kimi K3 uses a mixture-of-experts architecture and distributes 896 experts across a large number of graphics processors. A smaller key-value cache and other memory optimizations reduce the hardware resources required to process each token, potentially lowering costs when the model is deployed at data-centre scale.
Will cost savings push more US tech companies to adopt Chinese open-source AI models?
Microsoft has previously considered other Chinese open-source models for Copilot. Axios reported in June that the company was examining DeepSeek’s V4 or a similar model that could be hosted on Microsoft’s own infrastructure.
Hosting the models internally would give Microsoft greater control over data, security and deployment, while reducing its dependence on external model providers. It would also allow the company to select different models for different Copilot tasks rather than relying on a single supplier.
Any adoption of Kimi K3 could nevertheless create political tension in Washington. The Trump administration has sought to discourage US companies from relying on Chinese AI technology, citing national-security and economic-competition concerns.
Microsoft therefore faces a trade-off between economics and geopolitics. Chinese open-weight models could materially reduce the cost of operating Copilot, but deploying them inside one of the most widely used US enterprise software ecosystems would likely attract regulatory scrutiny.
