Microsoft tests Kimi K3 (Chinese model) for Copilot in effort to cut AI inference costs by $600 million.

Microsoft is preparing to test Moonshot AI's Kimi K3 model inside Copilot as the company seeks to reduce its reliance on more expensive models from OpenAI and Anthropic. Microsoft is in the process of adding Kimi K3 to its Azure cloud service, while engineers working on Copilot plan to evaluate whether the model can power features currently running on OpenAI and Anthropic systems, according to The Information.

The potential shift could reduce Microsoft's AI inference costs by as much as $600 million, according to the report. Inference refers to the computing used each time a model receives a prompt and generates a response. Costs can become substantial when an AI service operates across millions of users and processes large volumes of tokens. However, Microsoft has not publicly confirmed the cost estimate or detailed which Copilot features could move to Kimi K3.

Microsoft has previously tested DeepSeek models and earlier versions of Kimi for Copilot, according to a person familiar with its plans. These evaluations indicate Microsoft is exploring a broader mix of models rather than relying on a single provider across every Copilot workload. The tests do not mean Kimi K3 has already been deployed in Copilot. Microsoft engineers must first determine whether the model meets the product's requirements for response quality, reliability, safety and latency.

Moonshot released Kimi K3 on July 16. The Chinese AI startup describes it as a 2.8 trillion parameter model with native multimodal capabilities and a one million token context window, designed for coding, knowledge work and complex reasoning.

Kimi K3's open weight structure could give Microsoft more control over how the model is hosted and optimized on Azure. It could also allow the company to assign less demanding Copilot tasks to cheaper models while reserving more expensive systems for workloads requiring stronger reasoning. Microsoft has increasingly promoted this model selection approach through Foundry, its platform described as model agnostic that encourages developers to select models based on capability, safety, latency and cost rather than sending every request to the most powerful available system. Microsoft's public Foundry pricing page currently lists earlier Moonshot models, including Kimi K2 Thinking, Kimi K2.5 Thinking and Kimi K2.6 Thinking, but Kimi K3 has not yet appeared, supporting reports that its Azure integration remains in progress.

A successful Copilot test would deepen Microsoft's use of Chinese open weight models beyond offering them to Azure customers, placing Kimi K3 directly inside one of Microsoft's flagship AI products while increasing pricing pressure on the US model developers currently powering its features.