Open source AI models are significantly closing the performance gap with closed source models from companies like OpenAI, Anthropic, and Google. Earlier this month, Beijing-based Dark Side of the Moon Moonshot AI officially released the open source model Kimi K3, which has demonstrated superiority over several cutting-edge models in code writing, document processing, and agent benchmark tests.
Microsoft announced that Kimi K3 is now officially available on the Microsoft Foundry platform through Fireworks AI. With this integration, Azure customers can now take advantage of Fireworks AI's inference infrastructure optimized for open source models, while leveraging Microsoft Foundry to provide enterprise-grade deployment, security, monitoring and governance services. As early as March this year, Microsoft announced the public preview of support for Fireworks AI models on Azure. Fireworks' high-throughput inference stack will handle seamless requests, ensuring excellent performance while taking into account Azure-level management specifications.
As a Mixture-of-Experts model with 2.8 trillion parameters, Kimi K3 has native multi-modal capabilities and supports context windows of up to 1 million tokens. During inference, the model activates only 16 of the 896 experts on each token, effectively reducing the required computational overhead. With a large contextual window, developers can use the model to easily analyze entire code bases, lengthy legal documents, vast amounts of research material, and even hours of meeting transcripts.
In various benchmark tests, Kimi K3 performed equally well: it scored 81.2 points in the FrontierSWE test, surpassing GPT-5.6 Sol and Claude Opus 4.8. At the same time, it also achieved leading results in tests such as BrowseComp, SWE Marathon, AutomationBench, SpreadsheetBench 2, and OmniDocBench.
Currently, Microsoft Foundry customers can deploy Kimi K3 through Fireworks AI’s Data Zone. Its pricing standards are: $3.30 per million input tokens, $0.33 per million cached input tokens, and $16.50 per million output tokens.
It's worth noting that the Kimi K3 does not come under the permissive modified MIT license used by earlier Kimi models. For example, when a company's operating revenue exceeds $20 million in a 12-month period, if the company provides Kimi K3's inference or fine-tuning services to third parties, it must separately negotiate and enter into an agreement with Dark Side of the Moon.