Microsoft weighs Moonshot's Kimi K3 for Copilot

Microsoft is evaluating Moonshot AI's open-source Kimi K3 model for some Copilot functions as it looks for lower-cost alternatives to systems from OpenAI and Anthropic. According to reporting by The Information, the software company has already started adding Kimi K3 to certain Copilot features, with internal projections cited in coverage suggesting the transition could yield up to $600 million in inference-cost savings.
Kimi K3 architecture and capabilities
The Kimi K3 model is a multi-modal, open-source architecture featuring 2.8 trillion parameters and a 1-million-token context window. According to technical descriptions in industry reporting, the model uses specialized optimizations for data center hardware, positioning it as a candidate for large-scale enterprise deployment rather than only experimental use.
Kimi K3 climbs the AA-Briefcase benchmark
Kimi K3, developed by Moonshot AI, has been recognized as the second-best AI model on the AA-Briefcase benchmark, trailing only Claude Fable 5 from Anthropic. Despite its strong benchmark showing, the model faces significant operational cost challenges that industry coverage suggests could limit near-term deployment economics for some users.
Routing Kimi K3 with Fable 5 sets a new state of the art
Benchmarking by Fireworks AI over more than 1,000 agentic tasks demonstrates that routing between the open-source Kimi K3 model and the closed Anthropic Fable 5 model creates a new state-of-the-art approach that surpasses the performance of either model alone. Fireworks describes Kimi K3 as a frontier-quality open model available at a fraction of the cost of closed alternatives, and notes that the two systems complement each other predictably enough to enable task-based routing.
Open-source competition reshapes the AI landscape
Recent industry commentary has framed Kimi K3 as the latest Chinese-developed model to trigger concern about US AI dominance, drawing explicit comparisons to the market reaction that followed the release of DeepSeek R1. Coverage emphasizes that, regardless of national origin, open-weight models are now demonstrably competitive with leading closed systems on both benchmark scores and real agentic workloads.
Follow-up signal: Microsoft has not publicly confirmed a Copilot-wide Kimi K3 rollout; further announcements, integration details, or pricing commitments from Microsoft, Moonshot AI, or Anthropic are the next verifiable milestones to watch.
Share this article







