Moonshot AI, operating as Kimi.ai, has announced the release of Kimi K2, a new open-source agentic AI model. Launched on July 11, 2025, Kimi K2 features a Mixture-of-Experts (MoE) architecture with 1 trillion total parameters and 32 billion active parameters, designed to make advanced agentic intelligence more accessible. The company stated in its announcement, "With Kimi K2, advanced agentic intelligence is more open and accessible than ever."
Kimi K2 is built on a large-scale MoE model, pre-trained on an extensive 15.5 trillion tokens, utilizing a novel MuonClip optimizer for stable large-scale training. This architecture allows it to efficiently handle complex tasks while maintaining computational efficiency. The model supports long-context inference up to 128K tokens, a significant capability for advanced AI applications requiring extensive contextual understanding.
The new model demonstrates state-of-the-art (SOTA) performance across critical benchmarks among open models, including SWE Bench Verified, Tau2, and AceBench. It exhibits strong capabilities in coding and agentic tasks, excelling in tool use, reasoning, and autonomous problem-solving. According to Moonshot AI's technical reports, Kimi K2 "outperforms or matches Claude Sonnet 4, Opus 4, GPT-4.1, DeepSeek, and Gemini 2.4 Flash" in various evaluations.
Kimi K2 is available in two variants: Kimi-K2-Base, a foundational model for research and fine-tuning, and Kimi-K2-Instruct, a post-trained version optimized for general-purpose chat and agentic applications. The model is accessible via an API with a competitive pricing structure: "$0.15 / million input tokens (cache hit)," "$0.60 / million input tokens (cache miss)," and "$2.50 / million output tokens." This open-source release, with its accessible API and available weights, signifies a move towards more affordable and widespread AI autonomy.
Moonshot AI, a Chinese unicorn backed by major entities like Alibaba, positions Kimi K2 as a significant development in the open-source AI landscape. This release reflects a broader industry trend towards democratizing AI capabilities, enabling developers and researchers globally to build advanced applications. The company expressed anticipation for its impact, stating, "We can't wait to see what you build!"