Kimi K2 Release! Open-Source Agentic Model!
Related Tools
Described as a Mixture-of-Experts (MoE) model, Kimi K2 features 1 trillion total parameters and 32 billion active parameters.
It positions itself as a leader among open models, boasting state-of-the-art performance on benchmarks including SWE Bench Verified, Tau2, and AceBench.
The model is highlighted for its strong capabilities in coding and agentic tasks, though it notes that multimodal functionality and thought-mode are not supported at this time.
Alongside the model's launch, Moonshot AI emphasized the goal of making advanced agentic intelligence more open and accessible, expressing anticipation for the applications users will develop.
API access to Kimi K2 is now available through https://platform.moonshot.ai, with pricing structured as follows:
- $0.15 per million input tokens for cache hits
- $0.60 per million input tokens for cache misses
- $2.50 per million output tokens
Technical documentation and resources are accessible via a tech blog at github repo, while model weights and code can be found on Hugging Face and GitHub. Users can try Kimi K2 directly at http://Kimi.ai or via the API.
Moonshot AI also shared examples of "vibe tests" conducted with the model, including an interactive 3D mountain scene, a bouncing ball simulation in a hexagon, visual analysis of remote work and salary trends, a 3D particle galaxy simulation, and a Coldplay 2025 concert trip planner. It noted that agent capabilities are available via the API, with more tools set to launch on http://Kimi.ai in the future.
