Kimi K3 debuts with 2.8 trillion parameters

Kimi K3 is Kimi’s new flagship model, built with 2.8 trillion parameters, Kimi Delta Attention, Attention Residuals, native visual understanding, and a 1M-token context window. Kimi describes it as the world’s first open-source model in the 3-trillion-parameter class, aimed at long-horizon coding, knowledge work, and reasoning. The full model weights are scheduled for release by July 27, 2026.

The model is designed for sustained engineering tasks, large codebases, terminal-tool coordination, and workflows that combine software engineering with visual feedback, including game development, frontend engineering, and CAD. Kimi also says K3 improves agentic knowledge-work performance in internal evaluations based on recurring user-agent collaboration patterns.

API access requires a successful top-up with a minimum $1. Developers can use the OpenAI SDK with Python 3.9+, call the model as kimi-k3, configure reasoning_effort, stream separate reasoning and final-answer deltas, send image and video inputs, enforce structured output, continue from a prefix with Partial Mode, and use custom or dynamically loaded tools.

K3 always has thinking mode enabled, with reasoning_effort set to low, high, or max by default. Context caching is automatic when the previous request’s prompt tokens exceed 256, while max_completion_tokens defaults to 131072 and can be set up to 1048576. Web search is being updated and is not recommended for production workflows in the near term.