Moonshot AI Launches Kimi K3, an Open Frontier Model to Take on GPT-5.6 Sol & Fable 5

China’s AI startup Moonshot AI has unveiled Kimi K3, a 2.8-trillion-parameter open-weight AI model with a one-million-token context window. 

“It is the world’s first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning,” the company said in its blog post.

The model is available through Kimi.com, Kimi Work, Kimi Code and the Kimi API, while its full weights are scheduled for release on July 27.

The Chinese AI startup said Kimi K3 is built on its Kimi Delta Attention (KDA) and Attention Residuals (AttnRes) architectures and includes native vision capabilities. 

The company claims the model delivers frontier-level performance in coding, reasoning and knowledge work, although it acknowledged that it still trails proprietary models such as Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol in overall capability.

However, it outperformed OpenAI’s GPT-5.5 and Anthropic’s Opus 4.8 on several benchmarks.

According to Moonshot AI, Kimi K3 is the first open model to reach 2.8 trillion parameters. It uses a mixture-of-experts architecture with 896 experts, activating only 16 experts per token through its Stable LatentMoE framework. The company said the architectural changes improve scaling efficiency by about 2.5 times compared with Kimi K2.

Moonshot AI also highlighted the model’s coding capabilities, saying Kimi K3 can sustain long engineering sessions, navigate large code repositories and use terminal tools with minimal human oversight. 

In internal demonstrations, the company said the model built a Triton-like GPU compiler, designed GPU kernels, created browser-based 3D games, edited videos and reproduced computational astrophysics research by reviewing more than 20 papers and generating over 3,000 lines of Python code.

For enterprise users, Moonshot AI introduced two new features in Kimi Work. Widgets generate interactive components inside chats, while Dashboard organises these components into persistent workspaces connected to local data or external plugins.

The company said Kimi K3 currently operates in maximum-thinking mode by default, with low- and high-effort reasoning modes planned for future updates.