Chinese AI lab Moonshot AI released Kimi K3 on July 16, calling it the world’s first open 3-trillion-parameter-class AI model, according to the company’s official announcement. The model has 2.8 trillion total parameters and is built on a Mixture of Experts (MoE) architecture that activates only 16 of 896 experts per token, keeping inference costs manageable despite the model’s size.
Architecture and capabilities
K3 introduces two new techniques Moonshot calls Kimi Delta Attention and Attention Residuals, designed to improve how information flows through the network as it scales. The model supports a 1-million-token context window and adds native vision understanding, letting it process images alongside text in the same request. Moonshot said K3 was used to build its own tooling during development, including a compiler called MiniTriton that the company says matches the performance of Nvidia’s Triton, and completed a research task that previously took engineers two weeks in about two hours.
Moonshot acknowledges that K3 still trails Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol on overall benchmarks, but says it consistently outperformed other tested open models on its evaluation suite, particularly on long-horizon coding and agentic tasks.
Pricing and availability
K3 is live now through Kimi.com, the Kimi Work desktop app, the Kimi Code terminal tool, and Moonshot’s API, priced at $0.30 per million tokens for cached input, $3 for uncached input, and $15 per million output tokens (as of July 2026). Full model weights, which would let developers download and run K3 on their own hardware, are scheduled for release by July 27.