Moonshot AI launches 2.8-trillion-parameter Kimi K3

Chinese startup Moonshot AI has launched Kimi K3, a large multimodal model for coding, research and long-running knowledge work. The company calls it the first open model in the three-trillion-parameter class. K3 has 2.8 trillion parameters, native text and image capabilities, and a context window of up to one million tokens. It is Moonshot’s strongest attempt to bring an open-weight system close to leading proprietary models.

Kimi K3 is available through Kimi’s consumer service, Kimi Work, Kimi Code and an OpenAI-compatible API. Moonshot says the full model weights will be released by July 27. Developers will then be able to inspect, adapt and deploy the model outside the company’s hosted products. Open weights offer more control than a closed API, although running a model of this size remains far beyond ordinary local hardware.

K3 uses a Mixture-of-Experts architecture that activates 16 of 896 experts for a task. Moonshot combines this with Kimi Delta Attention and Attention Residuals, techniques designed to improve information flow across long inputs and deep networks. The company claims these changes make scaling roughly 2.5 times more efficient than with Kimi K2. It recommends server configurations with at least 64 accelerators, underlining the infrastructure required.

Moonshot’s own evaluations place K3 near the frontier across programming, tool use, visual reasoning and office work. Associated Press reported that it reached the top of Arena’s front-end coding ranking. These are meaningful signals, but benchmark results depend heavily on test settings, agent harnesses and reasoning budgets. Independent testing after the weight release should give a clearer picture.

Moonshot is unusually candid about the limitations. Its technical blog says K3 still trails Claude Fable 5 and GPT-5.6 Sol overall. The model can become overly proactive on ambiguous tasks, may behave unpredictably if a conversation switches models midway and works best when its full reasoning history is preserved. Makers should set explicit boundaries and test it carefully before allowing access to production systems or sensitive data.

For AI users and businesses, K3 increases competitive pressure on closed providers. Its API costs $3 per million uncached input tokens and $15 per million output tokens, with cheaper cached input. More important than one price comparison is the direction of travel: capable open-weight models are becoming credible foundations for serious agents, research workflows and specialized enterprise tools. Kimi K3 does not settle the open-versus-closed debate, but it makes the choice more substantial.