K3, the open model from Moonshot AI that matches the big players in several tests

Moonshot AI, the Chinese company behind the Kimi assistant, has introduced Kimi K3, an open-source model with 2.8 trillion parameters. According to the company, it is the first open model to reach this scale, and it is designed for coding, document analysis, and tasks that require holding context over long sessions, since it supports up to one million tokens at once.

The company acknowledges that, in overall performance, K3 still trails the most powerful commercial models, such as Claude Fable 5 and GPT-5.6 Sol. However, in several specific coding and office-work benchmarks, its results match or exceed other reference models, including Claude Opus 4.8 and GPT-5.5. In web browsing and automation, for example, it scores highest among all the models compared.

One point the company highlights is the balance between performance and cost: in several benchmarks, K3 achieves results similar to or better than Claude Opus 4.8 or Claude Mythos 5, at a notably lower cost per task. Access through the programming interface costs $0.30 per million input tokens already cached, $3 per million uncached input tokens, and $15 per million output tokens generated.

Regarding its internal design, Moonshot AI explains that the model combines two proprietary attention mechanisms with an expert-based architecture in which only a small portion of the 896 available experts activates for each task, reducing computational cost without losing capability. The company also warns of limitations: the model may make decisions the user did not request during long tasks, and it becomes unstable if switched to mid-session from another model.

Kimi K3 is already available in the Kimi app, in its desktop version for work, in the Kimi Code programming tool, and through the API. The full model weights will be released before July 27, 2026.

Key points

• Moonshot AI introduces Kimi K3, an open model with 2.8 trillion parameters.
• It is the first open model to reach this scale, according to the company.
• It supports up to one million tokens of context in a single task.
• In overall performance it trails Claude Fable 5 and GPT-5.6 Sol.
• It matches or beats Opus 4.8 and GPT-5.5 in several coding and office benchmarks.
• It scores highest in web browsing and automation.
• It offers a better performance-to-cost ratio than Opus 4.8 or Mythos 5.
• API pricing ranges from $0.30 to $15 per million tokens depending on use.
• It may make unrequested decisions during long tasks, the company admits.
• Full model weights will be released before July 27, 2026.