Moonshot AI is preparing to release the full open weights for Kimi K3 on July 27, giving developers a chance to download, run, and fine-tune the model locally. The planned release stands out because Kimi K3 has 2.8 trillion total parameters.
Before the full weights become available, users can already test Kimi K3 without a developer account. The model is accessible through the Kimi app on iOS and Android, the kimi.com website, and the Kimi Work desktop application.
Free access is subject to usage limits, but it offers a way to assess the model before the open-weight release. Developers seeking service integration can instead use Kimi K3 through its API.
| Service | Price | Availability |
|---|---|---|
| Free access | Usage-limited | Kimi apps, kimi.com, Kimi Work |
| API input | $3 per million tokens | API access |
| API output | $15 per million tokens | API access |
Moonshot AI says API pricing starts at $3 per million input tokens and $15 per million output tokens. The company also provides discounts through a caching mechanism.
A Massive Model Built Around MoE
Moonshot AI describes Kimi K3 as the largest AI model ever released with open weights. Its scale is managed through a Mixture-of-Experts, or MoE, architecture rather than by activating every part of the model for each request.
Kimi K3 activates 16 experts out of a total of 896 experts for an individual request. This means the model’s full 2.8 trillion-parameter capacity is not used in every response.
| Specification | Kimi K3 Detail |
|---|---|
| Total parameters | 2.8 trillion |
| Architecture | Mixture-of-Experts |
| Active experts per request | 16 of 896 |
| Context window | Up to 1 million tokens |
| Input support | Images and reportedly video |
The model also supports a context window of up to 1 million tokens, intended for long conversations and extensive source material in a single session. Beyond text, Kimi K3 supports image input and is reportedly equipped for video input as well.
Moonshot AI has added Kimi Delta Attention, a technology it says helps process long conversations more efficiently. The feature is relevant to a model designed to work with unusually large volumes of context.
Strong Benchmarks, With Product Work Remaining
Early benchmark results place Kimi K3 in a competitive position across several categories. It is still said to trail leading Western models, including Claude Fable 5 and GPT-5.6, in certain areas.
At the same time, Kimi K3 is reported to match or outperform other models in coding, agentic workflows, and knowledge-based work. Independent testing from Artificial Analysis also places the model in a strong overall position.
Gizmochina reported that Moonshot AI acknowledges it still needs to improve the overall user experience. Even so, the full open-weight release could create more room for self-hosting and customization among developers and AI enthusiasts.
The planned release also highlights the growing role of China’s AI ecosystem beyond the Western companies that have dominated advanced model development. For the open-source community, Kimi K3 may become a new option for experimenting with AI at a very large scale.
