Moonshot AI, the Beijing-based artificial intelligence startup backed by Alibaba, has introduced Kimi K3, described as the largest open-source AI model built to date. The company launched the model on July 16 through its API and hosted platforms, with full model weights scheduled to become available for public download on July 27 under a Modified MIT license.

Kimi K3 has 2.8 trillion parameters, placing it among the largest publicly available AI models and making it the first public model in the near-3-trillion-parameter range. Parameter count is not a direct measure of model quality, but it is a useful indicator of scale and of the computing resources likely required to run or adapt the system.

Kimi K3’s technical design

The model uses a mixture-of-experts, or MoE, architecture alongside what Moonshot calls Kimi Delta Attention technology. Rather than activating every parameter for each query, the system selects the most relevant subset from its large parameter pool. That design is intended to keep inference costs manageable despite the model’s overall scale.

Kimi K3 also supports a 1 million-token context window, allowing it to process roughly the equivalent of several full-length novels in a single prompt. Moonshot designed the capability for demanding tasks such as long-horizon coding, advanced knowledge work, and deep reasoning.

Under the Modified MIT license, developers can download, modify, and self-host the model freely once the weights are released. That distinction matters because hosted access alone leaves users dependent on the provider’s availability and pricing, while open weights allow organizations with sufficient infrastructure to inspect, adapt, and operate the model on their own systems.

Demand appeared to exceed Moonshot AI’s available capacity soon after launch. The company paused new subscriptions just days after introducing Kimi K3 because usage demand outstripped capacity.

US-China AI context

The release comes amid heightened US scrutiny of Chinese AI development. The Biden administration and the subsequent Trump administration have imposed export controls covering advanced chips shipped to China.

After the model weights become public on July 27, users with sufficient computing resources will be able to run, fine-tune, and deploy K3 regardless of jurisdiction. In practice, the scale of the model means adoption will likely depend not only on download access, but also on hardware availability, serving costs, and whether developers can efficiently integrate it into existing AI workflows.

Implications for AI companies and policy

If Kimi K3 delivers performance comparable with leading proprietary models from OpenAI and Anthropic while being available as a free download, it could complicate the revenue model for companies that charge premium prices for API access.

Moonshot AI’s backing from Alibaba also indicates that the model may continue to receive development, optimization, and support.

The release could draw further attention from Washington. If policymakers view open-source releases of frontier Chinese AI models as a security risk, they may consider restrictions on deployment, hosting, or integration of foreign-developed AI systems within US infrastructure. Key details to watch after the weight release include independent benchmark results, real-world deployment reports, and whether Moonshot provides further documentation or tooling to support self-hosted use.