Kimi K3.1 Pro Has Crossed China’s Export Threshold

What Moonshot AI actually released

I would view Kimi K3.1 Pro from two angles at once: it is a very large MoE model and, potentially, an object of export control. Moonshot AI's Hugging Face model card lists 2.8 trillion parameters, with 104 billion activated for each token. The stated context window reaches 1,048,576 tokens.

The distinction between total and active parameters is crucial here. A Mixture of Experts architecture does not turn every inference step into computation equivalent to a dense 2.8-trillion-parameter model, yet the complete set of weights remains enormous. This is where a technical specification unexpectedly becomes a legal question.

K3.1 Pro is described as a multimodal model supporting text, images, and video through its intended interfaces. Moonshot's API documentation also lists low, high, and max reasoning levels, while key use cases include long-context work, coding, tool use, and agentic task execution.

According to Moonshot's own results, the model scored 65.7% on Terminal-Bench 4.0. That is a notable figure, but I would still treat it as a developer claim: independent validation is limited, and some reported evaluations rely on internal datasets. Full reproducibility is also unavailable because the published materials do not establish that all training code and data composition have been released.

The export threshold changes access rules

The main risk is that Kimi K3.1 Pro formally exceeds China's proposed 2-trillion-parameter threshold. If regulators count every MoE parameter, the model meets the criterion; if they count only the 104 billion active parameters, the outcome changes.

As of November 4, 2026, available reports describe a draft or planned control system rather than an automatic worldwide ban. The treatment of APIs, weights, checkpoints, and technical support remains unclear, as does the precise licensing process.

For developers outside China, this could mean a shift from open distribution toward managed cloud access, regional interfaces, and user screening. The consequences may be even more difficult for researchers: fewer available weights, harder benchmark reproduction, and wider differences between regional model versions.

My engineering conclusion is straightforward: 2.8 trillion parameters are impressive, but model size is not the real story. Everything depends on which number the regulator ultimately recognizes as the size of an MoE system.

We previously covered Pony Alpha, a likely GLM-5 model available through OpenRouter with a 200K context window. Its release context helps frame how Chinese frontier models are reaching users through different channels amid export-control constraints.