AINEWS
Moonshot's Kimi is a free, open source model that benchmarks close to the top paid US AI models. The Trump AI team is publicly split on whether to ban it, compete with it, or let it run.
Two of the Trump administration's top AI advisers are trading public insults about each other and about US AI labs. The dispute under those insults is over what to do about Kimi, a free AI chatbot from Chinese startup Moonshot that is good enough to rival the paid US frontier.
David Sacks last week called Anthropic's models "lobotomized" and "woke" in an interview with The Hill about the rise of Chinese open-source AI. Emil Michael, a senior Pentagon official, separately labeled a new OpenAI executive the "supreme village idiot" in social posts captured in the MIT Technology Review's policy roundup. The Trump AI team is not on message, and the reason it cannot get on message is that nobody on it agrees on what to do about Kimi.
Kimi K3, released last week by Moonshot, is a free, open-weight model, meaning anyone can download, inspect, and run it on their own hardware. Moonshot published it without usage restrictions, alongside benchmark numbers that US labs usually gate behind paid APIs. Independent testing reported by Gizmodo puts it inside the range of paid frontier models from OpenAI and Anthropic, with the usual caveat that benchmark parity and product parity are not the same thing.
OpenAI's ChatGPT Plus and Anthropic's Claude Pro both retail at $20 per month for premium model access. Moonshot released Kimi K3 at no cost, with no usage cap. The price gap is what the public insults are about, because the venture capital underwriting the US frontier labs was priced for a world where the best model is also the most expensive one to run.
The administration's stated position is that it is "weighing" a ban on Chinese AI models, with officials publicly divided and no draft order on the table. Sacks's public comments frame Chinese open-source models as an economic and ideological threat. The more national-security-aligned wing argues a hard ban is both feasible and overdue. Neither side has a way to win the argument with the other, because the underlying policy question is harder than either position admits.
Regulators know how to ban a closed service: block the endpoint, pressure the app stores, the user experience breaks. They have no equivalent playbook for an open-weight model. Once the weights are downloaded, the model can be re-hosted on any cloud, run on local GPUs, and modified downstream in ways that make attribution difficult. Open-source coverage that tracks Kimi's release has been warning for two years that open weights do not respect borders. The US open-source community has been telling regulators the same thing in miniature. The Kimi moment is the first time the question is being asked at a level the White House cannot defer.
Sacks and Michael are not arguing about Kimi specifically. They are arguing about whether the answer to a free, open-weight foreign model is containment, competition, or a decision to outbuild and accept the consequences. No White House has produced a coherent answer. The visible infighting is what policy paralysis looks like when the people writing the policy are also the people talking about it on the record.
The next concrete test is whether the administration produces a draft executive action. If it does, US cloud providers and chip vendors will have to choose which models they will and will not serve, and the precedent will apply to every future foreign open-weight release from any country. If it does not, the Trump AI team is, for now, giving no answer on the record.