Moonshot’s 2.8 trillion parameter Kimi K3 raises open model pressure

Moonshot AI has released Kimi K3, a 2.8 trillion parameter open-weights model that it says performs near leading systems from OpenAI, Anthropic and Google on a range of benchmarks. The release matters for AI buyers and model providers because it strengthens the case that open models, including Chinese-built systems, are now competitive enough to influence U.S. policy and enterprise procurement.

Moonshot disclosed the model in a technical blog post that included benchmark charts and demos. The company did not disclose revenue, customer adoption, training cost or the compute budget behind Kimi K3. At its stated size, Kimi K3 is beyond what most enterprises could run on their own infrastructure, even if the open-weights release gives developers more access than closed API-only systems.

Kimi K3 is also notable because Moonshot says it is the largest open-weights model built to date. Its 2.8 trillion parameters put it well ahead of U.S. open-weights efforts cited in the same debate, including Thinking Machines Lab’s newly announced Inkling model, described as just under one trillion parameters, and Nvidia’s Nemotron 3 Ultra at 550 billion parameters.

Benchmarks revive the Chinese model debate

Kimi K3 follows a pattern seen with DeepSeek R1, Alibaba’s Qwen family, Z.AI’s GLM models and MiniMax’s M-Series: a Chinese lab publishes a model with benchmark results that appear to compare favorably with the top U.S. closed models. Moonshot’s materials make the same broader point those releases did, that U.S. restrictions on advanced American accelerators have not kept Chinese model developers far from the frontier.

Benchmarks remain an imperfect proxy for commercial usefulness, and Moonshot’s publication does not establish how Kimi K3 performs in production deployments, regulated settings or cost-sensitive workloads. The model’s size may also limit practical use by companies that cannot afford large-scale inference infrastructure.

The national origin of Kimi K3 is likely to draw as much attention as its architecture. Anthropic CEO Dario Amodei has previously framed advanced Chinese AI models as a national security concern. Last month, Amodei accused Alibaba of using Claude to improve its Qwen models through distillation. Anthropic has also agreed to pay $1.5 billion to settle claims that it used millions of pirated books to train Claude.

Policy pressure is building

The U.S. government’s stance toward frontier models has become more active. GPT-5.6 was delayed by the government, and Claude Fable 5 was pulled offline shortly after release while officials reviewed security concerns. One researcher later said the concern was tied to the model’s ability to complete practical software tasks when prompted with requests such as fixing code.

Those episodes have given closed-model labs more material for arguing that open access to advanced systems creates risk. The argument becomes sharper when the model is open weights and developed in China, because U.S. officials have less leverage over the provider than they do over domestic companies such as OpenAI and Anthropic.

White House officials are hearing calls to restrict access to Chinese AI models, according to reports, though no formal action has been taken. Separate reports say the U.S. and China are expected to meet later this year to discuss the risks created by each country’s AI programs.

Restrictions on Chinese models would benefit U.S. labs by reducing competition from some of the strongest open-weights alternatives. For enterprise buyers, the tradeoff would be narrower supplier choice at a time when open models are beginning to pressure closed-model pricing, performance claims and deployment options.

This story draws on original reporting from The Register.