Kimi K3 and Qwen3.8 pressure the US lead with open weights

Moonshot and Alibaba introduced Kimi K3 and Qwen3.8, models they claim approach the top systems from OpenAI and Anthropic. The bigger issue is distribution: China is pushing very large models toward open weights while the US frontier remains closed.

Moonshot and Alibaba put trillion-parameter models into the race

The Verge describes a two-part Chinese push: Moonshot AI unveiled Kimi K3 and Alibaba followed with a preview of Qwen3.8. Moonshot describes Kimi K3 as an open-source system with 2.8 trillion parameters. Alibaba, according to The Verge, says Qwen3.8 has 2.4 trillion parameters and is headed for an open-weight release.

Both companies claim their models are approaching the best closed systems from Silicon Valley. Kimi K3’s full weight release is expected on July 27, according to the article. Qwen3.8 is described more loosely: open weights are coming soon.

That needs caution. Parameters are not performance and vendor benchmarks are not reality. The Verge itself notes that the models will be hard to assess until they are fully released and independently tested.

Open weights shift geopolitics more than raw scores do

The strategic point is not only whether Kimi beats a particular Claude or GPT version. Open weights enable local deployment, fine-tuning, auditing and derivatives without routing every request through a foreign lab’s API.

For states, companies and research teams, that is a different kind of leverage. A closed model may be the better service. An open-weight model can become infrastructure that someone adapts, cheapens, speeds up or connects to private data without vendor permission.

The numbers look strong, but independent tests still matter

The main risk is overclaiming. Chinese models can be very good and still sit months behind the absolute closed frontier. Zvi Mowshowitz flags spotty access, token hunger and the possibility that Kimi K3’s benchmarks overstate practical performance.

Claims about cost need the same caution. “Fraction of the cost” is a powerful line, but without consistent measurement of price, latency, quality and reasoning length it can mislead. A cheaper model that needs more tokens or more time may not be cheaper in production.

Repos, API logs and outside evals will settle the first round

The next signals are simple: actual weights, licensing terms, inference cost and independent evals. If Kimi K3 and Qwen3.8 hold up outside marketing charts, US labs will face pressure on openness as well as capability.

If they mostly dominate narrow benchmarks or prove expensive to run, this will be less revolution than another round of fast following. Even that would be enough to make Silicon Valley stop treating the closed frontier as a permanent wall.

Lilith's verdict

China is not just placing a stronger model on the board. It is drawing escape routes for everyone tired of waiting at the American API gate.