The artificial intelligence race is splitting into two competitions, and the one China is winning may matter more than the one America thinks it controls.
On July 16, Chinese lab Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model that is now the largest publicly available AI system ever built. On the Artificial Analysis Intelligence Index v4.1, it scores 57.1 against GPT-5.6 Sol at 58.9 and Anthropic’s Fable 5 at 59.9. That gap is real, but it is the smallest it has ever been between an open-weight model and the best proprietary systems.
The model is built on two novel architectural features, Kimi Delta Attention for long-sequence scaling and a Stable LatentMoE router that activates 16 of 896 experts per token. It has a 1-million-token context window and matches or beats Western frontier models on several coding benchmarks. On the Arena WebDev leaderboard, Kimi K3 ranked number one with a score of 1,679, ahead of both Fable 5 and GPT-5.6 Sol.
The price difference is even starker. Kimi K3 charges $3 per million input tokens and $15 per million output tokens, roughly half the cost of Claude Opus 4.8. Full open weights are promised by July 27, meaning any developer anywhere will be able to self-host the model on their own infrastructure.
The broader story is structural. Eight of the top ten Chinese AI models are now open-weight, downloadable, self-hostable, and commercially usable under permissive licenses. Chinese-developed models account for roughly 30% of all open-model downloads globally on HuggingFace. Alibaba’s Qwen family has surpassed Meta’s Llama in cumulative downloads.
This is a geopolitical shift dressed up as a tech story. American AI dominance has rested on two assumptions: that the best models would stay closed behind paywalls, and that export controls on advanced chips would prevent Chinese labs from competing. Both assumptions are crumbling. China’s labs are publishing their weights openly, buying themselves adoption across the developing world, and increasingly in the West, while US companies keep their most capable models behind API endpoints that governments can shut down or surveil.
The open-weight model from Kimi K3 also arrived with no published system card, no safety documentation, and, according to developer Theo Browne, who tested it extensively, “no safety restrictions.” Browne completed a three-hour autonomous code port, a full security audit of an unlaunched cloud product, and generated a playable 3D game, all of which he said he had never seen from an open-weight system before.
For Washington, the calculus is uncomfortable. Trying to stop Chinese AI through export controls has not worked, the models keep getting better, and they keep getting published. The race is no longer about who builds the smartest model in a San Francisco lab. It is about whose model gets adopted by the rest of the world.