Chinese Open-Weight AI: Kimi K3 Challenges US Frontier Models

The Chinese startup Moonshot AI, supported by the Chinese internet and commerce group Alibaba, has released an open-weight model called Kimi K3 with 2.8 trillion parameters. According to internal benchmarks, it performs exceptionally well, ranking just behind the high-end versions of competitors Claude Fable 5 and GPT-5.6 Sol. In coding tasks, Kimi K3 wins two of six disciplines and ranks at least third in all others. As a universal agent, it performs even better, winning three of six tests, while Fable 5 leads overall. An independent evaluation by the lab Artificial Analysis found similar results: Kimi achieved 57 points on their Intelligence Index, jumping from rank 17 to rank 3, just three points behind Claude Fable 5 and two behind GPT-5.6 Sol. Kimi K3 performed particularly well in agentic and automation tasks, even beating Fable 5 in automation. Its hallucination rate increased significantly compared to the K2 version (51%), but Fable 5 (55%) and GPT-5.6 Sol (85%) performed worse in this metric. This follows DeepSeek's release in January 2025 as another fundamental challenge to the established multimodal AIs from OpenAI, Google, and Anthropic, suggesting that open-source models can compete with western frontier models despite the US chip embargo.