Moonshot AI Challenges the West with Kimi K3, but Flaws Emerge

Moonshot AI Challenges the West with Kimi K3, but Flaws Emerge

Moonshot AI, a Chinese AI company, has released the model weights and a technical report for Kimi K3. In addition to the model weights available on Hugging Face, the company is also opening parts of its infrastructure, including high-performance attention kernels, a MoE communication library, and tools to operate large-scale AI agents. Moonshot AI claims that the new architecture offers 2.5 times more intelligence per unit of computation. The technical report is available on GitHub.

Since its initial announcement in mid-July 2026, Kimi K3 has generated significant interest by achieving scores close to those of leading Western models such as Fable 5 and GPT-5.6 Sol on popular benchmarks, but at a slightly lower cost and now with open weights. However, an independent test conducted by the Cyber Institute in the UK revealed that the model's cyber capabilities are well below those of leading models. The same goes for its mathematical skills.

These two shortcomings could indicate that Kimi K3 relies on distillation, a technique where a smaller model learns from the outputs of a more powerful model. Chinese models are often faced with this accusation. At the same time, American advocates of open weights increasingly view distillation as a legitimate technique.