Have Chinese AI models caught up? Kimi K3 reopens the question
Kimi K3's benchmark sheet looks frontier-level — but a closer read of which tests were chosen, and what "the gap" even means, tells a more careful story. An analysis, via Lisan al Gaib.
✓ Verified
Source
Independent analysis (scaling01 · Substack)
⚑ Research
What sparked itA 2.8-trillion-parameter open model
On July 16, 2026, Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight model — roughly twice the size of DeepSeek-V4, and the first open model larger than the original 1.8-trillion-parameter GPT-4 that started the frenzy. Its scale and scores reopened a long-running question: how far behind the US frontier are Chinese models, really?
2.8TTotal parameters (open weights)
≈2×The size of DeepSeek-V4
3rdOn the independent AA Intelligence Index