Moonshot AI's Kimi K3: Largest open model at 2.8T, but still behind closed frontier

Moonshot AI's Kimi K3 is a 2.8 trillion parameter open-weight model, making it the largest open model to date. It achieves excellent benchmark scores, outperforming other open models, especially in agentic coding and 3D tasks. However, analysis suggests the model is roughly 4-6 months behind the closed frontier (Anthropic's Claude Opus, OpenAI's Mythos). Distillation from Claude, likely using Fable, contributes to its performance, but independent gains also come from its sheer size. The model is slow and token-hungry, trading efficiency for capability.

Despite strong benchmarks, practical performance is uneven and likely overestimated due to benchmark-specific tuning. Access is still limited. Moonshot AI is capitalizing on the buzz, planning an IPO in Hong Kong within six months. While Kimi K3 fills a niche for certain workflows, it won't replace smaller open models or top closed models in most scenarios. The release reignites debates about Chinese AI catching up to American leaders, but careful analysis shows the gap remains significant.