Nathan Lambert argues that Moonshot AI’s Kimi K3 (a 2.8-trillion-parameter MoE model whose weights were later released on July 27) marks a major escalation in the open-weights landscape. K3 ranks near the absolute frontier - roughly #2–3 on major independent indices, behind only Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol - while being dramatically cheaper. This shrinks the previously estimated 6–9 month open-to-closed (and US-to-China) performance gap to something closer to 3–5 months.
Lambert stresses that K3 demonstrates genuine Chinese capability rather than mere distillation of Western models. Moonshot is solving the same hard problems of data, architecture, training, and tooling as leading US labs, just with far less compute. The author notes the strong culture he observed at Moonshot during a visit to China and highlights Chinese labs’ capital efficiency (K3 delivers ~2.5× better scaling efficiency than its predecessor).
China’s open-source bet and the shifting balance of power
He places the current ranking roughly as: Claude Fable 5 > GPT-5.6 Sol > Kimi K3 (open) > Grok 4.5 > GLM-5.2 (open), with several American giants surprisingly lower.
The release, coinciding with Xi Jinping’s public recommitment to open-source AI at the World AI Conference, signals that China currently judges frontier models as low-risk enough to release openly. Beijing prioritizes rapid domestic adoption and global diffusion over the more precautionary stance common in US discourse.
Lambert sees open frontier models as economically “decelerationist” for closed labs (they erode pricing power and margins) but net positive for society: they accelerate diffusion, enable customization, and reduce concentration of power. He argues the healthy equilibrium is for open models to trail closed ones by a few months—providing a safety buffer—while still advancing capability and access. Heavy-handed US restrictions on Chinese open models would be counterproductive, creating asymmetries and slowing beneficial diffusion.
Overall, K3 inaugurates a more competitive, multipolar era in which open-weight models sit close enough to the frontier that coordination, evaluation capacity, and realistic risk assessment become urgent.