Moonshot AI launches Kimi K3, a 2.8 trillion-parameter open-weight model, claiming performance near US frontier labs
The AMW Read
K3 is the largest open-weight model ever, matching US frontier performance and upending expectations of US lead timeline; resolves open debate on when Chinese labs would catch up.
Moonshot AI launches Kimi K3, a 2.8 trillion-parameter open-weight model, claiming performance near US frontier labs
Beijing-based Moonshot AI unveiled Kimi K3, the largest open-weight model ever released, with benchmarks showing it competitive against Anthropic's Fable 5 and outperforming Opus 4.8 and GPT-5.6 Sol. The model is priced at $15 per million output tokens, significantly cheaper than Fable 5 at $50. The launch triggered a semiconductor selloff, with TSMC down 7%, SoftBank down 9%, and Nvidia falling 1.2%.
Why it matters: Kimi K3 represents a structural challenge to the hyperscaler distribution moat that US frontier labs have relied on. By releasing the largest open-weight model at a fraction of US pricing, Moonshot accelerates the capital-compression arc that began with DeepSeek, forcing US labs to justify premium pricing. The model's open-weight nature also feeds the acqui-licensing pattern, where downstream players like Cursor (already using Kimi 2.5) can integrate frontier capability without vendor lock-in to Anthropic or OpenAI.
Expert take: Paul Triolo notes China's AI ecosystem is stronger than widely assumed. Anthropic CEO Dario Amodei and Elon Musk had expected a six-month to Q1 2027 timeline for Chinese labs to approach US frontier—K3 compresses that to zero. The open-weight strategy also bypasses potential US export restrictions, as seen with z.AI's GLM-5.2 launch after US briefly restricted Anthropic model access. #MoonshotAI #KimiK3 #openweight #AIcompetition #ChinaAI #frontiermodels
