Moonshot is Chinese But Its AI Models Are From Another Planet

Performance and Benchmarks:

Moonshot's self-reported scores and independent analysis place Kimi K3 at or near the top of leaderboards, competitive with models like Anthropic's Mythos/Fable and OpenAI's GPT-5.6, and a step above Opus-4.8 and GPT-5.5. The jump from its predecessor, Kimi K2.6, is described as huge.

• Artificial Analysis: The firm's evaluation shows K3 is only slightly behind Fable and GPT-5.6. It highlights K3's cost-effectiveness, placing it in its "most attractive quadrant" for enterprise API users. Cost per task is $0.94 for Kimi K3, compared to $1.04 for GPT-5.6 Sol and $1.80 for Opus 4.8.
• Arena: An initial viral result showed K3 as "considerably better than Fable on frontend coding," though the author notes the chart's x-axis was truncated to exaggerate the difference. On Arena's overall text ranking, K3 falls considerably lower.
• Other Benchmarks: Kimi K3 ranked #1 on creative writing, #1 on Vercel’s Next.js agentic benchmark, and #3 on deepSWE, a long-horizon software engineering benchmark.

Architecture and Efficiency:

Kimi K3's capabilities are attributed to its large scale combined with novel efficiency techniques, developed in response to hardware constraints.

• Size: Kimi K3 has 2.8 trillion parameters, making it the first open-source model in the 3T parameter range.
• Architecture: It is an extremely sparse Mixture of Experts (MoE) model with almost 900 experts, of which only 16 are active for any given task.
• Efficiency Techniques: Moonshot implemented custom changes including Kimi Delta Attention and Attention Residuals. The training process was improved with INT4-native quantization and inference with expert parallelism.
• Overall Efficiency: These advances make Kimi K3 approximately 2.5 times more scale-efficient (at converting energy to intelligence) than its predecessor, Kimi K2.

Distillation Controversy:

Anthropic has accused Moonshot of using its models to improve their own, a practice known as distillation.

• Anthropic's Claims: In a February blog post, Anthropic CEO Dario Amodei stated they detected "industrial-scale campaigns" from Chinese actors. They specifically attributed an attack of 3.4 million exchanges to Moonshot, using hundreds of fraudulent accounts to extract data and reasoning traces from Claude models. Request metadata reportedly matched the public profiles of senior Moonshot staff.
• Counterarguments: The article notes that K3 outperforms teacher models on some benchmarks, which distillation alone cannot explain. It cites several Western AI figures, including OpenAI researchers, Replit CEO Amjad Masad, and researcher Nathan Lambert, who believe the performance reflects genuine innovation and talent. Lambert noted Chinese labs' willingness to do "grunt work," having less ego, and being open to new techniques.
• Conclusion on Controversy: The author suggests both narratives can be true: Chinese labs may be using distillation while also achieving legitimate breakthroughs.