Munsang AI Unveils 'Kimi K3' with 2.8 Billion Parameters... World's Third-Place Performance

Munsang AI has unveiled its flagship model 'Kimi K3' with 2.8 trillion parameters, achieving world third-place performance in benchmark tests against Anthropic's 'Claude 5' and OpenAI's 'GPT-5.5', signaling a shift in the US-centric frontier model competition. The company launched Kimi K3 via its official website, 'Kimi Work', 'Kimi Code', and API, and plans to open-source its full model weights and technical documentation. Notably, Kimi K3 scored 57 points in the Artificial Intelligence Index, trailing only Claude 5 (60) and GPT-5.6 Sol (59), marking China's best-ever performance. It also demonstrated strong capabilities in coding and agent tasks, outperforming Claude 5 in sandbox environments and developing its own compiler 'MiniTriton'. The model secured third place in the GDPval-AA v2 benchmark, outperformed Claude 4.8, and topped the Frontend Code Leaderboard in the Arena benchmark.