MoonshotAI: Kimi K2 0905

MoonshotAI's Kimi K2 0905, an update to the Kimi K2 0711 model, demonstrates a strong performance profile, particularly in reliability and specific task categories. With 1 trillion total parameters and 32 billion active per forward pass, this MoE model supports an extended context length of 256k tokens. It exhibits exceptional reliability, achieving a 100% success rate across all benchmarks, indicating consistent and stable operation. The model shows moderate speed performance, ranking in the 27th percentile, and offers moderate pricing, placing it in the 36th percentile. Kimi K2 0905 excels in Hallucinations (100% accuracy), Email Classification (99% accuracy), and Instruction Following (72% accuracy, 85th percentile), showcasing its precision in factual recall, categorization, and adherence to complex directives. Its agentic coding capabilities are evident in its strong Coding benchmark score (87% accuracy, 63rd percentile) and solid Reasoning performance (78% accuracy, 67th percentile). However, the model shows weaknesses in General Knowledge (77% accuracy, 23rd percentile), Ethics (80% accuracy, 20th percentile), and Mathematics (53% accuracy, 25th percentile), suggesting areas for improvement in broader factual recall and complex problem-solving outside of its specialized coding and reasoning domains.