Kimi K2.5 Tops Global Open-Source AI Rankings Within 24 Hours of Launch

On January 27, Chinese AI startup Kimi (Moonshadow AI) released its next-generation open-source model, Kimi K2.5. Within 24 hours, K2.5 topped multiple global AI rankings.

In the latest LMarena leaderboard, Kimi K2.5 ranks just below top closed-source models such as Claude 4.5 and Gemini 3 Pro in code and text tasks, making it the number one open-source model worldwide. On the independent evaluation list from Artificial Analysis, K2.5 placed 5th overall while remaining the top open-source model.

Built on a native multimodal architecture, Kimi K2.5 deeply integrates vision, reasoning, programming, and agent capabilities. By reconstructing the reinforcement learning infrastructure and optimizing training algorithms, the model introduces a new Agent cluster system, able to dispatch up to 100 specialized agents and process 1,500 steps in parallel. This cluster collaboration boosts efficiency for complex tasks like market research, multilingual document translation, and cross-domain literature reviews by more than ten times.

The launch generated a global buzz. News on X platform frequently trended, surpassing other tech announcements. Despite performance approaching top U.S. closed-source models like Claude, K2.5 is priced at only one-fifth of their cost. Programming tool competitor Kilo Code also announced K2.5 as its default model, free during the first week.

According to CNBC, "Kimi demonstrates astonishing real-world performance in long-text processing and intelligent assistant applications. The DeepSeek and Kimi cases show that while computing power matters, algorithm efficiency, data quality, and engineering capabilities are equally critical." Peng Bo reports, "K2.5 processes text, image, and video in a single prompt, aligning with the trend of 'universal models' set by industry leaders, surpassing open-source peers and narrowing the coding performance gap with top closed-source models."

Kimi founder and CEO Yang Zhilin stated: "We rebuilt the reinforcement learning infrastructure and optimized training algorithms to ensure K2.5 achieves ultimate efficiency and performance."