According to CoinMeta, Abu Dhabi IFM has released the K2 Horizon model, opening up the source code for 6 models at once, which are 0.9B, 3.7B, 7B, 32B, 36B- a4B, and 375B- a23B. Among them, the 0.9B model is designed for edge devices such as watches and glasses, while the 7B model can run on mobile phones. The largest 375B- a23B model uses the MOE architecture, and approximately 23B parameters are activated for each token generated. Officials claim that the 0.9B, 3.7B, and 7B models have refreshed multiple evaluation scores in their respective sizes. In addition to the final weights, the training code, data construction recipes, intermediate checkpoint, training logs, and evaluation records will also be made public. Some data is subject to license restrictions and cannot be directly redistributed, but the sources, construction methods, and ratios will be made available. The models and code use the Apache 2.0 protocol. Artificial Analysis awarded the 375B- a23B a comprehensive ability score of 47 points, which is close to the 45 points given to Minimax-M3. This model performs stronger on Agent tasks but is relatively weaker in terms of knowledge and complex reasoning.