Abu Dhabi IFFM Open Sources 6-Tier K2 Horizon Model: From $0.9B to $375B Full Coverage

Perceive Beating AI News Flash: Under the UAE's MBZUAI, the Institute for Foundational Models (IFM) has released K2 Horizon, open-sourcing 6 models in one go, namely 0.9B, 3.7B, 7B, 32B, 36B-A4B, and 375B-A23B. The 0.9B is designed for wearable devices such as watches and glasses, while the 7B can run on mobile phones. The largest 375B-A23B utilizes MoE, with approximately 23B parameters activated per generated token. The official statement mentioned that the 0.9B, 3.7B, and 7B have refreshed their evaluation scores in their respective sizes.

In addition to the full weights, training code, data or data generation recipes, intermediate checkpoints, training logs, and evaluation records will also be made public. Some data are restricted by licenses and cannot be directly redistributed, but the sources, construction methods, and proportions will be disclosed. The models and code are licensed under Apache 2.0.

Artificial Analysis has awarded the 375B-A23B an Intelligence Index score of 47, close to the MiniMax-M3's 45. It performs better on Agent tasks but is relatively weaker in knowledge and high-level reasoning.