Abu Dhabi's IFM Launches K2 Horizon: Six Fully Open Apache 2.0 Models from 0.9B to 375B
The Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in Abu Dhabi has officially launched K2 Horizon, a six-model family spanning edge devices to enterprise clusters published entirely under the permissive Apache 2.0 license. The K2 Horizon collection—now available on Hugging Face—includes models sized at 0.9B, 3B, 7B, 32B, 70B, and the flagship 375B-A23B Mixture-of-Experts (MoE) architecture. Unlike conventional open-weight releases that withhold pre-training sets, IFM has released K2 Horizon alongside its complete tokenized training datasets, data filtering pipelines, training code, and evaluation recipes, establishing what MBZUAI terms the industry's most transparent foundation model fleet. The flagship K2-Horizon-375B-A23B routes 23 billion active parameters per token across a 524,288-token (524K) context window. In evaluations published alongside the weights and benchmarked by Artificial Analysis, the model scored 47 on the Intelligence Index, achieved 87% on GPQA Diamond scientific reasoning, and scored 72% on Terminal-Bench 2.1, marking a 57-percentage-point gain over its predecessor, K2 Think V2. All checkpoints, deployment scripts, and datasets can be downloaded directly from Hugging Face for unrestricted commercial self-hosting, academic research, and downstream fine-tuning without licensing royalty fees.
