The Institute of Foundation Models, or IFM, under the UAE’s Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), has released K2 Horizon and open-sourced six models in one batch: 0.9B, 3.7B, 7B, 32B, 36B-A4B, and 375B-A23B. IFM said the 0.9B version is aimed at edge devices such as watches and smart glasses, while the 7B model can run on phones. The largest model, 375B-A23B, uses a mixture-of-experts architecture and activates about 23B parameters per generated token. According to the institute, the 0.9B, 3.7B, and 7B variants set new benchmark results within their respective size classes. Beyond final weights, IFM said it will also release training code, data or data-construction recipes, intermediate checkpoints, training logs, and evaluation records. Some datasets cannot be redistributed directly because of licensing limits, though the team said it will disclose sources, construction methods, and mixture ratios. The models and code are released under Apache 2.0. Artificial Analysis gave the 375B-A23B model an Intelligence Index score of 47, close to MiniMax-M3’s 45, and said it performs better on agent tasks while showing weaker results in knowledge and hard reasoning.
The Institute of Foundation Models (IFM), part of the UAE’s Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), has released K2 Horizon and open-sourced six models at once: 0.9B, 3.7B, 7B, 32B, 36B-A4B, and 375B-A23B.
IFM said the 0.9B model is built for edge devices such as watches and smart glasses, while the 7B version can run on a phone. The largest release, 375B-A23B, uses a mixture-of-experts, or MoE, architecture and activates roughly 23B parameters for each token it generates.
The institute said the 0.9B, 3.7B, and 7B models posted new benchmark highs within their respective size ranges. It also said the release will include more than final weights: training code, data or data-construction recipes, intermediate checkpoints, training logs, and evaluation records are set to be made public.
Some of the data cannot be redistributed directly because of license restrictions. IFM said it will still disclose the data sources, construction methods, and proportions used. The models and code are being released under the Apache 2.0 license.
Artificial Analysis assigned the 375B-A23B model an Intelligence Index score of 47, close to MiniMax-M3’s 45. It said the model is stronger on agent tasks, while its performance is relatively weaker in knowledge and difficult reasoning.
This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan. Disclaimer:
The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.
Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.
