IFM releases open source K2 Horizon AI model series with training logs

The Institute of Foundation Models at MBZUAI announced the launch of K2 Horizon, a new family of six open-source artificial intelligence models. The collection spans several sizes, ranging from a compact 0.9-billion parameter version designed for lightweight devices up to a massive 375-billion parameter flagship intended for complex tasks. Rather than releasing only the finished software, the organisation published the full training datasets, source code, evaluation logs, and intermediate checkpoints.

For everyday consumers, this degree of transparency helps accelerate the development of cheaper, privacy-focused software. Smaller versions in the lineup can run directly on consumer hardware like mobile phones, which could allow future smartphone apps to handle complex tasks offline. By sharing the entire recipe used to build the software, the release also gives independent researchers a rare opportunity to inspect how modern systems are built from scratch.

It remains uncertain how widely enterprise developers will adopt the larger models given the high computational costs required to run them independently. Additionally, providing unrestricted access to full training datasets may force the broader industry to re-evaluate how open projects manage intellectual property and data governance going forward.