IFM Releases K2 Horizon: Six AI Models with Open Training Data and Code

K2 Horizon Models Introduced

The Institute of Foundation Models (IFM), based in Abu Dhabi, introduced K2 Horizon, a collection of six AI foundation models. These models vary in size from 0.9 billion to 375 billion parameters. IFM states this release represents the "largest fully open-source fleet of AI models" made available to date.

Definition of 'Fully Open'

IFM's definition of "fully open" extends beyond just downloadable model weights. For K2 Horizon, the institute has committed to publishing training and evaluation code, training data (or detailed construction recipes if redistribution is not possible), configurations, logs, and intermediate checkpoints. This comprehensive release covers the entire training lifecycle, from pretraining through agentic post-training.

Developer Access and Reproducibility

The goal of this open approach is to allow developers to examine how the models were constructed, reproduce their development process, and adapt them for specific applications. Eric Xing, IFM founder, stated that this commitment supports open science by providing visibility into data, methods, and results, enabling reproduction and improvement.

Current Availability Status

While all six models have downloadable weights, not all promised components are available at launch. Model cards for the 0.9B, 32B, and 375B models indicate that some training data, code, or checkpoints will be released later. Additionally, the 32B model is currently a Stage 1 checkpoint, with the final version yet to be released.