IFM Releases K2 Horizon: A 'Radically Open' Fleet of Six AI Models
By providing the full training lifecycle, including data recipes and checkpoints, IFM moves beyond open weights toward a model of open science.
IFM has released K2 Horizon, a connected fleet of six open-weights models designed to bring unprecedented transparency to the development of frontier AI. The release marks a significant shift in the industry by providing the full training lifecycle rather than just the final model weights.
The K2 Horizon fleet consists of six models of varying scales: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B. According to IFM, the release is "radically open," encompassing not only the final weights but also intermediate checkpoints, training code, configurations, and fine-grained logs. The models and associated code are released under the Apache 2.0 license, while the training datasets are provided under applicable licenses, such as ODC-BY. This comprehensive package includes the open architecture, mixture compositions, and the specific data recipes used to train the models.
The Push for Transparency
This release arrives as the AI industry faces growing pressure to move away from the "black box" nature of closed-source frontier models. There is an increasing demand for "fully open" AI, where transparency extends to the data and processes used during creation. By exposing these elements, developers aim to reduce the potential for societal manipulation that can occur through hidden training biases in proprietary systems. When the training process is hidden, it is nearly impossible for external auditors to verify the safety or neutrality of a model's outputs.
Shifting to Open Science
By open-sourcing the entire training pipeline, IFM is providing a technical blueprint for the construction of frontier-level models. This level of disclosure allows the global research community to audit the training process, identify potential flaws, and attempt to reproduce or improve upon the results. This transition effectively shifts the open-source AI landscape from a model of "open weights"—where the result is shared but the process is secret—to a model of "open science," where the methodology is as accessible as the product.
Future Implications
As researchers begin to digest the K2 Horizon checkpoints and recipes, the industry will be watching to see if these models can be independently verified or surpassed using the provided blueprints. While IFM has released the tools necessary for reproduction, the extent to which the community can leverage these logs and configurations to accelerate the development of other open models remains to be seen. This release challenges other AI labs to move beyond marketing claims of openness and provide the actual evidence of their training methodologies.