IFM Launches K2 Horizon to Provide 'Radically Open' AI Training Blueprint

IFM has released K2 Horizon, a suite of six open-weights models designed to push the boundaries of transparency in artificial intelligence. The release is positioned as a "radically open" effort to provide the industry with a complete map of the AI training process.

The K2 Horizon fleet includes six distinct model sizes: 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B. Released under the Apache 2.0 license, the models are accompanied by a commitment to disclose the entire training lifecycle. According to IFM, this includes the release of intermediate checkpoints, training code, logs, and the specific data recipes used for pretraining, reasoning, and agentic post-training. IFM described the launch as their "most comprehensive open release to date."

Performance Claims

IFM claims that the smaller variants of the fleet—the 0.9B, 3.7B, and 7B models—set new state-of-the-art performance benchmarks for their respective size classes. Specifically, the company positions the 7B model as the best-performing model in the industry under 10 billion parameters, noting that it is capable of running locally on a smartphone.

The Push for Full Transparency

This release arrives as the AI community increasingly distinguishes between "open-weight" models and "fully open" stacks. While many companies release the final weights of a model, they typically keep the training data and the exact processes used to refine the model proprietary. By promising to open the full lifecycle, IFM is attempting to move beyond weights to provide a reproducible framework for frontier-level AI.

Industry Implications and Skepticism

If IFM delivers on its promise of radical openness, it would provide a rare and valuable blueprint for researchers and developers to understand how high-performing models are actually built. Such transparency could accelerate the development of smaller, more efficient models by revealing the specific data mixtures that drive performance.

However, the launch has not been without friction. Early community discussions on Hacker News have raised questions regarding the validity of IFM's performance claims when compared to competitors like Qwen. Furthermore, some early users reported that the promised open-source repositories were initially empty, leading to questions about whether the infrastructure for this "radical openness" was fully ready at the time of the announcement.

What to Watch

The primary metric for the success of K2 Horizon will be the actual availability and completeness of the promised training logs and data recipes. While the weights are available under Apache 2.0, the industry is waiting to see if the full training lifecycle is disclosed as promised. Until these repositories are fully populated and verified by third parties, the claim of "radical openness" remains a corporate promise rather than a verified industry standard.