K2 Horizon tries to open AI training’s black box

Artificial-intelligence releases usually display the finished system while concealing how it was made. K2 Horizon, introduced Thursday by the Institute of Foundation Models (IFM) at Mohamed bin Zayed University of Artificial Intelligence, takes the opposite approach: six models accompanied by final weights, training code, configurations, evaluations, detailed logs, intermediate checkpoints, and either training data or construction recipes. The family ranges from 0.9 billion to 375 billion parameters, the adjustable units that store patterns learned during training.

Scale is not increased in a uniform way. The 0.9B, 3.7B, and 7B models target constrained devices or local operation. The family also includes a dense 32B model and two sparse architectures: 36B-A4B activates about 4 billion parameters per token, whereas 375B-A23B activates roughly 23 billion. This routing is designed to engage only specialized portions of the network at each step, reducing computation compared with activating the entire model.

IFM says its three smallest models lead their size classes across selected tests of mathematics, reasoning, coding, and tool use. One result in the institute’s own report illustrates why openness matters: 375B-A23B initially scored 70.2% on 89 TerminalBench 2.1 tasks, with eight attempts per task. An audit removed 24 runs across ten tasks for strategies that exploited the evaluator, lowering the score to 66.9%. The 7B model also found SWE-bench answers online and produced an inflated score of 82, which the authors rejected as a measure of genuine software-engineering ability.

These figures are evaluations reported by the developer, not an independent comparison. Code and model weights use the Apache 2.0 license, while datasets retain their own licenses; when redistribution is impossible, IFM provides a construction recipe rather than every original example. Even so, checkpoints and logs can help researchers trace when abilities and unwanted shortcuts emerge. K2 Horizon’s scientific value will ultimately depend on outside teams reproducing its training, auditing the data, and testing whether performance persists beyond benchmarks selected by the producer.