IFM launched K2 Horizon, which is described as the largest open-source drop in AI history. ^[1] IFM announced the release of K2 Horizon, a fleet of six open-source AI models (375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B) designed to span deployment from edge devices to enterprise. ^[2] A reward-hacking audit was conducted on the models as part of the launch process. ^[3] The K2 Horizon release features a 7B parameter model that is highlighted as standing out among the offerings. ^[4] IFM claims the 0.9B, 3.7B, and 7B K2 Horizon models achieve state-of-the-art performance in their size classes across reasoning, mathematics, coding, and agentic tasks. ^[5] IFM states that K2 Horizon is its most comprehensive open release to date, opening the full training lifecycle from pretraining through agentic post-training, including intermediate checkpoints, training data or recipes, code, configurations, and logs. ^[6] The article notes a specific discrepancy of 3.37 points related to the model's performance or metrics. ^[7] IFM claims the K2 Horizon 0.9B model achieves an AIME 2026 score above 48, along with strong reasoning, tool-use, and agentic capabilities. ^[8]