IFM releases K2 Horizon, a connected fleet of six models ranging from 0.9B to 375B-A23B. Across reasoning, mathematics, coding, and agentic tasks, K2 Horizon delivers top-tier performance in every size class. The 0.9B, 3.7B, and 7B models set new state of the art at their respective scales. This release is the most comprehensive open model fleet to date, opening the full training lifecycle including intermediate checkpoints, data recipes, architecture, and final weights under the Apache 2.0 license. The 36B-A4B model features a new Mixture-of-Value-Attention mechanism, delivering exceptional capability per active parameter. The 32B and 375B-A23B models rank among the top in their respective classes. Together, the fleet provides competitive performance from edge devices to enterprise deployments. K2 Horizon is the first open model family to expose the complete development process through agentic post-training, enabling study and reproduction of reasoning and tool-use capabilities. The 0.9B model targets constrained environments like watches and glasses, while the 3.7B and 7B bring advanced capabilities to phones. The 32B and 36B-A4B serve local workstations and efficient serving. The 375B-A23B delivers strongest capabilities for demanding enterprise use. All six models include quantization support and share core architecture, vocabulary, and deployment tooling for dynamic routing and cross-scale study.

World-leading performance across scales defines the release. The smallest models achieve AIME 2026 scores above 48 and strong reasoning capabilities. The 3.7B and 7B extend these to software engineering environments. While complex exploration tasks remain challenging for smallest models, K2 Horizon moves the boundary of possibility at every scale, representing a new performance frontier across all parameter counts.