IFM Releases 'K2 Horizon', a Suite of 6 AI Models Exposing the Entire Training Process

IFM has released a suite of 6 interconnected AI models under the name K2 Horizon. This suite consists of 6 sizes in total, including 375B-A23B, 36B-A4B, 32B, 7B, 3.7B, and 0.9B, achieving top-tier performance at each scale. Notably, the 0.9B, 3.7B, and 7B models set new SOTA (State of the Art) benchmarks at their respective scales.

Transparency Across the Entire Training Process

K2 Horizon takes the most comprehensive approach among open model releases to date. It exposes the entire training lifecycle, from pre-training to inference and agent post-processing. By releasing not only final weights but also intermediate checkpoints, training data (or data composition recipes), architectures, mixture configurations, and configuration files, it enables researchers and developers to study and reproduce the model creation process. All models and code are distributed under the Apache 2.0 license.

Performance and Features by Scale

Each model is optimized for specific deployment environments.

• 0.9B: Designed for constrained environments such as watches and glasses. It demonstrates excellent performance in math, reasoning, and agent tasks.
• 3.7B and 7B: Suitable for software engineering and multi-step environments, proving strong capabilities in benchmarks such as TerminalBench.
• 32B and 36B-A4B: Models for local workstations and efficient serving. In particular, 36B-A4B applies a new Mixture-of-Value-Attention mechanism to deliver exceptional performance per active parameter.
• 375B-A23B: The highest-performance model for enterprise deployment.

These 6 models share a common architecture, vocabulary, training methodology, interfaces, evaluation infrastructure, and deployment tools, facilitating movement between scales and dynamic routing. All models support quantization.