IFM/K2-Horizon-375B-A23B

K2-Horizon-375B-A23B: A Frontier-Class Agentic MoE Model

IFM's K2-Horizon-375B-A23B is a 375 billion parameter sparse Mixture-of-Experts (MoE) model, activating 23 billion parameters per token. It boasts an impressive 512K token context window, maintained from mid-training stages. This model is distinguished by its "frontier-class agentic performance," matching or exceeding larger open-weight MoE models and competing with closed frontier models on benchmarks for agentic tool use, terminal tasks, and long-horizon workflows. IFM plans to release intermediate checkpoints, training data, and code, emphasizing its fully open nature.

Key Capabilities

• Exceptional Agentic Performance: Achieves high scores on agentic benchmarks like GDPVal-AA (1,441 Elo), Toolathlon Verified (65.3%), and Apex-Agents (24.8% pass@1).
• Extensive Context Window: Supports a native 512K token context, enabling complex and long-form interactions.
• Strong Coding Abilities: Demonstrates robust performance in agentic terminal use (Terminal-Bench 2.1: 70.2%) and software engineering tasks (SWE Bench Pro: 42.6%).
• Scientific Reasoning: Shows solid performance in graduate-level science QA (GPQA Diamond: 87.3%) and frontier physics reasoning (CritPt: 8.6%).
• Openness: Commitment to releasing training data, recipes, and code for transparency and research.

Good For

• Developing and deploying advanced AI agents requiring sophisticated tool use and workflow automation.
• Applications demanding long-context understanding and reasoning, such as deep web research or complex problem-solving.
• Software development and scientific research tasks that benefit from strong coding and scientific reasoning capabilities.
• Researchers interested in studying model evolution through intermediate checkpoints and open training artifacts.