IFM/K2-Horizon-375B-A23B

K2-Horizon-375B-A23B Overview

K2-Horizon-375B-A23B is IFM's flagship sparse Mixture-of-Experts (MoE) model, featuring 375 billion total parameters with 23 billion activated per token. It boasts an extensive native 512K token context window, making it highly capable for complex, long-form tasks. The model is fully open, with its checkpoints, training data, and code publicly released, allowing for detailed study of its development and capabilities.

Key Capabilities

• Frontier-class Agentic Performance: Achieves strong results on agentic tool use, terminal, and long-horizon workflow benchmarks, competing effectively with larger open-weight MoE models and closed frontier models. For instance, it scores 34.0 on tau3-Banking and 65.3 on Toolathlon Verified for agentic tool use, and 24.8 on Apex-Agents for long-horizon professional workflows.
• Extended Context Window: Supports a native 512K token context, enabling processing and generation of very long inputs and outputs.
• Openness and Reproducibility: Provides intermediate checkpoints, training data (TxT360-v2 dataset), and training code, facilitating research into capability evolution and enabling full reproducibility.
• Strong Coding and Scientific Reasoning: Demonstrates robust performance in coding benchmarks like Terminal-Bench 2.1 (70.2) and SWE-Atlas-QnA (48.4), and scientific reasoning tasks such as GPQA Diamond (87.3).

Good for

• Developing Advanced AI Agents: Ideal for applications requiring sophisticated agentic tool use, workflow automation, and complex task execution.
• Long-Context Applications: Suitable for tasks demanding extensive context understanding and generation, such as deep web research (72.8 on BrowseComp) and long-form content creation.
• Research and Development: Its fully open nature, including intermediate checkpoints and training data, makes it an excellent resource for academic and industrial research into large-scale MoE models and agentic AI.