IFM/K2-Horizon-7B
IFM/K2-Horizon-7B is a 9 billion parameter decoder-only language model from IFM, featuring a native 512K token context window. This model excels in agentic, coding, long-context, and reasoning benchmarks, outperforming several reference models in these areas. It incorporates a unique Final Token Preference Optimization (FTPO) stage to prevent 'doom loops' in long generations, making it particularly robust for complex agentic tasks and extended outputs. K2-Horizon-7B is fully open, with public training data, code, and evaluation resources.
Loading preview...
K2-Horizon-7B: A High-Performance, Long-Context Model
K2-Horizon-7B is a 9 billion parameter decoder-only model developed by IFM, distinguished by its exceptionally large native 512K token context window. It is designed to deliver strong performance across a range of demanding tasks, including agentic operations, coding, long-context understanding, and complex reasoning.
Key Capabilities and Differentiators
- Extended Context Window: Features a native 524,288-token context from midtraining stages, enabling processing of very long inputs and generating extensive outputs.
- Enhanced Reasoning and Coding: Demonstrates strong performance in mathematical reasoning (e.g., 73.3% on HMMT Feb 2026), software engineering (70.6% on SWE-bench Verified), and scientific coding (31.6% on SciCode), often surpassing comparable models.
- Agentic Task Proficiency: Achieves notable scores in agentic tool use (25.8% on tau3-Banking) and web browsing (59.0% on BrowseComp), indicating strong capabilities for automated tasks.
- Anti-Doom-Loop Training: Incorporates a unique Final Token Preference Optimization (FTPO) stage to mitigate degenerate repetition in long generations, significantly improving reliability for agentic and extended outputs.
- Openness: IFM provides public access to its training data, recipe, training code, and evaluation resources, fostering transparency and research.
When to Use This Model
K2-Horizon-7B is particularly well-suited for applications requiring:
- Long-form content generation or analysis: Its 512K context window is ideal for summarizing lengthy documents, generating detailed reports, or maintaining coherence over extended dialogues.
- Complex problem-solving: Excels in mathematical, scientific, and expert-level reasoning tasks.
- Code generation and software engineering: Strong benchmarks in coding suggest its utility for development assistance and automated code tasks.
- Agentic workflows: The anti-doom-loop training and strong performance in agentic benchmarks make it a robust choice for building AI agents that interact with tools or environments.