IFM/K2-Horizon-3.7B
IFM/K2-Horizon-3.7B is a 3.7 billion parameter dense decoder-only model from the K2-Horizon family, featuring a native 512K token context window. It demonstrates strong performance in agentic, coding, and reasoning benchmarks, outperforming several comparable models in areas like competition mathematics and software engineering. The model is designed for applications requiring extensive context and robust reasoning capabilities, with fully open training data, code, and evaluation resources.
Loading preview...
Overview
IFM/K2-Horizon-3.7B is a 3.7 billion parameter dense decoder-only model, part of the K2-Horizon family. It is notable for its exceptionally large native 512K token context window, developed through a multi-stage midtraining process. The model's training data, recipe, code, and evaluation resources are fully open, providing transparency and enabling further research.
Key Capabilities
- Extended Context: Features a native 524,288-token context window, allowing for processing very long inputs and generating extensive outputs.
- Strong Small-Model Baseline: Evaluated across agentic, coding, and reasoning benchmarks, demonstrating competitive performance against other dense models in its size class.
- Performance Highlights: Achieves 70.5% on HMMT Feb 2026 (competition mathematics) and 68.6% on SWE-bench Verified (software engineering), often outperforming comparable 3B-4B parameter models.
- Openness: Provides intermediate checkpoints, training data, and code, facilitating detailed study of its development and capabilities.
When to Use This Model
- Long-Context Applications: Ideal for tasks requiring processing or generating very long texts, such as document analysis, summarization of extensive reports, or complex conversational agents.
- Reasoning and Coding Tasks: Suitable for applications in mathematics, software engineering, and scientific reasoning, where it has shown strong benchmark results.
- Agentic Workflows: Designed with agentic capabilities, including tool use and function calling, making it suitable for automated task execution and complex interactive systems.
- Research and Development: Its fully open nature and availability of intermediate checkpoints make it valuable for researchers and developers looking to understand and build upon its architecture and training methodology.