IFM/K2-Horizon-32B

K2-Horizon-32B Overview

K2-Horizon-32B is a 32 billion parameter dense decoder-only model, part of the K2-Horizon family developed by IFM. A key differentiator is its native 512K token context window, which is integrated from early midtraining stages, making it highly capable for processing extensive inputs. The model emphasizes transparency, with its training data, recipe, and code all publicly available.

Key Capabilities

• Extended Context Handling: Features a native 512K token context window, enabling robust performance on long-context reasoning tasks.
• Agentic and Coding Performance: Evaluated across agentic, coding, and reasoning benchmarks, demonstrating strong capabilities in these areas.
• Openness and Transparency: Provides full access to training data, recipes, and code, fostering research and development.
• Intermediate Checkpoints: Releases intermediate checkpoints to allow for detailed study of capability evolution throughout the training process.

Good For

• Applications requiring extensive context: Ideal for tasks that benefit from processing very long documents or conversations, such as summarization, question answering over large texts, or complex code analysis.
• Research into model training and evolution: The availability of intermediate checkpoints makes it suitable for researchers studying how model capabilities develop over training stages.
• Developing agentic systems: Designed and benchmarked for agentic tool use and terminal interaction, making it a strong candidate for building AI agents.
• Coding and scientific reasoning tasks: Shows competitive performance in scientific coding and expert-level reasoning benchmarks.