IFM/K2-Horizon-0.9B
IFM/K2-Horizon-0.9B is a 0.9 billion parameter decoder-only model from IFM, featuring a 128K context window achieved through YaRN RoPE scaling. This compact model is optimized for reasoning across mathematics, coding, and scientific tasks, demonstrating strong performance in competition mathematics and code generation benchmarks. It is trained using multi-teacher distillation with domain experts for math, code, STEM, and instruction following, making it suitable for applications requiring robust analytical capabilities and long-context understanding.
Loading preview...
K2-Horizon-0.9B: Compact Reasoning with 128K Context
K2-Horizon-0.9B is a 0.9 billion parameter dense, decoder-only model developed by IFM, distinguished by its exceptionally long 128K (131,072 tokens) context window, enabled by YaRN RoPE scaling. This model is engineered for strong reasoning capabilities across various domains, including mathematics, coding, and scientific inquiry.
Key Capabilities & Features
- Extended Context Window: Supports up to 128K tokens, significantly enhancing its ability to process and understand long inputs.
- Specialized Reasoning: Achieves high performance in competition mathematics (e.g., 48.5% on AIME 2026) and code generation (e.g., 79.9% on HumanEval+), outperforming several reference models in its class.
- Multi-Teacher Distillation: Benefits from training with domain-specific teachers for math, code, STEM, and instruction following, contributing to its robust analytical skills.
- Fully Open: IFM plans to release the training data, recipe, and code, promoting transparency and further research.
- Agentic Capabilities: Includes support for tool-use and function calling, with a dedicated
k2_horizontool-call parser.
Recommended Usage
For optimal performance, IFM recommends using reasoning_effort="high", temperature=0.6, top_p=0.95, and allowing at least 32,768 output tokens to prevent truncation of reasoning processes. The model is well-suited for applications requiring deep analytical reasoning, complex problem-solving, and long-context understanding in technical and scientific fields.