Cannae-AI/ReasoningLlama-Math-1B-IT

Hugging Face
TEXT GENERATIONPricing:Input $0.108 / Output $0.804Concurrent Unit Cost:1Model Size:1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Nov 18, 2025License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Warm

ReasoningLlama-Math-1B-IT by Cannae-AI is a 1 billion parameter instruction-tuned Llama-3.2 model, fine-tuned for mathematical reasoning tasks. It leverages the OpenMathReasoning dataset, which was instrumental in winning the AI Mathematical Olympiad challenge. This model is specifically optimized for complex mathematical problem-solving and logical deduction, offering specialized capabilities in this domain.

Loading preview...

Model Overview

ReasoningLlama-Math-1B-IT is a 1 billion parameter language model developed by Cannae-AI, based on the unsloth/Llama-3.2-1B-Instruct architecture. Its primary distinction lies in its specialized fine-tuning for mathematical reasoning.

Key Capabilities

  • Mathematical Reasoning: The model is specifically fine-tuned on the unsloth/OpenMathReasoning-mini dataset, a derivative of the nvidia/OpenMathReasoning dataset. This dataset was notably used in the AI Mathematical Olympiad (AIMO) challenge, indicating its focus on advanced mathematical problem-solving.
  • Instruction-Tuned: It is an instruction-tuned model, designed to follow specific prompts and instructions for mathematical tasks.

Training and Origin

This model is a fine-tuned version of unsloth/Llama-3.2-1B-Instruct. The training data, unsloth/OpenMathReasoning-mini, is a smaller version of the dataset that contributed to a winning solution in the AIMO competition.

Recommended Usage

For optimal inference, the developers recommend specific settings: min_p = 0.1 and temperature = 1.5. These settings are suggested to enhance the model's performance in its specialized mathematical reasoning tasks.

Finetunes
2 models