PrimeIntellect/INTELLECT-MATH-step-47

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jan 17, 2025License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

PrimeIntellect/INTELLECT-MATH-step-47 is a 7.6 billion parameter model developed by PrimeIntellect, specifically optimized for advanced mathematical reasoning tasks. This model leverages a two-stage training process, including supervised fine-tuning on verified QwQ outputs and reinforcement learning with the PRIME-RL recipe. It demonstrates superior performance in mathematical benchmarks like MATH-500 and OLYMPIADBENCH, achieving 81.6% and 46.7% respectively, and matches previous state-of-the-art models with significantly faster training times.

Loading preview...

INTELLECT-MATH-step-47: Advanced Mathematical Reasoning

INTELLECT-MATH-step-47 is a 7.6 billion parameter model from PrimeIntellect, engineered for high-performance mathematical reasoning. It utilizes a unique two-stage training methodology, beginning with supervised fine-tuning (SFT) on high-quality, verified QwQ outputs, followed by a reinforcement learning (RL) stage using the PRIME-RL recipe.

Key Capabilities & Performance

This model's strength lies in its ability to achieve strong mathematical reasoning performance, outperforming its predecessors and matching state-of-the-art models with significantly reduced training time. Its SFT data quality is a key differentiator, enabling 10x faster RL training while maintaining high accuracy. Benchmarks highlight its capabilities:

  • MATH-500: Achieves 81.6%
  • OLYMPIADBENCH: Scores 46.7%
  • AIME 2024: Reaches 26.7%
  • AMC: Attains 57.8%

These results demonstrate its robust performance across various complex mathematical problem sets, making it a strong contender for tasks requiring precise and advanced mathematical problem-solving.

Good For

  • Mathematical Problem Solving: Excels in competitive math problems and complex reasoning tasks.
  • Research & Development: Ideal for exploring advanced RL techniques in language model training, particularly for mathematical domains.
  • Educational Tools: Potentially useful for generating explanations or solutions for challenging math problems.