DuoNeural/Qwen2.5-Math-NeuralMath-7B

TEXT GENERATIONConcurrency Cost:1Model Size:7.6BQuant:FP8Ctx Length:32kPublished:Apr 13, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

DuoNeural/Qwen2.5-Math-NeuralMath-7B is a 7.6 billion parameter, 32K context length language model fine-tuned from Qwen/Qwen2.5-Math-7B-Instruct. Developed by DuoNeural, this model specializes in advanced mathematical reasoning, particularly for competition and olympiad-level problems. It is optimized for generating deep chain-of-thought solutions and consistent \boxed{} answer formatting, making it ideal for complex math problem-solving applications.

Loading preview...

DuoNeural/Qwen2.5-Math-NeuralMath-7B Overview

DuoNeural/Qwen2.5-Math-NeuralMath-7B is a specialized 7.6 billion parameter language model, fine-tuned by DuoNeural from the Qwen2.5-Math-7B-Instruct base model. Its primary focus is on enhancing mathematical reasoning capabilities, particularly for challenging competition and olympiad-level math problems.

Key Capabilities and Differentiators

  • Enhanced Chain-of-Thought Reasoning: The model is trained on extensive, structured reasoning traces, enabling it to produce deeper and more coherent step-by-step solutions.
  • Competition Math Expertise: Fine-tuned using curated datasets like NuminaMath-CoT, which includes problems from AMC/AIME and olympiads, improving its performance on advanced mathematical challenges.
  • Consistent Output Formatting: Reliably generates answers in the \boxed{} format, crucial for automated evaluation and clear presentation.
  • QLoRA SFT Training: Utilizes QLoRA Supervised Fine-Tuning (SFT) with a 4-bit base and LoRA rank 16, trained over approximately 1.26 million tokens across three epochs on a specialized math dataset.

Use Cases and Limitations

This model is specifically designed for mathematical reasoning and problem-solving. It is not intended for general conversation. While it excels in English math problems, its performance in other languages is untested. Very long multi-step proofs exceeding 1024 tokens may be truncated due to its training sequence length. GGUF quantizations are available for various use cases, including q4_k_m for a balance of quality and speed, and f16 for full precision.