AceReason-Nemotron-7B: RL-Enhanced Math and Code Reasoning

NVIDIA's AceReason-Nemotron-7B is a 7.6 billion parameter model built upon the DeepSeek-R1-Distilled-Qwen-7B architecture, uniquely trained entirely through reinforcement learning (RL). This approach significantly boosts its capabilities in complex math and code reasoning tasks.

Key Capabilities & Differentiators

• Reinforcement Learning Focus: The model's core strength comes from a novel RL training process, starting with math-only prompts and then extending to code-only prompts. This method not only enhances math performance but also improves code reasoning.
• Strong Reasoning Performance: Achieves notable scores, including 69.0% on AIME 2024 (a 14.5% improvement over its base) and 51.8% on LiveCodeBench v5 (an 8% improvement). It also scores 53.6% on AIME 2025 and 44.1% on LiveCodeBench v6.
• Context Length: Supports a substantial context length of 131,072 tokens, beneficial for intricate reasoning problems.
• Systematic RL Study: The development includes extensive ablations to understand and optimize the RL training process, detailed in their technical report.

When to Use This Model

• Advanced Math Problem Solving: Ideal for applications requiring high accuracy in mathematical reasoning, as evidenced by its AIME benchmark results.
• Code Generation and Debugging: Excels in coding challenges, particularly those found in competitive programming or complex software development scenarios.
• Research in RL for LLMs: Provides a strong baseline and methodology for further research into applying reinforcement learning to enhance foundational model capabilities.