unsloth/AceReason-Nemotron-14B

TEXT GENERATIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 23, 2025License:otherArchitecture:Transformer Featherless Exclusive Cold

AceReason-Nemotron-14B is a 14.8 billion parameter language model developed by NVIDIA, specifically trained for advanced math and code reasoning. Starting from DeepSeek-R1-Distilled-Qwen-14B, it was entirely fine-tuned using reinforcement learning (RL) with a two-stage approach: math-only RL followed by code-only RL. This model achieves strong performance on benchmarks like AIME 2024 (78.6%), AIME 2025 (67.4%), LiveCodeBench v5 (61.1%), and LiveCodeBench v6 (54.9%), making it suitable for complex problem-solving in mathematics and programming.

Loading preview...

AceReason-Nemotron-14B: Advanced Math and Code Reasoning

AceReason-Nemotron-14B is a 14.8 billion parameter model developed by NVIDIA, specifically engineered for superior performance in mathematical and coding reasoning tasks. It is built upon the DeepSeek-R1-Distilled-Qwen-14B base and uniquely trained entirely through a reinforcement learning (RL) process. This RL approach involves distinct phases: initial training on math-only prompts, followed by training on code-only prompts, which has been shown to significantly enhance reasoning capabilities across both domains.

Key Capabilities

  • Enhanced Reasoning: Excels in complex problem-solving for both mathematics and programming.
  • Reinforcement Learning (RL) Driven: Utilizes a novel two-stage RL training methodology to push the limits of reasoning ability.
  • Strong Benchmark Performance: Achieves 78.6% on AIME 2024, 67.4% on AIME 2025, 61.1% on LiveCodeBench v5, and 54.9% on LiveCodeBench v6, outperforming several comparable models.
  • Code Generation: Demonstrates strong performance in generating correct and efficient code solutions.

Good For

  • Mathematical Problem Solving: Ideal for applications requiring advanced mathematical reasoning, such as competitive math or scientific calculations.
  • Code Generation and Debugging: Suitable for developers needing assistance with complex coding challenges and generating robust solutions.
  • Research in RL for LLMs: Provides a strong baseline and methodology for further research into reinforcement learning applications for language models.

For detailed training methodology and further evaluation results, refer to the technical report.

Finetunes
2 models