AceMath-7B-Instruct

AceMath is a family of frontier math-reasoning models from NVIDIA researchers (Zihan Liu, Yang Chen, Wei Ping, and colleagues), described in the 2024 paper 'AceMath: Advancing Frontier Math Reasoning with Post-Training and Reward Modeling' (arXiv:2412.15026). The training recipe applies supervised fine-tuning followed by targeted math-domain tuning on a curated set of over 2.3 million math prompts (1.3M initial plus 1M synthetically generated), plus a math-specialized reward model trained on the AceMath-RewardBench. The flagship AceMath-72B-Instruct was reported to greatly outperform Qwen2.5-Math-72B-Instruct, GPT-4o, and Claude 3.5 Sonnet on challenging math benchmarks, with 7B and smaller variants offering the same recipe at lower cost.

Limitations: all AceMath models are released for non-commercial use only under Creative Commons Attribution Non-Commercial 4.0 — there is no commercial-use path, which rules out product deployments. The synthetic training data was generated with OpenAI models, so its terms of use also bind downstream use. As a math specialist it is not a general assistant; use it for math reasoning and reward modeling rather than open-ended chat.