AceMath-72B-Instruct
NVIDIA's frontier math post-training suite — 72B instruct plus a record-setting math reward model, non-commercial license.
About
AceMath is a two-part play from NVIDIA's applied deep learning research lab: a family of instruction-tuned math models (1.5B through 72B) and a companion set of math reward models built on a new benchmark, AceMath-RewardBench. The recipe is careful rather than flashy — general-domain SFT first, then targeted math fine-tuning on curated prompts with synthetic responses.
The 72B instruct model averaged 71.8 pass@1 across the paper's math benchmarks, ahead of Qwen2.5-Math-72B-Instruct, GPT-4o and Claude 3.5 Sonnet, and paired with its reward model it set the paper's best rm@8 scores. The important caveat is the license: CC-BY-NC-4.0, so research and personal use are fine but commercial deployment is not.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (arXiv 2412.15084): 71.8 avg pass@1 across math benchmarks vs Qwen2.5-Math-72B-Instruct 68.2, GPT-4o 67.4, Claude 3.5 Sonnet 65.6
- AceMath-72B-RM sets the paper's best rm@8 scores across the math reasoning benchmarks
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER verified, metrics-v1). Evidence: NVIDIA ADLR paper (arXiv 2412.15084) with comparative benchmark tables; CC-BY-NC-4.0 license confirmed via HF API; open weights, training data and benchmarks promised at research.nvidia.com/labs/adlr/acemath. Strengths: strong empirical results, reward-model co-release is genuinely useful. Docked for: (1) CC-BY-NC-4.0 — no commercial use; (2) headline numbers are author-reported. License: CC-BY-NC-4.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · NVIDIA ADLR; released 2024-12-19
- · Paper arXiv 2412.15084; 71.8 avg pass@1 vs Qwen2.5-Math-72B-Instruct 68.2
- · Benchmarks are author-reported