LIMO
The 'less is more' reasoning model — 817 curated examples eliciting 57% AIME from a 32B base.
About
LIMO is the paper that made the reasoning world argue about data. Conventional wisdom said complex reasoning needs 100K+ training examples; the GAIR team showed that 817 meticulously curated demonstrations — long, high-quality chains of thought on hard math problems — could take Qwen2.5-32B-Instruct to 57.1% on AIME and 94.8% on MATH500.
The deeper claim is the Less-Is-More Reasoning Hypothesis: once pretraining has packed in the knowledge, post-training mostly needs to show the model how to organize its thinking, and a few perfect templates beat a mountain of mediocre ones. The model, dataset and eval code are all on Hugging Face. One flag for careful users: the repo declares no license, so the legal terms for reuse are genuinely unclear.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (arXiv 2502.03387): 57.1% AIME24, 94.8% MATH500 with 817 training samples (1% of prior data budgets)
- Same paper: +40.5pp absolute improvement over 100x-data baselines across 10 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 metrics, metrics-v1). Evidence: paper (arXiv 2502.03387) with 10-benchmark evals; open weights and 817-sample dataset on Hugging Face (GAIR/LIMO); influential in the RLVR/data-efficiency debate. Strengths: striking, well-documented result; fully reproducible artifacts. Docked for: (1) no license declared on the HF repo — reuse terms unclear; (2) headline numbers are author-reported; (3) repo activity is quiet. License: none declared.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 6.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · GAIR; released 2025-02-05
- · Paper arXiv 2502.03387; 57.1% AIME24 with 817 samples
- · Benchmarks are author-reported; repo quiet since release