Open weightsStanford/10 trustVerified

s1-32B

Name: s1-32B
Rating: 8

o1-level reasoning from 1,000 examples and budget forcing

Visit official page →

About

s1-32B comes from Stanford researchers who asked how little it takes to build a reasoning model: supervised fine-tuning on just 1,000 curated reasoning traces, trained in about 26 minutes on 16 H100s for roughly $50 of compute. Its signature trick is budget forcing at inference time — appending "Wait" to make the model keep thinking, or cutting it off to save time — which gives test-time scaling without retraining. The paper reports it beats o1-preview on math tasks like AIME 2024. It is a research artifact, not a product: no hosted API from the authors, and forcing extra thinking can push it into repetitive loops.

Specs

Context window

32K

Pricing

Free quota

API access

Self-host via Hugging Face weights (simplescaling org); vLLM with budget forcing supported

Weights

Open weights

Released

January 31, 2025

Benchmarks

Paper-reported: AIME 2024 56.7 with budget forcing (26.7 base before fine-tuning), MATH500 ~92.6-93.0

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

test-time scaling researchmath reasoningefficient reasoning fine-tuning

Trust score — 8/10

/10 trustVerified

Scored 8/10 (verified tier, metrics-v1). Evidence: Paper read: arXiv 2501.19393 's1: Simple test-time scaling' — 1,000 examples, 26 min on 16 H100s, budget forcing technique.; Benchmarks from paper: AIME24 56.7 with budget forcing, MATH500 ~92.6-93.0 — paper-reported.; Published Jan 31 2025 per multiple sources; GitHub simplescaling/s1 confirms artifacts.; Limitation: budget forcing can cause repetitive loops; research artifact without vendor API.; License: Apache 2.0 code and dataset per coverage..