s1-32B
The sub-$50 reasoning model that popularized test-time scaling
About
s1 is the Stanford and University of Washington reasoning model that made headlines for costing under $50 in compute, released January 2025. It's Qwen2.5-32B-Instruct fine-tuned on just 1,000 distilled reasoning traces, plus a 'budget forcing' trick that scales test-time thinking up or down. The paper reports AIME24 56.7 and MATH500 93.0 — near Gemini 2.0 Flash Thinking on math — though it doesn't match o1 or R1 on harder reasoning.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Paper-reported: AIME24 56.7, MATH500 93.0, GPQA 59.6 with budget forcing.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Paper released 2025-01-31 (arXiv:2501.19393) by Stanford/UW researchers; TechCrunch coverage confirms the sub-$50 training story; Qwen2.5-32B-Instruct fine-tuned on 1,000 distilled traces; budget-forcing test-time scaling is the core contribution; Benchmark figures (AIME24 56.7, MATH500 93.0, GPQA 59.6) are paper-reported and labeled as such; No verbatim official repo URL surfaced, so officialUrl left empty rather than guessed. License: Code and s1K data are Apache-2.0; weights inherit the Qwen2.5-32B-Instruct (Apache-2.0) lineage — confirm on the HF card..
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Paper released 2025-01-31 (arXiv:2501.19393) by Stanford/UW researchers; TechCrunch coverage confirms the sub-$50 training story
- · Qwen2.5-32B-Instruct fine-tuned on 1,000 distilled traces; budget-forcing test-time scaling is the core contribution
- · Benchmark figures (AIME24 56.7, MATH500 93.0, GPQA 59.6) are paper-reported and labeled as such
- · No verbatim official repo URL surfaced, so officialUrl left empty rather than guessed