DeepScaleR 1.5B
The 1.5B model that beat o1-preview on math with $4,500 of RL training.
About
Released February 2025 by the Agentica project at Berkeley's Sky Computing Lab, DeepScaleR-1.5B-Preview fine-tunes DeepSeek-R1-Distill-Qwen-1.5B with distributed reinforcement learning (GRPO) on 40,000 competition math problems, using iterative context lengthening from 8K to 24K tokens. It reached 43.1% Pass@1 on AIME 2024 — surpassing OpenAI's o1-preview — for roughly 3,800 A100 hours (~$4,500).
Limitations: math-specialized, not a generalist; 'Preview' release with a research-grade recipe rather than a polished product.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
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 (metrics tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 67; Trending score 72/100 from researcher assessment.