ulamai/Ulam-1-Small
Ulam-1-Small is a 3.086 billion parameter mathematical reasoning model developed by Ulam AI, built upon the Qwen2.5-3B lineage. This BF16 Transformers model is specifically designed for exploratory mathematical problem solving and proof-style reasoning. It excels in generating strong claims and purported proofs, making it suitable for research-oriented mathematical tasks requiring human review.
Loading preview...
Ulam-1-Small: A Mathematical Reasoning Model
Ulam-1-Small is a 3.086 billion parameter model developed by Ulam AI, specifically engineered for advanced mathematical reasoning. Built on the Qwen2.5-3B, Qwen2.5-Coder-3B, and WeiboAI/VibeThinker-3B lineage, this BF16 Transformers model focuses on exploratory mathematical problem solving and proof-style reasoning, rather than formal theorem certification.
Key Capabilities
- Mathematical Problem Solving: Designed to tackle complex mathematical problems and generate proof-style reasoning.
- High Context Length: Features a 131,072-token architectural context configuration, though practical serving lengths may vary.
- Robust Training: Benefits from a comprehensive training pipeline including full-weight outcome learning, supervised proof refinement, verified reward optimization, and direct preference optimization (DPO).
- Performance on Benchmarks: Achieved a mean score of 2.287 on the 226-item ErdosBench audit and 667/882 on SIMOBench, demonstrating strong performance in mathematical tasks.
Good For
- Exploratory Mathematical Research: Ideal for researchers exploring mathematical problems and generating proof-style reasoning that requires human review.
- Compact Reasoning Model Research: Suitable for studies on the efficiency and capabilities of smaller, specialized reasoning models.
- Local Inference: Supports standard Transformers and vLLM APIs for local deployment and experimentation.
- Commercial and Non-Commercial Use: Available under the MIT License for a wide range of applications.