QED-Nano: A Specialized Proof-Writing Model

QED-Nano is a 4 billion parameter model from lm-provers, specifically post-trained to enhance its capabilities in writing mathematical proofs. Despite its relatively small size, it demonstrates strong performance on challenging benchmarks, making it a notable model for domain-specific reasoning tasks.

Key Capabilities

• Exceptional Proof-Writing: Achieves a 40% score on the IMO-ProofBench benchmark, a significant +20% improvement over its Qwen3 base model, and matches the performance of much larger models like GPT-OSS-120B.
• Agentic Scaffold Integration: Utilizes an agent scaffold that scales inference-time compute to over 1 million tokens per problem, allowing it to approach the performance of Gemini-3-Pro in certain configurations.
• Specialized Training: Post-trained using a combination of supervised fine-tuning and reinforcement learning with a reasoning cache on a diverse set of Olympiad proof problems.

Good for

• Automated Theorem Proving: Ideal for tasks requiring rigorous mathematical proof generation.
• Mathematical Research Assistance: Can serve as an assistive tool for researchers and students working on complex mathematical problems.
• Domain-Specific Reasoning: Excels in highly specialized reasoning domains where logical deduction and proof construction are paramount.