UnipatAI/UniMath-35B-A3B

UniMath-35B-A3B: Olympiad-Level Mathematical Reasoning

UniPat AI's UniMath-35B-A3B is a 35.1 billion parameter model (with 3 billion activated per token) engineered for advanced mathematical reasoning, particularly at the olympiad level. It is built upon the Qwen3.6-35B-A3B architecture and further refined through a unique post-training process.

Key Capabilities and Differentiators

• Proof-Synthesis Data: The model is trained on detailed proof-evolution data, supervising not just the final proof but the entire trajectory, including initial attempts, self-critique, obstruction diagnosis, and targeted repair. This fosters a dynamic, self-evolving reasoning process.
• Test-Time Self-Evolution: During inference, UniMath-35B-A3B actively audits, repairs, and reconciles its own proof attempts. Instead of blind resampling, it refines candidate proofs by addressing identified obstructions and merges successful elements from different approaches.
• Adaptive Inference Compute: The self-evolution loop intelligently allocates computational resources, spending more on genuinely hard problems and short-circuiting to synthesis when sufficient self-verified proofs are accumulated.

Performance

UniMath-35B-A3B demonstrates exceptional performance in challenging mathematical competitions, reaching human gold-medal-contestant levels on IMO 2025 (35/42) and USAMO 2026 (36/42). It also scores 86.0% on IMO-ProofBench.

Use Cases

This model is ideal for applications requiring advanced, step-by-step mathematical problem-solving and proof generation, particularly in areas demanding high-level logical deduction and self-correction capabilities.