Models

← All models

QED-Nano

LM-Provers · 2026-02-12 · 4B parameters

QED-Nano is a dense 4-billion-parameter theorem-proving checkpoint based on Qwen3-4B-Thinking-2507 and post-trained with supervised fine-tuning and reinforcement learning on Olympiad-level proof problems. Its separately evaluated agent scaffold can scale inference-time computation up to two million tokens per problem, but the released model itself is the checkpoint recorded here.

Benchmark scores