Local LLM Runs for 63 Hours Attempting to Solve the Riemann Hypothesis

A Reddit user ran a 4-bit quantized Qwen 27B model continuously for 63 hours on a single RTX 3090 GPU in an attempt to solve the Riemann Hypothesis. Although the model failed to produce a proof, it generated over 50 million tokens of autonomous reasoning trajectories and self-corrections, which were published as a open dataset on Hugging Face. While using a quantized 27B model to tackle a Millennium Prize Problem is highly impractical, the resulting 50M-token trajectory dataset offers a unique resource for analyzing long-context retention and agent self-correction. It illustrates both the intriguing persistence of local LLM agents and the severe limitations of current open models on deep mathematical reasoning. The setup utilized a 100K context window and processed over 50 million tokens without hallucinating a false solution or crashing. The author noted that the model continuously attempted new strategies and self-corrected code errors, and the released dataset includes internal memories, code executions, and strategic planning trajectories.