Qwen AI Spent 63 Hours on a Math Problem Nobody Can Solve

A person who likes tinkering with AI gave a free, open-source model called Qwen a nearly impossible homework assignment: crack the Riemann hypothesis, a legendary math problem that has stumped experts for over 150 years and carries a $1 million prize. The twist is where it ran. Not in a giant data center, but on an RTX 3090, a graphics card many gamers already own. It ran for 63 hours straight, chewing through more than 50 million words worth of text.

It did not solve the problem. Not even close. But the way it failed is the interesting part. It never hallucinated an answer — meaning it never confidently made something up, the way chatbots sometimes do. Instead it kept generating fresh strategies, reorganized its own notes, and several times noticed and corrected its own errors. For 63 hours, with nobody watching, it just kept working the problem.

Why should you care? Because this is a small preview of AI that works alone for days instead of seconds. The tech world calls these agents — AI that can take actions on its own. If a cheap, open model can stay focused and stay honest for 63 hours on a consumer graphics card, that same skill could eventually handle long, tedious tasks for you: digging through contracts, monitoring accounts, chasing down answers across dozens of documents. Crucially, it knew when to say I do not know. That is the trait that makes an AI safe to leave running.

The catch is real. Nothing was solved. The math was never verified by experts, this was one person's experiment rather than a peer-reviewed study, and running a card hot for nearly three days is not free. The claim that a major unsolved math problem will fall to AI within 12 months is a hope, not a fact. Still, the direction is clear: AI is learning to grind, not just to chat.