Samsung’s AI lab in Montreal dropped a game-changer – the Tiny Recursive Model (TRM), a super small AI with just 7 million parameters that outsmarts huge models on tough reasoning tests. It smoked Google’s Gemini 2.5 Pro and OpenAI’s o3-mini on the ARC-AGI benchmark, scoring 45% on ARC-AGI-1 and 7.8% on ARC-AGI-2, while those giants hit just 4.9% and 3% on the harder one. Trained for under $500 on 4 NVIDIA H100 GPUs in two days, TRM shows you don’t need massive setups to nail complex puzzles like Sudoku or mazes.
How This Little Guy Works
TRM uses a simple trick: it starts with a rough guess, then loops back to fix mistakes up to 16 times, refining its thinking and answer step by step. With only two layers, it avoids overfitting and crushes benchmarks – 87.4% on extreme Sudoku and 85.3% on hard mazes. It’s way more efficient than big LLMs that chug through tokens and flop on logic puzzles. Built by researcher Alexia Jolicoeur-Martineau, it’s a tweak on older models but dials up the smarts without the bloat.
Who Can Try It and When?
The paper’s out now on arXiv for anyone to check – grab it at arxiv.org/abs/2510.04871. It’s research for now, but could hit phones or gadgets soon, since it’s tiny and cheap to run. Open to devs worldwide, with code likely coming if it follows AI trends.
Why This Flips AI on Its Head
TRM proves “less is more” – small models can beat beasts on specific jobs, cutting costs and power use. It could spark cheaper AI for startups, like custom tools for PDFs or forecasts, as investor Deedy Das points out. Samsung’s shaking up the race, pushing for smarter, not bigger, AI – one puzzle at a time.
