VentureBeat
Follow
Samsung AI researcher's new, open reasoning model TRM outperforms models 10,000X larger — on specific problems
Alexia Jolicoeur-Martineau of Samsung developed the Tiny Recursion Model (TRM), a small open-source AI model. TRM, with only 7 million parameters, rivals much larger models like OpenAI's o3-mini and Google's Gemini 2.5 Pro. The model excels in structured, grid-based reasoning tasks like Sudoku and puzzles. TRM utilizes a two-layer architecture with recursive refinement of its own predictions, replacing the need for larger, more complex models. This recursive approach allows it to achieve high performance on specific benchmarks. TRM's open-source availability under an MIT license enables widespread use and modification. TRM's success stems from minimalism, reducing complexity for better generalization and avoiding overfitting. The model's success sparked debate, with some praising its efficiency, while others questioned the scope of its applicability. Future research may explore generative variants and scaling laws for recursion, building on TRM's framework.
TRM proves that carefully designed, recursive thinking can be more effective than simply increasing model size.
