In a move that challenges the prevailing "bigger is better" philosophy in artificial intelligence, Samsung AI Labs has unveiled its Tiny Recursive Model (TRM), a compact AI model boasting only 7 million parameters. This open-source model rivals the performance of systems thousands of times larger, like OpenAI's GPT and Google's Gemini, demonstrating that smarter, more efficient architectures can achieve exceptional reasoning without relying on massive datasets or power-hungry infrastructure.
The TRM project explores whether recursion can stand in for sheer scale. Instead of stacking more layers, the model refines its reasoning step by step. Each pass through the network generates a new version of the answer, which then becomes input for the next pass.
TRM begins by examining a problem and providing a quick, initial answer. The model then utilizes a hidden "scratchpad" to contemplate and debate with itself, checking the validity of its reasoning. It continuously asks itself, "Is my answer satisfactory? Can I improve it?" at every step of sentence generation. Instead of providing a fixed answer, it mimics human-like deliberation. Deep supervision, which provides feedback at multiple steps, and adaptive halting, which allows the model to determine when to stop refining, further enhance the model's learning process.
The TRM has demonstrated impressive benchmark results. It achieved 44.6% accuracy on ARC-AGI-1 and 7.8% on ARC-AGI-2, outperforming DeepSeek R1, Gemini 2.5 Pro, and o3-mini in reasoning tests. It also showed 87% accuracy on Sudoku-Extreme. The model was tested on logic puzzles and reasoning challenges, and it performs well on solving hard Sudoku puzzles and finding paths through tricky mazes. In the ARC-AGI test, TRM scored 45% accuracy, surpassing Gemini (37%) and o3 (34.5%).
The research team revealed that the model was trained in two days using four NVIDIA H100 GPUs at a cost of approximately $500. Alexia Jolicoeur-Martino, the lead researcher, stated that a small model can achieve significant results without large costs by recursively learning from itself and updating answers over time.
TRM's high scores were in puzzle-type tasks with explicit input-output rules. While it is a research-stage model specialized for solving specific problems, analysts suggest it could be applied to on-device AI directly embedded in devices.
Samsung's TRM signals a potential paradigm shift in AI development, moving away from closed, billion-parameter ecosystems towards small, smart, and open AI models built for everyone. The open-source release of TRM on GitHub includes code and training details.
