Samsung has just stunned the technology world, it announced a super-small AI model (Tiny AI Model) that can surpass both Google Gemini 2.5 Pro and DeepSeek R1 in a logical thinking test. Despite only possessing 7 million parameters - 10,000 times smaller than large language models (LLM) - Samsung's new AI shows outstanding reasoning ability, proving that "size does not determine intelligence". According to information from the academic platform Arxiv, research from Samsung Institute of Integrated Technology (SAIT) led by Dr. Alexia Jouricœur Martino, introduces the Tiny Recursive Model (TRM) - a super-small AI architecture capable of self-learning and self-improvement. Instead of just predicting text, TRM can analyze, self-reflect, and edit its own results up to 16 times to minimize errors. In tasks like logic puzzles and mazes, TRM creates a hidden “inference memory” that continually re-evaluates previous results to get closer to the correct answer.

The results were astonishing: TRM scored 45% accuracy on the ARC-AGI-1 test, far ahead of Gemini 2.5 Pro (37%) and DeepSeek R1 (15.8%). With the new ARC-AGI-2 benchmark, TRM still leads with 8%, while Gemini only scored 4.9%. “The intelligence of AI is not in the size of the model, but in the ability to iterate the thought process to improve itself,” said Dr. Martino.

Unlike giant LLMs with thousands of processing layers, TRM has only two layers of neural networks, but takes advantage of the iteration mechanism to achieve higher performance. This approach is simpler than advanced techniques like Chain-of-Thought or Test-Time Computing that Google and OpenAI are testing, but its effectiveness is no less.

After a period of silence since the launch of the Gauss AI line two years ago, Samsung's new research shows that the company is still quietly strengthening its core AI capabilities. Although there are no plans to commercialize it, TRM is assessed to have strong potential for application in autonomous cars, robots, mobile devices, and on-device AI, opening up a new direction for high-performance, low-energy AI - an area where Samsung can completely lead.