A researcher at Samsung SAIT AI Lab has shown that small models can be just as powerful as giants. Her Tiny Recursion Model (TRM), with just 7 million parameters, outperformed systems like Gemini 2.5 Pro, Claude Opus 4, and even GPT-5 (Low) at solving complex logic and recursion problems. TRM is built on the principle of recursive thinking — the model doesn’t just produce an answer, but refines it several times, feeding the results of the previous step back into its own neural network. This approach allows the system to learn from its own mistakes in real time, minimizing the need for billions of parameters.
In the ARC-AGI-1 tests, which test the ability of AI to solve problems that are understandable to humans but difficult for machines, TRM scored 45%, almost equal to GPT-5 (Low). And in the more complex ARC-AGI-2, it scored 8%, leaving leading commercial models behind.
The most impressive thing was the price-performance ratio: it cost only a fraction of a cent to complete the tasks, while other systems cost from 25 cents to $1.
The model’s author, Alexia Joliqueur-Martineau, notes:
> “The idea that only huge models with billions of parameters can solve complex problems is a trap. Sometimes in artificial intelligence, ‘less is more.’”
The development of TRM is an attempt to return AI to efficiency, not uncontrolled scaling. In recent years, leading companies have invested billions of dollars in training large language models, but such systems often have problems with flexibility, power consumption and overtraining.
Samsung has shown that a focus on algorithmic optimization can produce the same results as large LLM systems, while consuming thousands of times fewer resources. Samsung’s work is a vivid example that the future of AI is not just in size, but in intelligence. The Tiny Recursion Model concept could fundamentally change the approach to creating new systems — from mobile assistants to industrial analytical tools.
If this direction develops, then in a few years “small AIs” could replace large cloud models, making intelligence available even on simple devices.
