Samsung AI Labs unveils its open source tiny recursive model (TRM), a 7-million-parameter system that rivals giants like GPT and Gemini by thinking recursively instead of scaling endlessly.
In a bold move that challenges the “bigger is better” philosophy of artificial intelligence, Samsung AI Labs in Montreal has unveiled the open-source Tiny Recursive Model (TRM), a compact 7 million-parameter model that performs on par with systems thousands of times larger.
Released on GitHub alongside a research paper on arXiv, TRM represents a new direction in AI evolution, where intelligence is enhanced through recursive reasoning rather than computational scale. The model’s core innovation lies in its ability to refine its own answers iteratively, asking, “Is my answer good? If not, can I make it better?”
Despite its remarkably small size, nearly 10,000× smaller than leading LLMs such as OpenAI’s GPT, Google’s Gemini, and Anthropic’s Claude, TRM has demonstrated impressive benchmark results: 44.6% accuracy on ARC-AGI-1, 7.8% on ARC-AGI-2 (where most LLMs score below 5%), and 87% on Sudoku-Extreme. It even outperformed DeepSeek R1, Gemini 2.5 Pro, and o3-mini in reasoning tests, all while running efficiently on a laptop.
By simplifying the complex multi-network structures used in earlier recursive systems like the Hierarchical Reasoning Model (HRM), TRM proves that smarter, smaller architectures can achieve exceptional reasoning without relying on massive datasets or power-hungry infrastructure.
As Samsung emphasises, “Instead of building bigger models, one can build smarter, more efficient ones that think recursively.” The open-source release of TRM signals a paradigm shift, from closed, billion-parameter ecosystems to small, smart, and open AI models built for everyone.
