Samsung’s Tiny AI Model Outsmarts Gemini 2.5 Pro and o3-mini — Trained on Just 4 GPUs!

Samsung Tiny Recursive Model

In a surprising twist to the AI race, Samsung’s AI Lab in Montreal has proven that bigger isn’t always better.
Their new creation — a Tiny Recursive Model (TRM) with only 7 million parameters — has outperformed some of the world’s most powerful AI systems, including Google’s Gemini 2.5 Pro and OpenAI’s o3-mini-high, on a critical reasoning benchmark.

The results mark a milestone in AI research, showing that small, efficient models can reason and adapt as effectively — and sometimes more — than trillion-dollar behemoths.

Small but Smarter: Meet TRM

Samsung’s Tiny Recursive Model (TRM) isn’t a massive, resource-hungry model. It’s compact — trained with just 4 NVIDIA H100 GPUs over two days, costing less than $500.

Despite its size, TRM scored an impressive 45% accuracy on the ARC-AGI 1 benchmark — a test designed to evaluate human-like reasoning, pattern recognition, and abstract thinking.

To put that into context:

  • Gemini 2.5 Pro scored 37%
  • OpenAI’s o3-mini-high managed 34.5%
  • DeepSeek-R1 hit 15.8%

On the tougher ARC-AGI 2 benchmark, TRM achieved 7.8%, still outperforming Gemini (4.9%) and o3-mini (3%).
Currently, xAI’s Grok 4 leads the charts with 66.7% and 16% on ARC-AGI 1 and 2, but Samsung’s result is groundbreaking given the scale difference.

The Secret: Thinking in Loops

So what makes TRM special?
It’s not about size — it’s about structure.

Instead of dumping all reasoning into one forward pass like most models, TRM uses a recursive approach — it thinks, checks, and improves.
The model starts with an initial guess, reviews it, then refines the answer step-by-step — a bit like how humans revise their work.

According to the study,

“This recursive process allows the model to progressively improve its answer, addressing errors from previous iterations — achieving higher reasoning ability while minimizing overfitting.”

In simple terms: TRM doesn’t need more neurons. It just needs to think smarter.

A $500 Breakthrough in the Billion-Dollar AI Race

The training details alone have stunned the AI community.
Alexia Jolicoeur-Martineau, the paper’s author from Samsung’s Advanced Institute of Technology, confirmed that TRM was trained in just two days on 4 H100 GPUs — at a total cost of under $500.

Compare that with the tens of millions required to train models like GPT-4 or Gemini, and the achievement feels almost impossible.

AI researcher Sebastian Raschka reacted on X (formerly Twitter):

“Yes, it’s still possible to do cool stuff without a data center.”

This experiment demonstrates that innovation, not infrastructure, will define the next phase of AI progress.

Why It Matters: “Less Is More” in AI

The research paper — fittingly titled “Less Is More” — argues that architectural innovation can outperform brute-force scaling.
TRM proves that intelligent recursion and task specialization can make small models outperform general-purpose giants.

Industry experts are taking notice.
Deedy Das, Partner at Menlo Ventures, commented:

“Most AI companies today rely on huge general-purpose models. For specific tasks, smaller models may not just be cheaper — they can actually be better.”

He predicts a coming wave of micro-AI models — lightweight systems fine-tuned for narrow tasks like PDF extraction, time-series forecasting, or chat moderation.
These “tiny experts” could complement large LLMs, improve performance, and allow startups to train proprietary AI for under $1000.

A Shift in the AI Mindset

Samsung’s achievement marks a philosophical change in how the AI world measures intelligence.
For years, success meant building larger models with more data and parameters. TRM challenges that belief.

By relying on recursive reasoning instead of raw size, Samsung’s researchers have unlocked a way to make AI both smarter and sustainable.
And that’s critical — because as AI becomes more embedded in devices, from phones to IoT, efficiency will matter more than scale.

In essence, TRM shows that the next AI breakthrough may not come from a trillion-dollar cluster — but from a clever design that teaches machines how to think like humans do — one step at a time.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top