What Samsung’s Tiny Model Taught Me About Task Fit

The AI industry keeps returning to the same lever: bigger models, more parameters, larger training sets. Samsung’s 7-million-parameter model pulled a different one. On specific abstract-reasoning benchmarks, it beat models with thousands of times more parameters.

The Tiny Recursive Model (TRM) doesn’t solve problems in one pass. It repeatedly updates a latent state and candidate answer, trained with supervision over repeated improvement steps. That is numerical computation over a structured task, not a written critique or a conversational scratchpad.

On the ARC-AGI benchmark, a set of abstract reasoning tests that trip up even the best LLMs, TRM scored 45% on ARC-AGI-1 and 8% on ARC-AGI-2. Better than DeepSeek-R1, o3-mini, and Gemini 2.5 Pro, all running billions of parameters. On Sudoku-Extreme with 1,000 training examples, it hit 87.4% accuracy versus its predecessor’s 55%.

Alexia Jolicoeur-Martineau, the researcher behind TRM, argues that these results challenge the assumption that difficult structured tasks always need expensive foundation models. The benchmark scope matters to that argument.

You don’t write down the first answer that pops into your head when solving a hard problem. You draft something, spot the holes, rethink it, try again. TRM does this in a two-layer neural network. Autoregressive LLMs generate tokens through repeated forward passes, and reasoning systems can also revise or search. TRM’s distinction is a small network trained to refine a structured answer, not recursion versus a universal one-pass baseline.

TRM won’t write your emails or debug your code. It handles structured, grid-based reasoning: Sudoku, mazes, abstract pattern recognition. For these problems, the paper shows that a 7-million-parameter model can beat much larger cloud models. That is a narrower and more interesting result than claiming small models win at everything.

I’ve been testing Apple’s Foundation Models framework. What I like most is the privacy. It runs entirely offline. Your prompts stay on your device. Apple doesn’t use your data to train models.

A cloud model requires sending the request to a remote service; retention varies by provider, product, and configuration. Apple’s 3-billion parameter model runs natively in iOS 26, iPadOS 26, and macOS 26. It feels like it’s mine, not something I’m renting access to.

Samsung’s TRM is open source under an MIT license. Apple’s Foundation Models framework also gives me a practical way to build with an on-device model. Neither replaces the general-purpose models I use for coding, but both make the same point: for a bounded problem, I would rather run the smallest model that can do the work.