Samsung’s tiny recursive model beats monster LLMs

Samsung’s tiny new recursive model beats monster LLMs, has a mystery at its heart

Pay attention but beware hype-mongers.

Edward Targett

Oct 08, 2025 - 3 min read

A “tiny recursive model” released by a Samsung researcher just outperformed DeepSeek-R1 and Gemini 2.5 Pro on the ARC-AGI reasoning benchmark, whilst being 0.01% of the size; a mere 7-million parameters.

The hiccup: It took a heck of a long time to do as it made numerous “recursive” loops: “Experiments on ARC-AGI were ran [sic] for around 3 days with 4 H100 with 80Gb of RAM,” an October 6 paper showed.

This post is for paying subscribers only

Join peers managing over $100 billion in annual IT spend and subscribe to unlock full access to The Stack’s analysis and events.

Subscribe now

Already a member? Sign in