OmniMouse: Scaling properties of multi-modal, multi-task Brain Models on 150B Neural Tokens

A team of 21 researchers from Baylor College of Medicine, University of Tübingen, and other institutions has released OmniMouse, a multi-modal, multi-task brain model trained on an unprecedented dataset of 3.1 million neurons from the visual cortex of 73 mice. The dataset spans 323 recording sessions and includes 150 billion neural tokens captured during natural movies, images, parametric stimuli, and behavioral tasks. OmniMouse supports three regimes at test time: neural prediction, behavioral decoding, and neural forecasting—or any combination thereof. It achieves state-of-the-art performance, outperforming specialized baselines across nearly all evaluation regimes.

Crucially, OmniMouse inverts the standard AI scaling story. In language and vision, massive datasets make parameter scaling the primary driver of progress. Here, performance scales reliably with more data, but gains from increasing model size saturate. This suggests that brain modeling—even in the relatively simple mouse visual cortex—remains data-limited despite vast recordings. The authors note that this systematic scaling raises the possibility of phase transitions in neural modeling, where larger and richer datasets might unlock qualitatively new capabilities, paralleling emergent properties seen in large language models. Published at ICLR 2026, the code is available on GitHub.