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

A team led by Konstantin Willeke has built OmniMouse, a multi-modal brain model trained on more than 150 billion neural tokens from the mouse visual cortex, achieving state-of-the-art performance across neural prediction, behavioral decoding, and neural forecasting tasks [1][2]. The model was trained on recordings from 3.1 million neurons in 73 mice across 323 sessions, using natural movies, images, parametric stimuli, and behavioral data [1][2]. OmniMouse supports three test-time regimes — neural prediction, behavioral decoding, and neural forecasting — and can handle any combination of the three [2]. It outperforms specialized baselines across nearly all evaluation regimes, according to the paper posted on arXiv, an open-access repository for electronic preprints that has hosted more than two million articles as of late 2021 [2][6]. The researchers report that performance scales reliably with additional data, but gains from increasing model size saturate [1][2]. This finding inverts the standard scaling story observed in large language models, where massive datasets make parameter scaling the primary driver of progress [2][8]. In brain modeling, even in the relatively simple mouse visual cortex, models remain data-limited despite the vast recordings now available [2]. The paper, submitted on April 20, 2026, and revised on June 14, 2026, raises the possibility of phase transitions in neural modeling [1][2]. The authors suggest that larger and richer datasets might unlock qualitatively new capabilities, paralleling the emergent properties seen in large language models [2][8]. Code for the project is available on GitHub [2]. The work appears on arXiv under the Quantitative Biology section, with the submission history showing an initial version of 2,823 KB and a revised version of 2,824 KB [1]. The repository, which began on August 14, 1991, now receives about 24,000 submissions per month as of November 2024 [6].