Meta FAIR's TRIBE model wins Algonauts 2025

TRIBE (Trimodal Brain Encoder) is an open-source 1B-parameter neural network from Meta FAIR that just took first place at the Algonauts 2025 brain modeling competition. It predicts fMRI brain responses across multiple modalities, cortical areas, and individuals as people watch movies.

How it works

Instead of building sensory feature extractors from scratch, TRIBE combines pretrained representations from three of Meta's foundation models: Llama 3.2 for text, Wav2Vec2-BERT (from Seamless) for audio, and V-JEPA 2 for video.

The system was trained on the Courtois NeuroMod dataset, which features roughly 80 hours of high-resolution spatio-temporal fMRI brain recordings per subject during movie watching. It marks the first deep neural network trained to handle this breadth of sensory inputs, brain regions, and subjects simultaneously.

The team has published the complete source code on GitHub alongside the research paper on arXiv.

Why it matters

Neuroscience has traditionally relied on separate, handcrafted models for visual or auditory neural encoding. TRIBE shows that stitching together general-purpose multimodal foundation models is enough to beat specialized brain architectures at predicting actual neural responses.

For cognitive scientists and neural engineering teams, it delivers a powerful open baseline right out of the box. That said, dealing with 80 hours of fMRI data per participant alongside multiple heavy foundation models means it remains firmly a research-grade tool rather than something you run on a laptop.