A Foundation Model of the Brain — Inside Meta FAIR's TRIBE v2
— by Vivax
Meta FAIR's TRIBE v2 is a tri-modal foundation model that predicts high-resolution fMRI brain activity from any video, audio, or text — trained on over 1,000…
Meta FAIR has released TRIBE v2, a tri-modal foundation model of the human brain. TRIBE v2 is trained to predict high-resolution fMRI brain activity from any video, audio, and/or text stimulus — a single model that captures whole-brain responses across naturalistic movies, podcasts, and controlled experiments. Built on the architecture that won first place out of 263 teams at the Algonauts 2025 challenge, it supersedes traditional linear encoding models by several-fold in accuracy, and generalizes zero-shot to novel stimuli, tasks, and subjects.
The design is elegant. TRIBE v2 wraps three frozen, pretrained encoders — V-JEPA 2 for video, W2v-BERT for audio, and Llama 3.2 for language — and feeds their embeddings into a shared Transformer, followed by a per-subject 'Subject Block' that predicts that individual's brain activity. Only about 1 billion parameters are learnable; the encoders stay frozen. Trained on a unified corpus of over 1,000 hours of fMRI across 720 subjects, the model learns how the brain integrates sight, sound, and meaning into a coherent response to the world. In that sense TRIBE v2 is, in part, a world model: it does not just label a stimulus, it predicts how a biological system reacts to the unfolding world — and, run in silico, it recovers results that took decades of empirical neuroscience to establish.
This is the same thesis Vivax is building on. Our foundational clinical world model aims to understand the patient and the physical clinical environment as one continuous, grounded picture — and models like TRIBE v2 show that predicting a living system's response to rich multimodal input is now tractable at foundation-model scale. TRIBE v2's code, weights, and an interactive demo are all open, which makes it a powerful reference point for the in-silico, patient-grounded AI we are developing at Vivax.