Meta’s FAIR team released TRIBE v2 on March 26: a foundation model that predicts human brain activity. Feed it video, audio, or text, and it outputs predicted fMRI responses. It was trained on brain-imaging data from more than 700 healthy volunteers, makes zero-shot predictions for new subjects, languages, and tasks it has never seen, and delivers roughly 70x the resolution of similar models. Weights, training and evaluation code, the paper, and an interactive demo are all public.
For AI researchers and neuroscientists, this pushes “in-silico neuroscience” a step further: to test a hypothesis, you no longer need to book scanner time and recruit participants — you can run the simulation first. Meta calls it a digital twin of neural activity.
What TRIBE v2 Is
TRIBE stands for TRansformer for In-silico Brain Experiments. It is a tri-modal encoding model: state-of-the-art text, audio, and video encoders are merged into a single Transformer that maps multimodal features onto the cortical surface, predicting brain responses while a person watches movies, listens to podcasts, or reads. Its focus is the ventral visual stream and the auditory stream. The name is literal — the repository positions it as a tool for running brain experiments computationally, much as in-silico screening replaced parts of wet-lab drug discovery.
The predecessor was far smaller. The award-winning Algonauts 2025 version was trained on low-resolution fMRI from four individuals. TRIBE v2 scales the same technical line to foundation-model size — absorb enough subject-stimulus combinations, then predict for arbitrary new inputs.
Training on 700 People and Three Modalities
The training data comes from more than 700 healthy volunteers, with stimuli spanning images, podcasts, videos, and text — naturalistic material across vision, audition, and language rather than artificial lab stimuli. That scale is the precondition for zero-shot capability: only a model that has seen enough person-stimulus pairs can predict for people it has never encountered.
The engineering choices are pragmatic. Predictions live on the fsaverage5 standard mesh (roughly 20,000 vertices) and are uniformly shifted back five seconds to compensate for hemodynamic lag. In other words, it predicts the response pattern of an “average” brain, not one individual’s full neural activity.
Zero-Shot Prediction and 70x Resolution
Two numbers stand out. First, a roughly 70x resolution increase over similar models, which lets the model distinguish finer-grained neural differences instead of coarse regional activation. Second, zero-shot generalization: for new subjects, new languages, and new tasks, it predicts directly, with no retraining or recalibration. Meta reports it consistently outperforms standard modeling approaches.
Zero-shot is the dividing line for researchers. Most prior brain-encoding models had to be retrained for every new experiment; only a model that transfers to unseen people can plausibly serve as a simulator. The economics matter too: scanning a single new subject costs thousands of dollars and weeks of scheduling, while rerunning a simulation costs cents of compute — which is exactly the gap “in-silico neuroscience” is meant to close.
What Was Released, and the License Catch
The release is thorough: model weights on Hugging Face (facebook/tribev2), training and evaluation code on GitHub (facebookresearch/tribev2), the arXiv paper, and an interactive demo with a Colab notebook at aidemos.atmeta.com. The license is CC BY-NC 4.0 — free for research and redistribution, commercial use prohibited. Research is clearly the intended lane, and the hosted demo lowers the barrier for neuroscientists who do not manage GPU clusters.
The license matters: this is a research-tool release, not product-grade open weights. Teams that want to wrap it into commercial neurointerfaces or attention-monitoring products will need a separate arrangement.
Practical Uses for In-Silico Neuroscience
Meta lays out three application directions. Neuroscientists and clinical researchers can test hypotheses without scanning human subjects, using simulation to compress research cycles. Brain insights can flow back into improving AI systems. And for neurological disorders affecting hundreds of millions of people, simulation can shorten the iteration loop on treatment research. NeuroscienceNews’ coverage also points to potential value for brain-computer interface work and disorders such as aphasia.
One boundary deserves emphasis: TRIBE v2 does encoding — predicting how the brain responds to input — not decoding private thoughts. It answers “how will auditory cortex react to this clip,” not “what is this person thinking.” That line is what keeps this class of model on solid ethical footing.
Sources
- Introducing TRIBE v2: A Predictive Foundation Model — Meta AI
- facebookresearch/tribev2 — GitHub
- Meta TRIBE AI brain decoding — Neuroscience News
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.