Summary: Meta’s FAIR team released TRIBE v2 on March 26, a foundation model that predicts fMRI responses to video, audio, or text. It trained on brain-imaging data from more than 700 healthy volunteers, supports zero-shot prediction for unseen subjects, languages, and tasks, and reaches roughly 70x the resolution of similar models. Weights, code, a paper, and a demo shipped under CC BY-NC 4.0, barring commercial use.
On March 26, Meta’s FAIR team shipped TRIBE v2, a foundation model built to forecast human brain activity: hand it video, audio, or text, and what comes back is a set of predicted fMRI responses. The model learned from brain-imaging data gathered from more than 700 healthy volunteers, predicts zero-shot for subjects, languages, and tasks it has never encountered, and reaches roughly 70x the resolution of similar models. Model weights, training and evaluation code, the paper, and an interactive demo are all openly available.
For AI researchers and neuroscientists alike, the move carries “in-silico neuroscience” another step forward: before booking scanner time and recruiting participants to test a hypothesis, you can run the simulation on the model first. Meta describes it as a digital twin of neural activity.
What Kind of Model TRIBE v2 Actually Is
TRIBE is short for TRansformer for In-silico Brain Experiments. Architecturally, it is a tri-modal encoding model: state-of-the-art text, audio, and video encoders are folded into a single Transformer, whose multimodal features are then projected onto the cortical surface to predict what the brain does while someone watches a movie, listens to a podcast, or reads. The ventral visual stream and the auditory stream get the spotlight. The name is meant literally — the repository frames it as a way to run brain experiments computationally, echoing how in-silico screening took over part of wet-lab drug discovery.
The predecessor was a great deal smaller: the award-winning Algonauts 2025 version trained on low-resolution fMRI from only four individuals. TRIBE v2 keeps that technical line and scales it up to foundation-model size — absorb a large pool of subjects and stimuli first, then predict for whatever new input arrives.
The Training Base: 700-Plus People, Three Modalities
Training drew on more than 700 healthy volunteers, and the stimuli span images, podcasts, videos, and text — naturalistic material covering vision, hearing, and language rather than artificial laboratory stimuli. Scale is what makes zero-shot possible at all: a model has to observe enough person-stimulus combinations before it can predict for a stranger.
The engineering decisions are pragmatic. Predictions sit on the fsaverage5 standard mesh (roughly 20,000 vertices), and every output is shifted back five seconds to offset hemodynamic lag. What the model captures, in other words, is the response pattern of an “average” brain — not any one individual’s full neural activity.
Two Numbers to Note: 70x Resolution and Zero-Shot Transfer
Two numbers deserve attention. One: resolution improves roughly 70x over similar models, letting the model tell apart finer neural differences instead of issuing broad regional activation. Two: zero-shot generalization — new subjects, new languages, and new tasks it never saw during training all get predictions directly, with no retraining or recalibration. Meta’s report says it consistently outperforms standard modeling approaches across settings.
For researchers, zero-shot is the true watershed. Brain-encoding models historically needed retraining for nearly every new experiment; only a model that carries over to unseen people can plausibly act as a simulator. The economics reinforce this: scanning a single new subject costs thousands of dollars and weeks of scheduling, while a simulation rerun costs cents of compute — exactly the gap “in-silico neuroscience” exists to close.
What Shipped, and Where the License Bites
The package is complete: 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. It ships under CC BY-NC 4.0 — research and redistribution are free, commercial use is barred. Research is plainly the intended lane, and the hosted demo lowers the entry bar for neuroscientists who don’t run their own GPU clusters.
The license deserves a close read: what shipped is a research tool, not product-grade open weights. Teams hoping to wrap it into commercial neurointerfaces or attention-monitoring products will have to negotiate separately.
What In-Silico Neuroscience Is Good For
Meta names three directions. First, neuroscientists and clinical researchers can put hypotheses to the test without scanning human subjects, letting computational simulation accelerate the research timetable. Second, what the brain teaches us can flow back into better AI systems. Third, for neurological disorders affecting hundreds of millions of people, simulation can tighten the iteration loop on treatment research. NeuroscienceNews’ coverage also flags potential value for brain-computer interface work and conditions such as aphasia.
One boundary has to be drawn: TRIBE v2 performs encoding — predicting how the brain responds to input — not decoding of private thought. It can answer “how will auditory cortex react to this clip”; it cannot answer “what is this person thinking.” Holding that line is what keeps this class of model defensible in ethical debate.