Meta's Brain Model Counts 720 Subjects. It Learned From 25.

Meta released TRIBE v2 on March 26, and the paper reached arXiv on May 5. The model takes a video, a soundtrack, or a transcript and predicts how a human cortex would respond. The abstract says it was built with “a unified dataset of over 1,000 hours of fMRI across 720 subjects.” Meta’s announcement contrasts the first TRIBE, “trained on the low-resolution fMRI recordings of four individuals,” with a new model that uses “a massive dataset of more than 700 healthy volunteers.” A reader could easily take that as four training subjects growing to seven hundred.

Table 1 of the paper says something else. The model was trained on 25 subjects. The other 695 are the people it was tested on.

Deep to learn, wide to check

The paper splits its eight datasets in two. Four “deep” ones, where a few people spent a long time in the scanner, were used for training: Courtois NeuroMod (4 subjects, 268.7 hours), Lebel2023 (8 subjects, 85.8 hours), BoldMoments (10 subjects, 61.9 hours), and Wen2017 (3 subjects, 35.2 hours). That is 25 subjects and 451.6 hours, about 18 hours each. Four “broad” ones, where many people were scanned briefly, were held out for testing: Narratives (321 subjects), LPP (112), HCP (176), and NNDb (86). That is 695 subjects and 666.1 hours, under an hour each.

TRIBE v2's 720 subjects split into 25 for training and 695 for testing Two stacked horizontal bars built from Table 1 of the TRIBE v2 paper. Subjects: 25 in the four training datasets and 695 in the four held-out test datasets, 720 in all. Hours of fMRI: 451.6 in training and 666.1 in testing, 1,117.7 in all. By the writer's arithmetic that is about 18.1 hours per subject in training and 0.96 hours per subject in testing. Below the bars, the four training datasets are listed as Courtois NeuroMod with 4 subjects and 268.7 hours, Lebel2023 with 8 subjects and 85.8 hours, BoldMoments with 10 subjects and 61.9 hours, and Wen2017 with 3 subjects and 35.2 hours; the four test datasets are Narratives with 321 subjects and 146.6 hours, LPP with 112 subjects and 180.2 hours, HCP with 176 subjects and 178.7 hours, and NNDb with 86 subjects and 160.6 hours. TRIBE V2 · WHO THE 720 SUBJECTS ARE · TABLE 1 OF THE PAPER Subjects fMRI hours Hours per subject 25 trained Training: 25 subjects across four deep datasets Held-out test: 695 subjects across four broad datasets 695 held out Training: 451.6 hours of fMRI Held-out test: 666.1 hours of fMRI 451.6 h trained 666.1 h held out 18.1 in training 0.96 in testing TRAINING, DEEP DATASETS (SUBJECTS, HOURS) CNeuroMod 4, 268.7 · Lebel2023 8, 85.8 · BoldMoments 10, 61.9 · Wen2017 3, 35.2 TEST, BROAD DATASETS (SUBJECTS, HOURS) Narratives 321, 146.6 · LPP 112, 180.2 · HCP 176, 178.7 · NNDb 86, 160.6 Hours per subject is the writer's arithmetic from Table 1; every other figure is as printed there.
The 720 counts how many people the model was checked against. The number that shaped its weights is 25.

Held-out sets are usually where I start poking, and this one is built the right way round. The authors say they kept the wide datasets for testing because training on deep data works better and because a reliable group-average response needs many subjects. The appendix states the comparison plainly: the first TRIBE trained on 4 participants doing one task, and this one trained on 25 participants across four tasks and was tested on 695 held-out participants. My complaint is with the launch copy. The announcement and the demo page never give the 25.

A small head on borrowed features

Little of what makes it a foundation model was learned from those brains. TRIBE v2 runs text through Llama-3.2-3B, audio through Wav2Vec-BERT 2.0, and video through V-JEPA 2 Giant, all frozen, then passes the combined embeddings through an 8-layer transformer and a subject-specific linear layer that maps onto the cortex. Extracting those features took 24 hours on 128 V100 GPUs. Training the brain-facing part took 24 hours on one. By my arithmetic that is 3,072 GPU-hours against 24. The authors also average the video features across image patches, which they say discards position and should hurt prediction in early visual areas that map space directly. What the 25 brains supply is a mapping from representations that other models already learned.

TRIBE v2 pipeline: three frozen pretrained encoders feeding a small transformer trained on brain data A left-to-right flow diagram in two zones. The left zone, frozen and pretrained elsewhere, holds three encoders: Llama-3.2-3B for text, Wav2Vec-BERT 2.0 for audio, and V-JEPA 2 Giant for video, with video patches averaged. An arrow leads to the right zone, trained on 25 subjects: the embeddings are concatenated and passed through an 8-layer, 8-head transformer, then a subject block, or an unseen-subject layer for the group average, then the outputs of 20,484 cortical vertices plus 8,802 subcortical voxels, 29,286 targets in all by the writer's sum. Below the zones, feature extraction took 24 hours on 128 V100 GPUs, 3,072 GPU-hours by the writer's arithmetic, while training took 24 hours on one V100, or 24 GPU-hours. TRIBE V2 · BORROWED VS. TRAINED-ON-BRAINS FROZEN, PRETRAINED ELSEWHERE Text: Llama-3.2-3B timed word embeddings Audio: Wav2Vec-BERT 2.0 60-second chunks Video: V-JEPA 2 Giant 64 frames, patches averaged TRAINED ON 25 SUBJECTS Concatenate, then 8-layer, 8-head transformer modality dropout, positional and subject embeddings Subject block, or an unseen-subject layer the unseen-subject layer gives the group average 20,484 cortical vertices + 8,802 subcortical voxels 29,286 targets in all (writer's sum) Feature extraction, the frozen side: 24 hours on 128 V100 GPUs, 3,072 GPU-hours (writer's arithmetic). Training, the brain side: 24 hours on one V100, 24 GPU-hours. Compute figures are the authors' Methods; the paper reports them in hours, not GPU-hours.
Most of the compute, and most of the learned representation, sits in models other people trained on video, audio, and text.

Two other numbers

The launch materials describe the output in a way the paper does not. Meta’s demo page says the model “predicts whole-brain activity across 70,000 voxels,” a far sharper picture than the “1,000 cortical predictions” of the first version, and the blog claims “a 70x resolution increase as compared to similar models.” The paper’s methods list 20,484 cortical vertices plus 8,802 subcortical voxels as the prediction targets, which sum to 29,286, about 29 times the first version’s 1,000 parcels. The released code returns roughly 20,000 vertices on the fsaverage5 cortical mesh. I couldn’t find 70,000 in the paper’s text, and the launch pages don’t say what it counts.

Prediction targets per time point: 1,000 for TRIBE v1, 29,286 by the paper, 70,000 by the demo page A horizontal bar chart on a linear scale. TRIBE v1 predicted 1,000 cortical parcels. TRIBE v2, by the paper's Methods, predicts 20,484 cortical vertices plus 8,802 subcortical voxels, 29,286 targets in all by the writer's sum, about 29 times v1. TRIBE v2, by Meta's demo page, predicts 70,000 voxels, 70 times v1. The paper's text as read does not contain the 70,000 figure. PREDICTION TARGETS PER TIME POINT · LINEAR SCALE TRIBE v1, paper TRIBE v2, paper's Methods TRIBE v2, demo page TRIBE v1: 1,000 cortical parcels TRIBE v2, paper: 20,484 vertices + 8,802 subcortical voxels = 29,286 TRIBE v2, demo page: 70,000 voxels 1,000 29,286 70,000 Paper: 20,484 cortical vertices plus 8,802 subcortical voxels (Methods), summed by the writer. Demo page: "70,000 voxels." The paper's text, as read, does not contain that figure.
Against the first version's 1,000 parcels, the paper's own target count is about 29-fold. The launch pages say 70.

The model’s name for itself varies too. The blog calls it “a digital twin of human neural activity,” the demo “a digital mirror,” and the paper a “digital model.” Only the last is one the paper’s evidence supports. The code’s README says predictions are for the “average” subject, and a twin of a particular person would not be an average. The paper does show that fine-tuning on up to an hour of a person’s own data improves scores two- to four-fold over a linear encoder trained from scratch on that person. Read off the paper’s figure, though, those fine-tuned scores average between roughly 0.06 and 0.12 in correlation across the test datasets, so several-fold gains start from a low base.

What it predicts well

The strongest result is about the group, not the individual. Given a stimulus, TRIBE v2’s zero-shot output tracks the group-average response better than most individual subjects’ own recordings do. In the 7T HCP data, which has the best signal-to-noise ratio of the four test sets, the correlation is near 0.4, about twice the median subject’s. The demo puts it in plain terms: raw scans are noisy, and the model predicts “a canonical brain response.” That is a legitimate use, and the authors suggest piloting naturalistic studies with it before scanning anyone. But a single scan is a noisy sample of that average, which is what the number reflects. It does not show that the model can stand in for one person’s brain.

In silico, with stand-ins

The paper’s headline test replicates classic localizer experiments in silico. Predicted responses to flashed images recover the fusiform face area for faces, the parahippocampal place area for places, the extrastriate body area for bodies, and the visual word-form area for written characters. Language contrasts recover regions such as the temporo-parietal junction and Broca’s area. The results section presents the agreement as a qualitative match plus a significant spatial correlation across 360 cortical parcels, shown in figures rather than as a single coefficient in the text.

Rows of example stimuli (faces, places, body parts, written characters, tools), then brain surface maps of the activation TRIBE v2 predicted for each category, then maps measured in the Individual Brain Charting dataset. Predicted and measured maps look alike for the first four categories and diverge for tools.
Top row, example images; middle row, what TRIBE v2 predicts; bottom row, what the scanner measured. The paper's scatter plots beneath these maps print a correlation across 360 cortical parcels for each category: 0.64 for faces, 0.79 for places, 0.74 for body parts, 0.60 for characters, and 0.12 for tools (p=0.03). Image: cropped from Figure 4 of d'Ascoli et al., arXiv:2605.04326 (Meta FAIR, 2026), reproduced for commentary.

The results text names four of the five categories in that figure. The fifth, tools, is where the prediction mostly fails, and the paper doesn’t comment on it in the text I read. One weak category out of five is not damning, but it shows that “recovers” is doing different work for different stimuli.

The stimuli are stand-ins in places. The language sentences in that battery were in French. The authors translated them to English, synthesized speech with a text-to-speech model, then ran a speech-to-text tool over the audio to get word timings, and fed the result to TRIBE v2. For one task the original movie wasn’t available, so they contrasted speech and non-speech segments from a different dataset instead. None of that is improper, since the model needs timed audio and text. But “recovers a variety of results established by decades of empirical research” describes approximate reproductions of classic contrasts, not reruns.

An independent pair of researchers, Bladon and Bent, probed the model a different way. They ran feature visualization on TRIBE v2 and found a visible progression across the early visual areas, face-like features for the face area, and line patterns for the place area. They also found that images optimized for the face-area response drove it about four times as hard as a natural face photograph, “adversarial super-stimuli rather than canonical exemplars,” and they note that held-out prediction accuracy “tells us an encoder fits the data without telling us whether it has internalized the functional organization of the brain.” I read that as the boundary of what a localizer replication shows. Agreement with findings already in the literature is not evidence the model would get an unknown one right.

The scaling claim

The paper reports a log-linear rise in encoding accuracy with no plateau, and the demo says performance “increases log-linearly as we train it with more data.” In the paper the axis is hours of training data per participant, on the four-subject Algonauts dataset. That says more deep recordings help; it doesn’t say more subjects do. Deep recordings are the scarce kind. The Courtois NeuroMod project scanned six participants for what the paper calls “an unprecedently large recording volume of over 80 hours of fMRI per subject,” and TRIBE v2 trains on a four-subject subset of it curated for the Algonauts 2025 competition, 268.7 hours in all, about 67 hours a person.

What it’s being used for

Meta’s own authors (Benchetrit and colleagues) have already used TRIBE v2 to generate synthetic fMRI for training image decoders, reporting up to a 68 percent improvement in top-10 image retrieval and above-chance decoding from synthetic data alone in some settings. That’s a first-party result. As of today the Hugging Face weights show 42,415 downloads in the past 30 days and the GitHub repository 3,229 stars. Download counts measure curiosity more than use, as I’ve argued before. The license is CC BY-NC 4.0, so noncommercial, and the clinical uses the announcement gestures at remain prospective; the paper itself says integrating “clinical pathology remains a primary goal.”

Of the 720 subjects in the headline, 25 shaped the model’s weights. The other 695 are the reason to believe what it learned, and the paper says so. The announcement and the demo page never do.

References

  1. d’Ascoli, S., Rapin, J., Benchetrit, Y., Brooks, T., Begany, K., Raugel, J., Banville, H., & King, J.-R. A foundation model of vision, audition, and language for in-silico neuroscience. arXiv:2605.04326, May 5, 2026. Also on Meta AI Research, dated March 26, 2026.
  2. Meta AI. Introducing TRIBE v2: A Predictive Foundation Model Trained to Understand How the Human Brain Processes Complex Stimuli. Meta AI blog, March 26, 2026.
  3. Meta. TRIBE v2 demo page. Undated; read September 19, 2026.
  4. Meta FAIR. facebookresearch/tribev2, README and license (CC BY-NC 4.0).
  5. Bladon, S., & Bent, B. Feature Visualization Recovers Known Cortical Selectivity from TRIBE v2. arXiv:2605.13904, May 13, 2026.
  6. Benchetrit, Y., Careil, M., Dahan, S., Banville, H., d’Ascoli, S., & King, J.-R. Boosting Brain-to-Image Decoding with TRIBE v2 Data Augmentation. arXiv:2606.06345, June 4, 2026.
  7. Hugging Face, facebook/tribev2, and the GitHub repository, download and star counts read September 19, 2026.