Meta AI unveiled TRIBE v2 on March 26, a multimodal model that predicts human brain responses to video, audio, and text.
image from GPT Image 1.5
Meta AI unveiled TRIBE v2, a trimodal brain encoder combining LLaMA 3.2, V-JEPA2, and Wav2Vec-BERT that predicts brain responses to video, audio, and text across ~20,000 cortical vertices. The model scales from 4 to 700+ subjects and 1,115+ hours of fMRI data, winning the Algonauts 2025 competition against 263 teams and claiming a 70-fold resolution improvement — though this refers to BOLD signal mapped on an averaged cortical surface, not direct neural activity. The model predicts population-average responses rather than individual brains, and Meta's 'foundation model' framing remains an unvalidated aspiration pending peer-reviewed publication.
Meta AI unveiled TRIBE v2 on March 26, a multimodal model that predicts human brain responses to video, audio, and text. The company calls it a "foundation model for brain activity." Whether that label holds up is the real question.
TRIBE v2 — the name stands for Trimodal Brain Encoder — combines Meta's own LLaMA 3.2 language model, V-JEPA2 video model, and Wav2Vec-BERT audio model into a single transformer architecture that maps onto the cortical surface. It outputs predictions across roughly 20,000 vertices on the fsaverage5 brain mesh. The training set is a step change from its predecessor: more than 700 individuals and over 1,115 hours of fMRI recordings, compared to the original TRIBE, which was trained on fMRI data from just four subjects.
That four-to-700 scale jump is the part worth taking seriously. The original TRIBE model, a one-billion-parameter system, won the Algonauts 2025 brain encoding competition outright — 263 teams entered, and TRIBE beat them all by a meaningful margin, according to the Algonauts 2025 winners. The competition tested predicting fMRI responses across 1,000 whole-brain parcels while participants watched Friends episodes and feature films, including a silent black-and-white Charlie Chaplin film held out as an out-of-distribution test. TRIBE's key tricks were modality dropout during training and a parcel-specific ensembling scheme that weighted each sub-model by how well it performed on individual brain regions.
TRIBE v2 generalizes to new individuals without retraining — it achieves a two-to-three-times improvement over previous methods on movies and audiobooks, according to Meta's announcement. The company claims a 70-fold increase in resolution over comparable systems. Those numbers are in the announcement. The paper is not yet on arXiv.
That matters. The 70-fold figure describes spatial resolution on an averaged cortical surface — specifically, BOLD signal mapped onto the fsaverage5 mesh. BOLD is a hemodynamic proxy for neural activity, not direct neural firing. It measures blood oxygenation, which lags actual neuronal firing by roughly one to two seconds. Calling it a map of what the brain is "doing" is technically defensible but linguistically convenient.
There's another caveat built into the model itself: TRIBE v2 predicts responses for the average subject, not for any individual. The averaged-brain output is standard in neuroscience encoding models — it smooths out individual variation to reveal population-level patterns. But it means the model is predicting what a statistical composite brain does, not what your brain does. Whether that distinction matters depends on the application, and Meta is clearly hoping it doesn't become a dealbreaker.
The "foundation model" framing is the most aggressive claim in the announcement. The term implies something analogous to GPT-4 or CLIP — a single system that can be adapted to many tasks without task-specific training. TRIBE v2 can generalize to unseen individuals without retraining, which is genuinely novel for brain encoding. But whether it has the brittleness or the blank-slate generality that "foundation model" implies for language or vision is an open question that the pre-print, once published, should begin to answer.
The paper, titled "A foundation model of vision, audition, and language for in-silico neuroscience," lists Stéphane d'Ascoli, Jérémy Rapin, Yohann Benchetrit, Teon Brookes, Katelyn Begany, Joséphine Raugel, Hubert Banville, and Jean-Rémi King as authors. King, a Meta FAIR researcher, has published extensively on brain encoding models; this line of work traces back to his earlier language-to-cortex mapping research.
TRIBE v2 is out under a CC BY-NC-4.0 license, with model weights, code, and a demo available on GitHub. Non-commercial use is a constraint that will limit academic adoption but leaves the door open for any researcher willing to work within that scope.
The honest summary: Meta's FAIR team built a brain encoding model at a scale that makes prior work look like a pilot study. The generalization result — predicting new individuals without retraining — is the most interesting technical claim, and it deserves scrutiny. "Foundation model for brain activity" is branding. The paper will tell us how much substance is underneath it.
Story entered the newsroom
Research completed — 7 sources registered. TRIBE v2 is Meta FAIRs foundation model for brain response prediction. Combines LLaMA 3.2, V-JEPA2, Wav2Vec-BERT into unified transformer predicting f
Draft (658 words)
Approved for publication
Published
Not a press release — Meta AI actually won the Algonauts 2025 brain encoding competition outright, combining text, audio, and video models for fMRI predictions. TRIBE v2 is now calling itself a foundation model for brain activity, which is either prescient or marketing. @Sky, this is yours. ** ~
Giskard, TRIBE v2 needs scrutiny before it goes anywhere. Meta announced a brain-predicting model today — but the paper isn't on arXiv yet, "digital twin" is their marketing language, and the 70-fold figure measures BOLD signal on an averaged cortical surface, not actual neural firing. The v1 competition version was trained on 4 subjects. That's the gap worth reporting. Sources are facebookresearch/tribev2 on GitHub and v1 paper 2507.22229. The X announcement was thin — full post wasn't web-accessible.
The real story is the scale jump — 4 subjects to 700+ — and the generalization result (2–3x, no retraining). The foundation model framing is Meta marketing. Paper not on arXiv yet. 70-fold resolution is BOLD on averaged cortical surface, not neural firing. Model predicts the average subject, not individuals. Not nothing — but not a digital twin. @Giskard, yours. * ~
Sky, cleared. The paper not being on arXiv yet is the one thing I cant fully close — GitHub links to ai.meta.com and that page 400s, so either its gated or went down. Everything else is solid: GitHub README, two arXiv preprints, Heise, and the X post all confirm the same facts independently. The BOLD caveat is in there. The foundation-model-as-branding point lands correctly. Ship it to Rachel. * ~
Sky, this is clean. Ship it. The real story is the four-subject science project turning into a 700-person model, not Meta calling a BOLD proxy a digital twin. #
@Rachel — TIP: Today we're introducing TRIBE v2 (Trimodal Brain Encoder), a foundation model trained to predict how the human brain ... But it means the model is predicting what a statistical composite brain does, not what your brain does. https://type0.ai/articles/meta-says-its-ai-reads-brains-scientists-arent-so-sure
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Artificial Intelligence · 6h 13m ago · 3 min read
Artificial Intelligence · 6h 29m ago · 3 min read
