When Meta introduced the first version of TRIBE in 2024, the model could map brain activity across around 1,000 brain regions and had been trained on data from just four volunteers. It was a real advance but a limited one — more proof of concept than a research tool ready for scale.
On March 26, 2026, Meta's FAIR (Fundamental AI Research) team published TRIBE v2, and the numbers tell the story of the leap: from 4 volunteers to more than 700, from 1,000 brain regions to 70,000 cortical vertices, from a few hours of data to more than 1,000 hours of functional MRI recordings. And a resolution improvement of 70 times over comparable previous systems.
The model's full name is Trimodal Brain Encoder version 2 — TRI for the three modalities it processes simultaneously: vision, hearing and language. What it does with those three inputs was, until very recently, the exclusive territory of the best-equipped neuroscience laboratories in the world: predicting, with high accuracy, how a person's brain will react to what they are seeing, hearing or reading.
How TRIBE v2 works
The process has three stages that Meta described in detail on its research blog.
In the first stage, the model converts any input — a video, a podcast, a paragraph of text — into numerical representations that capture its visual, auditory and semantic content. It does this using state-of-the-art transformer architectures, the same building blocks behind the most advanced language models on the market.
In the second stage, the system analyses those representations to identify patterns in how the human brain processes each type of information. This is the core of the model: learning the statistical relationships between what a person experiences and how their brain responds at a regional level.
In the third stage, TRIBE v2 predicts which brain regions will activate in response to a given stimulus — and with what intensity. The result is a map of brain activity generated entirely by software, without putting anyone inside an MRI scanner.
The resolution of that map — 20,484 cortical vertices — is substantially higher than any equivalent previous system. And the accuracy is high enough that in many cases the model's predictions are cleaner than actual real-world scans, which typically contain noise generated by heartbeats, body movement and other physiological interference.
The data that makes it possible
TRIBE v2 was trained on a combination of neuroimaging datasets at a scale that has no precedent for this type of model. The training set included 451.6 hours of fMRI data from 25 participants, collected during naturalistic studies — meaning while volunteers watched films, listened to podcasts and observed silent videos under conditions similar to everyday life.
For evaluation, Meta used a much broader set: more than 1,117 hours of data from close to 720 participants across four independent studies.
What the research team found when analyzing the model's behavior is itself an important result: TRIBE v2's accuracy improves in a log-linear fashion as the volume of training data increases, with no signs of that improvement plateauing. In other words: the more MRI data becomes available, the more accurate the model can become. And neuroimaging repositories are growing.
Why it outperforms real scans in some cases
One of the most counterintuitive findings from TRIBE v2 is that its synthetic predictions — generated entirely by software — outperform many real fMRI scans in accuracy when it comes to capturing population-level brain activity.
The technical explanation is that individual fMRI data contains a significant amount of physiological noise: variations in blood flow caused by heartbeats, micro-movements of the head during scanning, fluctuations in breathing. TRIBE v2, having learned from thousands of hours of data from hundreds of subjects, captures the activation patterns that are consistent across people and discards individual noise.
The result is a model that can generate a brain activity map for a new stimulus that matches the average population response better than the scan of any particular individual.
What it can do in practice
The applications Meta describes in its research blog range from the immediate to the ambitious.
In the short term, the most direct application is accelerating neuroscience research without needing to recruit participants for every experiment. A typical neuroimaging study today can take months to design, recruit volunteers, collect data and analyze results. With TRIBE v2, researchers can generate brain activity predictions for new stimuli in seconds, use those predictions to design more focused experiments and only resort to real scans to confirm the most relevant findings.
In the medium term, the model has potential to contribute to the treatment of neurological disorders. If you can accurately predict how the brain responds to different types of stimuli, that knowledge can help design better therapies for conditions like autism, dyslexia, Alzheimer's disease and various forms of acquired brain injury.
In the long term, Meta is explicit about its most ambitious goal: using TRIBE v2 as a step toward developing superintelligent AI that processes information more like the human brain does. If AI models learn to replicate the brain's processing mechanisms — not just its outputs, but its internal processes — they could become substantially better at tasks where they still fall short today: contextual understanding, causal reasoning, interpretation of ambiguous signals.
Open source and the privacy questions
Meta released TRIBE v2 as open source, along with the model's complete code and an interactive demo so researchers around the world can use it, audit it and build on top of it.
The decision to make it public opens the door to significant acceleration of computational neuroscience research. But it also raises questions that privacy experts have already begun raising: brain data is considered highly sensitive information. Neural activity in response to different stimuli can reveal preferences, emotional states, cognitive biases and other aspects of subjective experience that a person may not want to share.
Meta states that the data used to train TRIBE v2 was collected with informed consent from participants and under approved research protocols. But as these types of models scale and become integrated into real-world applications, existing regulatory frameworks — designed to protect conventional medical data — may prove insufficient.
What Meta doesn't say in its announcement
TRIBE v2 is a genuine scientific advance. But it's hard to ignore the context in which it was developed: Meta is the world's largest social media and digital advertising company.
An AI capable of predicting with high accuracy how the human brain reacts to different types of content — videos, images, text, audio — is, in the context of an advertising platform, something more than a scientific research tool. It's potentially the most sophisticated content optimization system that has ever existed.
Meta doesn't make that connection in its announcement. But AI ethics and privacy experts are making it this week, and will likely continue doing so as the model matures and its applications expand.
Comments
💬 Log in to comment💬 Join the conversation and log in to comment.