Meta Has Built an AI That Predicts How Your Brain Responds to Content — and Its Own Ad Business May Be the First Beneficiary
Meta's Fundamental AI Research team released TRIBE v2 on 26 March 2026, a foundation model that predicts how the human brain responds to video, audio and text. Trained on more than 1,100 hours of functional MRI recordings from 720 volunteers, the model has been fully open-sourced, with code, weights, and an interactive demo released under a non-commercial licence.
The headline figure is a 70-fold increase in spatial resolution over its predecessor, scaling from roughly 1,000 brain voxels to approximately 70,000. In zero-shot tests on an independent dataset, the model's group-averaged predictions outperformed individual human brain recordings. Meta's researchers described the result plainly: "TRIBE v2 reliably predicts high-resolution fMRI brain activity, enabling zero-shot predictions for new subjects, languages, and tasks, and consistently outperforms standard modelling approaches."
TRIBE v2 is not a mind-reading tool. It does not decode thoughts or reconstruct private experience. Given a stimulus, a film clip, a spoken sentence or an image, it predicts the pattern of neural activation that the stimulus is likely to produce across the cortex. That distinction matters enormously for how this technology is understood, and collapsing it serves no one.
What the Model Actually Does
The architecture draws on three frozen foundation models as feature extractors: LLaMA 3.2-3B for text, V-JEPA2 for video, and Wav2Vec-BERT for audio. A temporal transformer processes the time-evolving nature of naturalistic stimuli, while a subject-specific prediction block outputs predicted fMRI responses at whole-brain resolution. The model follows log-linear scaling laws, with predictive accuracy rising steadily as training data increases and no performance plateau is currently in sight.
Five functional networks emerged as properties of the model's internal representations without being explicitly programmed: primary auditory, language, motion, default mode, and visual. These are well-established categories in neuroscience. The fact that they emerged from a model trained purely on stimulus-response pairs, rather than on labelled anatomy, is the finding the research community is likely to find most substantive.
Meta said the goal of releasing the model publicly was to "accelerate neuroscience research that will unlock scientific and clinical breakthroughs for the greater good." Observers in the computational neuroscience community noted that the scale of the training dataset, combined with zero-shot generalisation across unseen subjects and languages, represents a meaningful departure from prior brain-encoding models, which were typically narrow in scope and required per-subject retraining.
How Disruptive Is This, Realistically?
For academic neuroscience, the near-term impact is concrete. Each fMRI session can cost several hundred dollars and requires specialist facilities. A model that approximates those results computationally changes the economics of brain research in a way that prior tools have not. Fine-tuning TRIBE v2 on just one hour of new subject data produces a two- to four-fold improvement over linear models trained from scratch. For under-resourced research teams, that is a material change in what research is feasible.
For brain-computer interface development, the zero-shot capability is the relevant variable. BCIs have long faced a scaling problem: models trained on one person's neural data perform poorly on another's. TRIBE v2 does not solve that problem in a single step, but it narrows the gap in a credible, well-documented way.
For enterprise technology, the disruption is real but indirect. The most immediately plausible commercial application is neuromarketing. Marketing teams could eventually test how content and advertising affect neural engagement without recruiting fMRI participants, thereby compressing a research process that currently incurs high costs and requires specialist access. Meta's own commercial interest here is not incidental: a company whose revenue depends on advertising effectiveness has built a model that simulates neural engagement with visual and auditory content.
Further applications in diagnostic neurology and treatment planning are credible, as the research pathway is now more plausible. They are not imminent. Clinical translation operates on timescales spanning years to decades and involves regulatory frameworks that TRIBE v2 does not sidestep.
The Open-Source Question
Meta's decision to release the model weights, codebase and demo is the mechanism by which the research becomes useful beyond Meta's own walls. It also accelerates independent replication, identifies limitations, and generates the critical engagement that improves the underlying science. The CC BY-NC licence restricts third-party revenue extraction but does not restrict Meta's internal use.
The Algonauts 2025 competition win that preceded this release is the material context. Brain encoding competitions provide a standardised benchmark for comparing different architectures on the same prediction task. First place in that environment is peer validation, not marketing copy. TRIBE v2 builds on a result that was independently evaluated before Meta decided to scale it into a foundation model. As the company's researchers noted in a release: "By sharing this work, we hope to help accelerate neuroscience research that will unlock scientific and clinical breakthroughs for the greater good."
For technology leaders, the near-term question is not whether to deploy TRIBE v2; it is a research release, not a product. The question is whether the category of AI capability it represents belongs in their threat and opportunity modelling. It does. The combination of multimodal AI and high-resolution neural prediction is no longer speculative. It exists, it scales, and it is openly available. The use cases that follow will be shaped by whoever moves fastest to understand what the model actually does, and by the regulatory environment that does, or does not, develop around it.