Meta unveils TRIBE v2 for zero-shot prediction of high-resolution fMRI activity

Meta described TRIBE v2 on March 26, 2026 as a predictive foundation model that acts as a digital twin of human neural activity. The company says the system can predict how the brain responds to sight, sound, and language, and that it delivers a 70x resolution increase over similar approaches. Rather than framing it as another narrow benchmark model, Meta is positioning TRIBE v2 as research infrastructure for neuroscience and clinically relevant hypothesis testing.

The scale claim is central to the release. Meta says TRIBE v2 was trained on data from more than 700 healthy volunteers who were exposed to images, podcasts, videos, and text. On that basis, the company says the model can make zero-shot predictions for new subjects, languages, and tasks while consistently outperforming standard modeling approaches. If those results hold up in wider use, the model could reduce how often researchers need to collect fresh human-subject data during the exploratory phase of a study.

Why this matters

Brain-response modeling is usually constrained by expensive data collection, limited resolution, and weak transfer across individuals. Meta is explicitly arguing that TRIBE v2 can compress that cycle: researchers could test ideas in a computational model first, then reserve scarce lab time for the most promising questions. That matters for basic neuroscience, but it also matters for work on neurological disorders, where faster iteration can help refine research directions earlier.

• Meta says TRIBE v2 predicts high-resolution fMRI activity rather than coarse behavioral outputs.
• The company highlights zero-shot generalization to unseen subjects, languages, and tasks.
• Meta is releasing model weights, code, a paper, and an interactive demo for the research community.

There is still a clear boundary between prediction and explanation. A model that approximates brain activity does not automatically reveal causal mechanisms, and Meta notes that the goal is to help researchers test hypotheses faster, not replace experiments outright. Even so, the release stands out because it combines a large multimodal brain dataset, a deployable predictive model, and public research artifacts in one package. For AI and IT readers, the broader takeaway is that foundation-model techniques are continuing to move into domains where measurement is expensive and expert review is the main bottleneck.