According to AI at Meta, TRIBE v2 predicts individual brain responses without any retraining and delivers a 2–3x improvement over prior methods across movies and audiobooks, with the model, codebase, paper, and demo now released for researchers. As reported by Meta’s AI team, the open resources (paper at go.meta.me/210503, model at go.meta.me/ea1cff, code at go.meta.me/873d02) enable labs to build generalizable encoding models, accelerate computational simulation for neurological disease diagnosis, and transfer brain insights into better AI architectures. According to Meta, this zero-shot generalization across unseen individuals lowers data collection costs, expands cross-subject benchmarking, and creates opportunities for healthcare imaging vendors, neurotech startups, and foundational model builders to integrate brain-aligned representations into product pipelines.

According to AI at Meta on X, Meta introduced TRIBE v2, a trimodal brain encoder foundation model trained to predict human brain responses to almost any sight or sound using 500+ hours of fMRI from 700+ participants (source: AI at Meta). According to Meta’s announcement page, the model builds on its Algonauts 2025 award-winning architecture to create a digital twin of neural activity and generalize in zero-shot to new subjects, languages, and tasks (source: go.meta.me/tribe2). As reported by AI at Meta, a public demo is available, signaling practical applications for neuroscience-informed AI, multimodal alignment, and personalized neuroadaptive interfaces in research and healthcare (source: AI at Meta).