Meta open‑sources TRIBE v2 brain simulator
Meta released TRIBE v2, a brain simulator model that predicts brain responses (including dopamine triggers) from scans — a research tool with implications for personalized UX and neuromodelling. The open release may accelerate cross‑discipline experimentation in adaptive interfaces and measuring physiological responses. (x.com)
The release bundle published March 26, 2026 includes the TRIBE v2 model checkpoint (best.ckpt ~709 MB), configuration files, and an interactive demo, all hosted on Meta’s GitHub repository and a Hugging Face model card. (github.com/facebookresearch/tribev2 ) (huggingface.co/facebook/tribev2 ) The training corpus spans roughly 720 human volunteers and on the order of 1,000 hours of fMRI recordings, per the project coverage and release notes. (awesomeagents.ai ) Spatial granularity was scaled from about 1,000 parcels in earlier work to roughly 70,000 voxels in TRIBE v2, a stated ~70x increase in resolution for cortical predictions. (digitalapplied.com ) Meta reports TRIBE v2 achieves zero-shot generalization to unseen subjects, languages and tasks and delivers roughly 2–3× predictive accuracy gains on movie and audiobook benchmarks versus prior encoding methods. (aidemos.atmeta.com/tribev2 ) The architecture reuses pretrained backbones—LLaMA 3.2 for text, V‑JEPA2 for video and Wav2Vec‑BERT for audio—fused by a transformer that maps multimodal latent features onto cortical meshes. (davidborish.com) (github.com/facebookresearch/tribev2 ) The codebase and weights were released under a Creative Commons CC‑BY‑NC‑4.0 license, which disallows commercial use without separate permission, and repository issues show early commercial licensing queries from outside groups. (huggingface.co/facebook/tribev2/blob/main/LICENSE ) (github.com/facebookresearch/tribev2/issues ) A three‑part executive briefing template circulating among engineering leads when presenting TRIBE v2: (1) concrete release facts — assets, host URLs and date (GitHub/Hugging Face, Mar 26, 2026), (2) validation metrics — ~720 subjects, ~1,000 hours, ~70,000 voxels, 2–3× gains, and (3) commercialization & compliance milestones — CC‑BY‑NC‑4.0 limits and a plan for licensing or independent validation. (github.com/facebookresearch/tribev2 ) (awesomeagents.ai )
