Artificial intelligence continues to break into scientific and medical fields that until recently were the exclusive domain of complex and costly laboratory experiments. In a groundbreaking and impressive move, global technology company Meta has unveiled the second version of its multimodal artificial intelligence model, known as the “Tribe” model, which is specifically designed to predict how the human brain responds to various visual, auditory, and linguistic stimuli and triggers. This prominent development marks a fundamental turning point that could accelerate the pace of neuroscience research like never before, while simultaneously reducing the need for expensive brain scanning experiments that strain the budgets of research and medical institutions around the world and require significant time and effort.
An advanced digital twin for neural activity
The second version of this innovative model, developed by Meta’s Fundamental AI Research team, relies on a highly innovative and advanced architectural design to build what resembles a digital twin of the brain. This architecture proved its merit and overwhelming superiority when it claimed first place in the 2025 Algonauts brain decoding competition, outperforming more than 260 research teams participating from around the globe. To achieve this high level of accuracy, the model was trained on precise functional magnetic resonance imaging (fMRI) recordings derived from more than 700 volunteers. These volunteers were asked to watch a diverse set of action-packed, colorful videos, listen to complex radio programs and audio files, and read long texts in multiple languages, all while physically inside a scanner to monitor any changes in brain activity in real time.
This new release represents a massive leap and a major expansion compared to its predecessor, which was considered somewhat experimental. The first version relied on data derived from only four individuals and about 80 hours of audio and visual recordings, predicting activity across approximately 1,000 brain regions. The new and current version draws its formidable power from over 1,000 hours of fMRI data taken from 720 participants, and can predict neural responses across approximately 70,000 brain voxels, representing about a 70-fold increase in spatial accuracy and superior monitoring capabilities that capture the finest details of neural interaction inside the human skull.
The three-stage model workflow
This advanced model relies on a technical pipeline consisting of three carefully designed core stages to extract and analyze complex data coming from the senses. In the first stage, the model extracts fine-grained properties and features through separate, pre-trained encoders dedicated to processing video, audio, and written text individually. In the second stage, these three diverse inputs are integrated through a sophisticated fusion unit that relies primarily on transformer network technology. In the third and final stage, the resulting maps are rendered with extreme precision onto the virtual cortical surface, providing researchers with a comprehensive and clear picture of how the human brain interacts with those combined stimuli and how it processes various information in fractions of a second.
Computer simulation experiments and reducing medical costs
One of the central capabilities of this unique model is what researchers and scientists call “zero-shot prediction,” which is its superior ability to predict brain responses for entirely new individuals, languages the model has not previously recognized or processed, and even brand-new mental tasks without the cumbersome need to retrain the model from scratch every time. Meta reported accuracy improvements ranging from two to three times compared to previous traditional methods and approaches, which is particularly evident for individuals the model never encountered during the initial training phase, proving the model’s ability to generalize and deeply understand neural functions.
This technical breakthrough opens the door wide to what the company describes as “advanced computational experiments” conducted entirely inside high-performance supercomputers. When given a specific written sentence, the model can simply predict how the brain would be active as if inside an fMRI scanner, tracing the signal path through known language processing regions and reproducing patterns typically observed in real human studies. This innovative approach can allow researchers and scientists to test hypotheses and theories medically and computationally before committing to expensive and complex human trials; as it is well known, a single fMRI session can cost thousands of dollars, consume a lot of time, and require strict approvals from institutional review boards and medical ethics committees to ensure the safety of patients and volunteers.
Making the model available to researchers and the future of neuroscience
In a move aimed at accelerating open scientific collaboration and speeding up the wheel of development, Meta has made the model’s weights, codebase, research paper, and an interactive demo available under a Creative Commons license dedicated to non-commercial use, which is equivalent to a license allowing thousands of academics to begin using and experimenting with it immediately. In its official announcement of this exciting launch, the company explained that this strategic step represents a pivotal and fundamental milestone toward applying insights derived from brain study to build and develop better, more efficient artificial intelligence systems in the future, alongside accelerating medical discoveries in the diagnosis and treatment of intractable neurological diseases that have long puzzled doctors.
Despite these impressive successes and promising horizons, a number of specialized researchers pointed out an important point, which is that brain decoding models in their current state still explain only a modest portion of the vast variance in complex brain imaging data. Whether this new version will prove its practical usefulness and actual superiority outside pure academic benchmarks will depend on what independent scientists discover when they run it on their own real data in their laboratories. However, with the largest publicly available dataset in brain decoding, coupled with an open-source and accessible codebase, the global neuroscience community today possesses a powerful new infrastructure to test these possibilities and elevate our future understanding of the astonishing human mind and its limitless capabilities.
Frequently Asked Questions
Q: What is the second version of the Tribe model launched by Meta?
G: It is a multimodal artificial intelligence model specifically designed to predict how the human brain responds to visual, auditory, and linguistic stimuli without the need for direct human intervention.
Q: How does this model contribute to reducing neuroscience research costs?
G: The model can computationally simulate brain activity and test medical hypotheses with extreme precision, significantly reducing the need to conduct expensive and costly fMRI sessions on humans.
Q: Can the model predict brain activity for people whose data it was not trained on?
G: Yes, the model features “zero-shot prediction” capability, allowing it to read, analyze, and anticipate brain responses for new individuals, languages, and tasks it has not been previously trained on with the highest degree of accuracy and effectiveness.
