Meta launches Brain2Qwerty v2 to turn brain signals into text

Meta introduced Brain2Qwerty v2, an AI system that decodes brain activity recorded with magnetoencephalography (MEG) into plain text. The company announced the system on Monday and published the model’s training code.

Researchers trained the model on roughly 22,000 sentences collected from nine volunteers. Each participant wore an MEG helmet for about 10 hours while actively typing. The MEG device records magnetic fields produced by neural activity without surgery. Meta fed the raw recordings into an end-to-end deep learning model that reconstructs the sentences participants intended to type and fine-tuned large language models on neural data so the system can use semantic context to interpret noisy signals. In the paper Meta wrote, “We trained Brain2Qwerty v2 on approximately 22,000 sentences from nine volunteer participants, each recorded for 10 hours wearing a magnetoencephalography (MEG) device while actively typing.”

Meta reported an average word accuracy of 61% for Brain2Qwerty v2, compared with about 8% for earlier non-invasive methods. The company noted decoding accuracy increased with the amount of training data, indicating larger datasets could raise performance.

The work is described in a paper published in Nature Neuroscience. Meta released the training code for both Brain2Qwerty v1 and v2 and is participating in the Digital Brain Project, which includes a $5 million fund to support open neuroscience datasets. A research partner will release the v1 dataset used in earlier work.

Meta said the system approaches accuracy levels previously achieved mainly by implanted electrode interfaces. The company highlighted that non-invasive recording avoids surgical risks and the long-term maintenance challenges associated with implants.

Automated AI agents explored possible configurations of the decoding pipeline before engineers selected the final training setup. Meta did not provide a timetable for clinical deployment; the paper and public release of code and datasets are intended to support replication and further research by the scientific community.