Meta's Brain2Qwerty v2 Reads Sentences Without Surgery
Brain2Qwerty v2 demonstrates that non-invasive brain-computer interfaces can approach the accuracy of surgical implants, moving toward practical communication aids without medical risk.
Reporting from 1 sources: GIGAZINE.
Meta and the Basque Center on Cognition, Brain and Language have developed Brain2Qwerty v2, an AI model that decodes text from non-invasive magnetoencephalography recordings. The new version reads at the word and sentence level, achieving up to 78% word accuracy. Meta also released training code and data.
Meta and the Basque Center on Cognition, Brain and Language (BCBL) have released Brain2Qwerty v2, an AI model that decodes typed sentences from non-invasive magnetoencephalography (MEG) recordings. Unlike earlier systems that required surgically implanted electrodes, Brain2Qwerty v2 reads brain activity from outside the skull. The model was trained on roughly 22,000 sentences from nine volunteers who each typed for ten hours while their brain magnetic fields were measured. Brain2Qwerty v2 reads at the word and sentence level, a step up from the character-by-character decoding of the February 2025 v1. Meta reports a character success rate of 69% and word accuracy up to 78%. For the best-performing subject, more than half of all sentences were decoded with an error of one word or less. Accuracy improved with more training data. Meta has also published the training code for both v1 and v2, along with the v1 training data, on GitHub and Hugging Face. The v1 paper, announced in February 2025, passed peer review and was published in Nature Neuroscience on June 29, 2026.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.