Meta has developed Brain2Qwerty v2, a noninvasive system that reads brain signals and converts them to text, achieving a 61% average word accuracy without any surgical implant — roughly eight times better than the previous best for implant-free approaches, according to research published June 29, 2026, in Nature Neuroscience.
How it works
Participants wear a magnetoencephalography (MEG) device — a scanner that detects magnetic fields produced by the brain’s electrical activity — while typing sentences. An end-to-end deep learning model processes the raw neural signals and a fine-tuned large language model uses sentence context to sharpen accuracy.
Nine volunteers each wore the device for 10 hours, producing around 22,000 training sentences in total. The best-performing participant reached 78% word accuracy, with more than half of their decoded sentences containing one error or fewer.
Why it matters
Earlier noninvasive methods topped out at roughly 8% word accuracy. Brain-computer interfaces that rely on surgical implants can do better, but the procedure limits how widely they can be adopted. Brain2Qwerty v2 narrows that performance gap without the risks of surgery, pointing toward practical communication aids for the millions of people with brain injuries that prevent normal speech or movement.
Meta reports that accuracy scales log-linearly with data volume, suggesting there is significant room for further improvement as more training data becomes available.
What’s next
Meta is publishing the full training code alongside the paper and has committed $5 million to its Digital Brain Project, aimed at expanding open neuroscience datasets. A partnership with the Basque Center on Cognition, Brain and Language (BCBL) will add to those datasets. The company is also releasing three companion tools — Tribev2 (a perception encoding model), NeuralSet, and NeuralBench — to support further academic research.