Why now

Meta’s FAIR lab just released Brain2Qwerty v2, a system that turns raw brain activity into typed sentences with no implant, hitting 61% word accuracy on average and 78% for its best participant. That’s up from the roughly 8% other non-invasive methods manage, and it starts to close in on what surgical brain interfaces achieve. Meta is also open-sourcing the training code for v1 and v2, and its partner BCBL is releasing the v1 dataset.

What happened

The model learned from about 22,000 sentences typed by nine volunteers, each recorded for 10 hours inside a magnetoencephalography (MEG) scanner, the machine that reads the faint magnetic fields your neurons give off. Instead of hand-built pipelines that flag neural events one at a time, v2 runs end-to-end deep learning straight from the raw signal, then fine-tunes a language model on top so semantic context can clean up noisy guesses. Meta says AI agents helped search the optimization space, with engineers choosing the final settings. For the best participant, more than half of all sentences came out with one word error or fewer. The team also reports that accuracy climbs log-linearly with data, so more recording hours keep moving the number up.

Why it matters

This is aimed at a specific group: the millions of people with brain lesions who’ve lost the ability to speak or move, where the only working option today is neurosurgery to implant electrodes. A non-invasive route that reaches surgical-level accuracy would change who can actually get that kind of communication back, and the open code plus a $5 million open-data fund is built to pull academic labs into the problem.

But look at who generated the data. Every one of those nine participants was a healthy volunteer who could read the sentence on a screen and physically type it, and Meta’s own v1 analysis traced most of the decoding to motor processes, the brain’s finger-movement signals. The detail getting lost in the coverage is that the patients this is being built for, the ones who can’t move or type, are precisely the people who don’t produce that signal. So the 61% is measured on the easiest possible subject performing a task the intended user can’t do. The gap worth watching runs between a volunteer typing in a lab and a patient who has no keystrokes left to read, not between MEG and surgery.