Meta's Brain2Qwerty v2 Decodes Typing From MEG Brain Scans

Brain-computer interfaces have mostly been a story about implants. Stereotactic electroencephalography, electrocorticography, neural arrays, all of which can work but require surgery and are hard to scale. Meta's latest update on Brain2Qwerty is a small step in a different direction, decoding typed sentences from magnetoencephalography (MEG) recordings, with no implant involved.

The numbers Meta reports are what make it worth a second look. Across nine volunteer participants who each wore the MEG device for 10 hours while actively typing, the v2 system reached a 61% word accuracy rate on roughly 22,000 sentences. The best individual participant hit 78% word accuracy, with more than half of all sentences decoded with one word error or less. For context, Meta says other non-invasive methods sit around 8% word accuracy, so the jump is the headline even if absolute performance is still well short of fluent communication.

Two things in the method are worth flagging. The pipeline is end-to-end, learning from raw signals rather than older hand-crafted features, and it fine-tunes a large language model on neural data so it can lean on semantic context when the signal is noisy. Meta also claims decoding accuracy improves log-linearly with data volume, which is the part that hints at where the curve might go if more participant hours can be collected.

The honest caveat is hardware. MEG is not something you wear out of the lab, and the post does not address how that constraint gets resolved on the way to a usable assistive device. It also does not give independent replication, end-to-end latency, or how long per-user calibration takes. Take the specifics as reported, not settled.

Where this matters in practice is the open science wrapped around it. Meta is releasing the full training code for Brain2Qwerty v1 and v2, the Basque Center on Cognition, Brain, and Language is releasing the associated datasets, and there is a $5 million fund through Meta's Digital Brain Project to seed more open data. For academic labs working on non-invasive BCI, that combination is more useful than any single accuracy number.