Meta's Brain2Qwerty v2, its strongest non invasive brain to text decoder, reaches best session 78% word accuracy — averaging around 61% across all sessions — on a magnetoencephalography (MEG) scanner that costs several million dollars and lives in shielded rooms.
Locked-in syndrome, anarthria, and advanced ALS leave people fully conscious but unable to speak or move. The assistive devices that try to give them a voice are typically invasive (requiring brain surgery), slow (eye-tracking at a few words per minute), or both. On Tuesday, Meta open-sourced Brain2Qwerty v2, its strongest non-invasive brain-to-text decoder yet, and the gap between those two facts is the actual story.
The release put the model code on GitHub under facebookresearch/brain2qwerty and accompanied it with a research paper on arXiv describing the architecture and training. The headline number, 78% word accuracy in the best session, is the highest Meta has reported for the system. None of that changes what the decoder actually runs on.
Brain2Qwerty v2 reads the magnetic fields produced when groups of neurons fire together, a technique called magnetoencephalography, or MEG. Unlike electroencephalography (EEG), which picks up electrical noise through the scalp, MEG detects the much weaker magnetic signature of synchronized neural activity, giving it higher spatial resolution without penetrating the skull. The trade-off is that MEG scanners are rare, expensive, and inflexible. A single system can list for several million dollars, requires a magnetically shielded room, and forces the user to sit still inside a helmet-sized sensor array for the duration of a session.
The training gap
Nine healthy volunteers between the ages of 25 and 56 typed more than 2,500 sentences each across ten sessions at the Basque Center on Cognition, Brain, and Language in San Sebastián, Spain. The model learned to map the resulting brain activity to the sentence being typed. In its best configuration, it reached 78% word accuracy; the average across all sessions was closer to 61%. Both numbers are short of the bar clinical communication aids need to clear.
The intended users, people with anarthria, locked-in syndrome, and ALS, were not the training population. Whether a decoder trained on intact motor cortex activity generalizes to the atrophied or reorganized neural patterns of a paralyzed patient is an open question the companion paper flags but does not answer. Reporting on prior non-invasive decoders has noted that accuracy can degrade substantially when systems trained on healthy volunteers are moved into clinical populations.
The hardware problem
The deeper issue is what the open-source release does not touch. MEG scanners exist in only a handful of research hospitals worldwide. A patient who could benefit from a brain-to-text decoder would have to travel to one of those centers, be seated inside the sensor array, and remain still for the duration of a session. That is not an accessibility device. It is a research instrument, and Tuesday's release does not change that.
In a blog post announcing the work, Meta framed it as part of an "open-source AI for science" push to "advance neuroscience to identify, diagnose, and treat neurological disorders faster than in silos." The framing is fair. Open code lets outside teams test the model, train it on new data, and benchmark it against alternatives. It does not, by itself, get the hardware into a clinic or a home.
What to watch
Two things will determine whether Brain2Qwerty v2 matters as more than a research artifact. First, whether the accuracy ceiling moves substantially in the next round of training, and whether any improvement comes from more data, larger models, or both. Second, whether anyone can move the acquisition side of the pipeline, the sensors and the shielded rooms, onto hardware that exists outside a research hospital.
The software half of the problem is now, by Meta's own decision, a public good. The hardware half is still a physics problem with no obvious software solution. Until that changes, the most honest reading of this release is that it narrows the unknown in a hard problem. It does not solve the problem.