Skip to content
News Research

Meta's Brain2Qwerty v2 decodes typed sentences from brain signals at 61% word accuracy without surgery

· by Pondero Newsdesk

The short version

Meta FAIR published Brain2Qwerty v2 on June 29, using a non-invasive MEG helmet and end-to-end deep learning to decode what nine participants typed directly from brain activity, reaching 61% average word accuracy and 78% for the best participant.

Meta's Brain2Qwerty v2 decodes typed sentences from brain signals at 61% word accuracy without surgery

Nine people sat in a magnetically shielded room wearing a clinical MEG helmet while they typed sentences on a keyboard. Meta's FAIR lab published the result on June 29: a non-invasive brain-computer interface that decoded what they typed from their brain activity alone, hitting 61% average word accuracy with no surgery required.

What

Brain2Qwerty v2 is a brain-to-text pipeline trained on roughly 22,000 typed sentences collected from nine volunteers, each recorded for about 10 hours per the MarkTechPost coverage of the paper. The system reads magnetoencephalography (MEG) signals, which capture the magnetic fields produced by neuron firing without touching the brain. An end-to-end deep learning stack then turns those signals into text: a convolutional encoder processes the raw MEG stream, a transformer models longer-range structure across the recording, and a character-level language model constrains the output to plausible words and sentences.

Average word accuracy across participants reached 61%, or a 39% word error rate. The best participant reached 78%, and more than half of that participant's sentences had one word error or fewer. Prior non-invasive BCI systems had reached about 8% word accuracy per MarkTechPost's summary of the paper. Accuracy also scaled log-linearly with training data volume, meaning more recording hours translated to predictable accuracy gains in the tested range.

Meta released the full training code for both v1 and v2 under a CC BY-NC 4.0 license at github.com/facebookresearch/brain2qwerty. The v2 dataset remains under embargo until the paper clears peer review; the v1 study was published in Nature Neuroscience in 2025.

Why it matters

The 61% figure is the operative number for anyone watching the BCI field. Prior non-invasive systems hovered around 8% word accuracy. Surgical implants such as stereotactic EEG or electrocorticography deliver higher accuracy but require neurosurgery, a meaningful barrier that limits how many patients can access the technology. Brain2Qwerty v2 does not close the gap with implants entirely, but it narrows it considerably using nothing more invasive than a helmet.

The log-linear scaling result carries a practical implication. It suggests that accuracy is, to a significant degree, a data problem rather than an architecture problem. A research team with more recording hours could, per the paper's findings, expect continued accuracy improvement without redesigning the model. That is a different kind of claim than most BCI papers make.

The primary application target is communication assistance for people who cannot speak or move due to brain injuries, ALS, or other conditions. Millions of people globally fit that description. A system that requires only external MEG equipment (rather than an implanted array) is a different accessibility proposition than what surgical approaches offer, even though the MEG hardware is clinical-grade and currently requires a shielded room.

Context and reactions

Brain2Qwerty v1 shipped in February 2025 and worked at the character level. That version used both MEG and EEG recordings across 35 participants and demonstrated that MEG produced roughly twice the decoding accuracy of EEG signals. Version 2 narrowed to MEG, dropped to nine participants, and moved to sentence-level word accuracy as the primary metric. The two versions measure different things, so the character accuracy of v1 and the word accuracy of v2 do not map directly onto each other.

Data collection for v2 was done in collaboration with Spain's Basque Center on Cognition, Brain and Language (BCBL). The dataset belongs to that research center, which is why it remains under embargo pending journal acceptance. FAIR is releasing the model and training code separately.

For the broader AI community, the technical approach is notable for a different reason than the BCI application: it replaces hand-crafted neural event detection with a fully end-to-end deep learning pipeline. That design pattern transfers to other biosignal decoding tasks where hand-engineered feature extraction has historically been the bottleneck.

What to watch next

The immediate milestones are peer review of the v2 paper and the eventual release of the v2 dataset under the BCBL embargo timeline. The scaling result will attract follow-on work: if log-linear accuracy gains hold as data volume grows, groups with larger MEG recording programs should be able to push word accuracy further using the released training code. Watch for replication attempts and whether the pattern holds at higher data volumes than the current study tested.

Sources