The Surgery-Free Path to Brain-to-Text Decoding Gets a Major Upgrade

The Surgery-Free Path to Brain-to-Text Decoding Gets a Major Upgrade

For people who have lost the ability to speak or move after a brain injury, the most reliable way to restore communication has involved neurosurgery. Electrodes placed on or inside the brain can feed clean signals to a decoder, and recent implant systems have produced fluent text and speech. The problem is that surgery carries real risk and does not scale to the millions of people who might benefit. Meta's research group has spent two years chasing an alternative that works from outside the skull, and its latest result suggests the gap between the two approaches is narrowing faster than most expected.

The system is called Brain2Qwerty v2, released on June 29, 2026. It decodes the sentences a person types using only magnetoencephalography, a method that measures the faint magnetic fields produced by neural activity through a helmet that surrounds the head without touching the brain. No implant, no incision. The model reconstructs what the person typed by reading their brain activity as they type it.

What the numbers show

Meta trained the model on roughly 22,000 sentences collected from nine volunteers. Each participant wore the MEG device for about 10 hours while actively typing, which produced the large, dense recordings the model needed to learn from.

Across all participants, Brain2Qwerty v2 reached an average word accuracy of 61%, which the research paper frames as a 39% word error rate. For the single best participant, accuracy climbed to 78%, and more than half of that person's sentences were decoded with one word error or fewer. To anchor those figures, Meta compares them against an earlier non-invasive benchmark that achieved 8% word accuracy on a related task.

The first version of Brain2Qwerty, published in early 2025, reported its results as character error rates, reaching 32% on average with MEG and 19% for its best participant, while electroencephalography, the cheaper scalp-electrode method, trailed badly at a 67% character error rate. The shift to word-level accuracy in v2 reflects a system now good enough to be measured by whole words rather than individual keystrokes.

One finding in the paper carries more weight than the headline accuracy. Meta reports that decoding accuracy improves log-linearly with the volume of training data. In plain terms, every time the researchers fed the system substantially more recorded brain activity, accuracy rose in a predictable way rather than flattening out. That pattern is familiar from large language models, where more data has reliably bought more capability, and it suggests the remaining distance to surgical-implant performance might be closed partly by recording more data rather than inventing entirely new methods.

How the system works

Brain2Qwerty processes brain signals through a pipeline with three stages. A convolutional module first reads the raw MEG recording in short windows and extracts the temporal and spatial features buried in the noise. A transformer module then takes those windowed features and works out the likely sequence of characters, producing candidates and their probabilities. Finally, a language model cleans up the output, using its grasp of how words and sentences fit together to correct errors the neural decoder makes on its own.

The third stage is where v2 separates itself from the first version. Rather than decoding one character at a time, the newer system reasons across characters, words, and entire sentences at once. Meta fine-tuned large language models directly on neural data so the system could use semantic context, the ordinary statistical structure of language, to recover a coherent sentence from a recording that is genuinely noisy. When the neural signal alone is ambiguous between two readings, the language model's sense of what sentence a person was likely to type helps settle the question.

Meta also describes an unusual element of its development process. The team deployed AI agents to explore optimizations to the decoding pipeline, iterating on the code automatically, with engineers selecting the final training configurations by hand. The research used AI not only inside the model but as a tool for building the model.

Where Brain2Qwerty fits among other brain interfaces

Brain2Qwerty is one entry in a much larger field of devices trying to read intention from the brain, and the differences between them mostly come down to a trade between signal quality and physical risk. The systems closest to clinical use today sit at the invasive end of that spectrum, planted directly in or on neural tissue.

Neuralink's approach implants a chip with hundreds to thousands of hair-thin electrode threads directly into the motor cortex during open brain surgery. Patients in the company's trial have used the resulting signal to control a computer cursor, with early participants reaching roughly 9.5 bits per second of information transfer, comparable to able-bodied mouse use, and the company has reported a typing-focused trial in which a participant mapped imagined finger movements to a keyboard layout and reached 40 words per minute. A second-generation device with a denser electrode array has reportedly pushed one participant's typing speed past 100 words per minute, faster than most people type by hand. The signal is excellent because the electrodes sit against individual patches of cortex, but getting there requires a craniotomy and carries the surgical risks that come with any procedure inside the skull.

A related approach skips speech-as-typing entirely and decodes attempted speech itself. Researchers at Stanford recorded from intracortical microelectrode arrays in a person with ALS and decoded attempted speech at 62 words per minute with error rates under 10% on small vocabularies, a result that began to approach the pace of natural conversation. A separate UCSF and UC Berkeley team took a different angle, implanting an electrode array over the brain's speech and motor regions in a woman who had not spoken in 18 years following a stroke, and used it to drive a synthesized voice along with a digital avatar's facial movements in something close to real time. Both systems depend on electrodes placed directly on the brain's surface and require the same surgical exposure as Neuralink's approach, with the tradeoff that they aim to restore actual speech rather than typed text.

Synchron has staked out a middle position. Its Stentrode device is a small mesh stent, not unlike the ones used in cardiac procedures, threaded up through the jugular vein and guided into a blood vessel that runs along the motor cortex, where it reads neural signals through the vessel wall rather than from tissue contact. The implant procedure avoids opening the skull, takes about 20 minutes, and has shown a clean safety record through more than a year of follow-up in early trials. The cost of avoiding surgery is signal quality. Reading through a blood vessel wall picks up a coarser, lower-resolution signal than direct cortical contact, and Synchron's patients have generally achieved slower, less precise cursor control than Neuralink's. The company is using that safety advantage to pursue a broader patient population and has raised substantial funding to move toward commercialization.

Brain2Qwerty sits at the opposite end of the spectrum from all three of those systems. It requires no implant of any kind, invasive or endovascular. The signal comes entirely from outside the head, through a MEG helmet that detects the magnetic fields neurons produce without touching skin or skull. That is also where its limitations come from. MEG signal is far noisier than anything recorded from inside the brain, which is why Brain2Qwerty needs hours of training data per participant and still trails the best intracortical systems on raw accuracy and speed. It also depends on equipment that, unlike a surgical implant, cannot go home with the patient. A MEG scanner is a room-sized, magnetically shielded installation, not a wearable device, which puts a hard ceiling on where the technology can currently be used regardless of how good the decoding gets.

That distinction matters more than it might first appear. The other major category of non-invasive brain reading, scalp EEG, is genuinely portable but produces a much weaker signal, which is part of why Brain2Qwerty v1 reported a 67% character error rate on EEG against 32% on MEG. Consumer EEG headsets used in academic and hobbyist brain-computer interface work tend to perform worse still on anything as complex as full sentence decoding.

MEG occupies an unusual spot: non-invasive like EEG, but with signal quality that approaches what implants can offer, at the cost of being stuck in a lab. Meta's bet is that the accuracy gains from better AI decoding, not better hardware, can be banked now while sensor engineering catches up separately. The paper's finding that performance scales log-linearly with data volume is the evidence for that bet. It implies the AI side of the problem keeps paying off as more recordings come in, independent of whether the MEG hardware itself ever gets smaller.

The hardware problem has not gone away

The accuracy gains are real, but the practical barriers that limited the first version remain. MEG machines are large, expensive, and sensitive to interference. They are not wearable in any everyday sense, and they exist mostly in research hospitals and labs rather than homes or clinics. A system that works inside a magnetically shielded room with a multi-ton scanner is a long way from a device a patient could use at the kitchen table, and it is a long way, too, from an implant a Neuralink or Synchron patient simply carries with them after a single procedure.

The study population also matters. The nine volunteers were healthy participants typing memorized sentences, not people with the motor or speech impairments the technology is ultimately meant to help. Decoding the brain activity of someone who is physically typing is a different problem from decoding the intentions of someone who cannot move at all, and the paper's results do not yet speak directly to that harder case. The v1 work made the same caveat, and it still holds. The implant-based systems described above, by contrast, have already been tested in their target population, people with paralysis or ALS, which is part of why they currently lead on real-world deployment even though Brain2Qwerty is closing the accuracy gap on paper.

There is also the question of what "real-time" means here. Meta describes v2 as capable of real-time sentence decoding, an advance over v1, which had to process complete sentences after the fact. Independent coverage notes that exact latency figures and the viability of cheaper EEG-based recording were not detailed in the release, leaving some of the distance between lab demonstration and clinical use unspecified.

Why releasing the code matters

Meta published the full training code for both Brain2Qwerty v1 and v2, and its research partner, the Basque Center on Cognition, Brain, and Language, released the v1 dataset of Spanish-language recordings. That decision lets other labs replicate the pipeline and build on it without reconstructing it from scratch, which tends to accelerate progress in a field where high-quality brain data is scarce and expensive to collect. Meta frames the release as part of a broader effort to build open foundational models of the brain, alongside related projects for modeling perception and evaluating brain-decoding systems, and a $5 million fund aimed at stimulating open datasets.

The openness cuts against the instinct to keep neural-decoding capability proprietary, and it invites the obvious privacy concern that comes with any technology that reads brain activity. Public reaction to the release has reflected both responses, with some observers focused on the assistive potential and others wary of a large advertising company developing tools that interpret neural signals.

Where this leaves the field

Set against implanted systems, Brain2Qwerty v2 is not yet the fastest or most accurate way to get text out of a brain. Neuralink's typing trial and the Stanford speech neuroprosthesis both move faster, and Synchron already has patients living with an implanted device outside a lab. What Brain2Qwerty offers instead is a path that asks nothing of the patient's body. No craniotomy, no vessel navigation, no foreign object left behind. For a technology meant to eventually reach millions of people with brain lesions, that distinction may matter as much as the accuracy numbers, since a meaningful share of potential patients will never be candidates for elective brain surgery in the first place.

The log-linear scaling result is what makes the non-invasive route hard to dismiss as a permanent runner-up. Capabilities that first appear under tightly controlled conditions, with specialized equipment and narrow tasks, have a way of migrating outward as the supporting technology matures and the data accumulates. A skater landing a trick once in a padded practice setting tells you the trick is physically possible before anyone lands it in competition.

Brain2Qwerty v2 is that kind of demonstration: proof that non-invasive decoding can approach accuracy levels once thought to require surgery, recorded in a setting that does not yet resemble where it will eventually be used. The path from here runs through sensor engineering, larger datasets, and testing with the patients who stand to gain the most, and none of those steps is guaranteed. But the question has shifted from whether non-invasive decoding can approach implant-level accuracy to how quickly the remaining engineering, on both the AI and the hardware side, can be done.