Meta's Brain2Qwerty v2 hits 61 per cent word accuracy decoding typed sentences from MEG

By Inside BCI · June 30, 2026

Nine healthy adult volunteers in Spain sat under a magnetoencephalography (MEG) scanner for roughly ten hours each, typing 22,000 Spanish sentences on a custom non-ferromagnetic keyboard while Meta’s artificial intelligence research team recorded the magnetic fields above their scalps.

Meta AI on 29 June 2026 published a research system called Brain2Qwerty v2 that decoded those typed sentences from the MEG signals at 61 per cent average word accuracy, equivalent to a 39 per cent word error rate as reported in the underlying paper, with 78 per cent accuracy reported for the best individual participant. The system is the next iteration of Meta’s non-invasive brain-to-text research line, building on a Brain2Qwerty v1 paper that appeared the same day in Nature Neuroscience.

Neither version has a consumer-product path. The MEG hardware required is room-installed and needs a magnetically shielded room. Meta FAIR Brain and AI team lead Jean-Rémi King told MIT Technology Review in February 2025 of the v1 system that he did not think “there is a path for products because it’s too difficult.” The v2 paper is currently an arXiv preprint with co-first authors Mingfang (Lucy) Zhang and Jarod Levy at Meta FAIR, Jean-Rémi King as senior author, and the recordings carried out in partnership with the Basque Center on Cognition, Brain, and Language (BCBL) in Spain.

What Brain2Qwerty v2 decodes and how

The decoding task in both v1 and v2 is overt typing under MEG, not silent inner speech or imagined typing. Volunteers see a Spanish sentence presented word by word, memorise it during a fixation interval, and then type it on a non-ferromagnetic keyboard inside the MEG scanner with no visual feedback. The v1 paper describes the BCBL setup as a MEGIN-class 306-channel scanner (102 magnetometers plus 204 planar gradiometers) recording at 1 kHz, and the v2 study was carried out at the same lab with the same hardware class. Mean typing speed in the v1 cohort was approximately 152 characters per minute with around 3.6 per cent keystroke errors at the keyboard.

Brain2Qwerty v2 uses a three-stage neural network: a Conformer module that detects individual character keystrokes from the MEG signal, an Aligner module that converts the keystroke sequence into word-level embeddings, and a fine-tuned large language model that reconstructs the sentence. This architecture is a substantive change from v1, which used a convolutional module, a transformer module, and a 9-gram KenLM character-level language model trained on Spanish Wikipedia. The v2 paper additionally reports that artificial intelligence agents were used to iteratively refine the decoding pipeline through automated code development, with final training configurations selected by the human engineering team.

The v2 cohort differs from v1’s in scale. v1 recruited 35 volunteers, of whom 19 contributed to the MEG analysis. v2 used 9 volunteers for approximately 10 hours each, totalling approximately 22,000 typed sentences. Meta reports in the v2 paper that decoding accuracy improves log-linearly with data volume, without a detected plateau, which the team identifies as the path to narrowing the gap with invasive systems.

The 61 per cent number and what it actually measures

The headline “61 per cent average word accuracy” appears in the Meta blog post. The underlying paper expresses the same statistic as “an average word error rate of 39 per cent.” Word error rate counts substitutions, insertions, and deletions against the ground-truth sentence; 100 per cent minus 39 per cent equals 61 per cent, so the two numbers are identical metrics expressed in opposite framings. Independent readers should expect to see both forms.

The “78 per cent accuracy” for the best participant appears in the blog post. The corresponding paper abstract states that the model accurately decoded half of the best participant’s sentences with one word error or less, without printing the 78 per cent figure directly. The v1 paper reported character error rate rather than word error rate (29 per cent MEG average and 18 per cent best participant in the Nature Neuroscience published version, 32 per cent and 19 per cent in the earlier February 2025 arXiv preprint). v1 and v2 therefore cannot be compared on a single number, because the metrics changed.

Meta’s blog frames the v2 result against a baseline of “8 per cent word accuracy from other non-invasive methods,” citing a 2023 Nature Neuroscience paper. Earlier non-invasive brain-to-text approaches based on P300 spellers, SSVEP visual responses, and motor-imagery decoders have existed for decades. Earlier work from the same Meta FAIR Brain and AI team includes the 2023 Défossez et al. paper on non-invasive speech-perception decoding using contrastive learning, published in Nature Machine Intelligence: that paper decoded what listeners heard from MEG and EEG, a sister problem to v2’s typing decoding rather than a direct predecessor of the Brain2Qwerty system.

Why v2 is still locked in the MEG lab

The MEG hardware Brain2Qwerty v2 runs on is a MEGIN-class room-installed scanner inside a magnetically shielded room. MIT Technology Review’s February 2025 coverage of the v1 system characterised the scanner as weighing roughly half a tonne and costing in the order of two million US dollars per installed system. Meta’s own v2 project page acknowledges that the MEG device used in the study “consists of a large scanner, i.e. a setup inaccessible to most patients.”

The blog frames v2 as “the highest-performing end-to-end pipeline capable of real-time sentence decoding from non-invasive brain recordings.” The technical description in the v2 paper reframes this as “online sentence generation” from continuous recording, which is a step forward from v1’s explicitly offline sentence-level output but is not the same as keystroke-by-keystroke decoding in true real time. Readers should treat “real time” in the Meta blog as the company’s framing of an online but post-sentence pipeline.

A portable MEG path exists in principle through optically pumped magnetometer (OPM) MEG systems being developed commercially by Cerca Magnetics, FieldLine, and QuSpin. Meta cites the OPM line in its v2 materials as the longer-term hardware route. Brain2Qwerty v2 itself was not recorded on portable MEG hardware. Whether the v2 decoding architecture transfers to OPM-quality signals at smaller channel counts is an empirical question that the paper does not answer.

v2 in the BCI commercial map

The invasive intracortical BCI cohort has published higher-throughput decoding results than Brain2Qwerty v2 reports. The 2024 Card et al. New England Journal of Medicine paper from the UC Davis at-home setup reported approximately 56 words per minute at 99.2 per cent accuracy on a 125,000-word vocabulary in a single ALS participant. The 2023 Willett et al. speech neuroprosthesis paper from Stanford Neural Prosthetics Translational Laboratory reported decoding throughput in the same order of magnitude using intracortical electrode arrays in paralysed participants. Brain2Qwerty v2 does not report words per minute and uses a 128-sentence closed-vocabulary task structure, so the throughput comparison is not direct, but the cohort consensus is that invasive intracortical decoding remains faster and more accurate per session than current non-invasive decoding.

The non-invasive cohort has multiple competing approaches at different stages of commercial readiness. Aleph Neuro published what it described as the first 3D ultrasound localization microscopy image of a living human brain through an intact skull on 24 June 2026 (covered separately by Inside BCI). Forest Neurotech is building an implantable acoustic-window approach using the Butterfly Network ultrasound chip. Iconeus in Paris sells preclinical functional ultrasound scanners. Kernel has shipped OPM-MEG (Kernel Flux) and TD-fNIRS (Kernel Flow) commercial systems since 2020. Meta’s Brain2Qwerty sits in the same broad non-invasive lane as these systems but uses MEG rather than ultrasound or near-infrared, and is research-only rather than commercial.

The relationship between Brain2Qwerty (Meta FAIR research, MEG, Paris-led, open source) and the Meta Neural Band (Meta Reality Labs commercial product, surface electromyography on the wrist, shipped September 2025 in a bundle with Meta Ray-Ban Display at 799 US dollars) is one of common parent company only. Meta’s v2 blog makes no reference to the Neural Band. The two streams sit in different organisational units, read different signal classes from different body regions, and have different commercial trajectories.

What to watch

v2 is currently an arXiv preprint hosted on Meta’s content delivery network. Peer review at a high-tier journal is the next regulatory gate. v1 published the same day in Nature Neuroscience, suggesting that the Meta FAIR Brain and AI team operates on a roughly 12-to-18-month preprint-to-Nature-Neuroscience cadence. A v2 journal version in 2027 would convert the architecture claim from a Meta-hosted PDF to a refereed result.

Cerca Magnetics, FieldLine, and QuSpin develop commercial OPM-MEG systems that could in principle host the v2 decoding architecture in a smaller and less expensive form factor than the MEGIN-class room-scale scanner. Whether OPM signal quality at smaller channel counts is enough to recover the 61 per cent word-accuracy figure is the empirical question the next dataset will need to answer.

Replication by independent laboratories is the third gate. Brain2Qwerty has been an open-source project from v1: the v1 dataset is on Hugging Face via the BCBL partnership, and the v1 and v2 training code are on GitHub at facebookresearch/brain2qwerty. The Caltech, Stanford Neural Prosthetics Translational Laboratory, UCSF, and Tsinghua groups are the obvious replicators. A second-group result in 2026 or 2027 would convert v2 from a Meta-internal milestone to a field consensus.

Meta’s Digital Brain Project sits behind the v2 release as a 5 million US dollar open-neuroscience fund and a sister release of TRIBE v2 (perception encoding), NeuralSet, and NeuralBench. If the dataset and benchmark layer Meta is building becomes adopted as the de facto standard for non-invasive decoding research, Meta will have built the infrastructure that the next decade of non-invasive BCI research runs on, regardless of where the consumer product layer eventually settles.

Sources

research Meta-AI Meta-FAIR Brain2Qwerty non-invasive-BCI MEG Jean-Remi-King brain-to-text BCBL Conformer-architecture

What Brain2Qwerty v2 decodes and how

The 61 per cent number and what it actually measures

Why v2 is still locked in the MEG lab

v2 in the BCI commercial map

What to watch

Weekly BCI Brief in your inbox