Summary: Meta announced Brain2Qwerty v2 on June 29, 2026, a non-invasive MEG system that decodes typed sentences in near real time. Nine participants wore the device for about 10 hours each and typed about 22,000 sentences, yielding 61% overall word accuracy versus roughly 8% for earlier non-invasive methods. Meta released training code and dataset support, while noting accuracy varied and the study covered controlled typing.
Brain2Qwerty v2 arrived on June 29, 2026, in a Meta announcement describing a system that decodes sentences from non-invasive brain recordings in near real time: a participant types while wearing a magnetoencephalography (MEG) device, and the model turns the brain signals it reads straight into the sentence being typed. For someone whose brain lesion has taken away communication, the path back has, until now, run almost exclusively through invasive surgery.
Historically, pulling usable text out of brain activity called for implanted electrodes — stereotactic EEG or electrocorticography. As neuroprostheses these have truly handed individual patients their communication back, yet the cost and danger of surgery keep the approach from spreading. The non-invasive alternative has languished at single-digit accuracy: electrodes on the scalp pick up signals the skull has blurred, and the faint magnetic fields MEG measures get lost in noise all too easily. Brain2Qwerty v2 moves that line forward by a wide margin.
How the Study Worked: Nine Participants, 22,000 Sentences
Nine volunteers enrolled, each spending roughly ten hours wearing the MEG device and, between them, typing about 22,000 sentences. Plenty of brain-computer interface studies manage with only a few hours of data; at this scale, deep models finally get enough training material. The size also reflects a deliberate wager: rather than engineering cleverer features for small datasets, Meta collected enough data for a deep learning approach to carry the work.
What matters even more is the scaling pattern the team documented: accuracy climbs log-linearly as data volume grows. Meta reasons from this that closing the gap with invasive approaches may demand nothing more than additional data — no new surgical technique. For the non-invasive route, that is the most optimistic signal there is. Every extra hour of recording now buys a predictable, measurable gain, which recasts progress as an engineering and data-collection problem instead of a waiting-for-a-breakthrough problem.
Deep Learning End to End, with Language Models Tuned on Neural Data
Instead of the customary hand-crafted feature pipeline, the system runs end to end: a deep neural network takes raw brain signals and predicts the text itself. Noise is the stubborn problem, so the team fine-tunes large language models on neural data, and the semantic context an LLM supplies repairs garbled decodings until they read as coherent sentences. Here lies the work’s understated architectural idea: the language model functions as a prior on what plausible sentences look like, which frees the decoder from having to pull every word out of the signal by itself — it only has to get far enough for that prior to finish the job. Meta adds that AI agents helped search for pipeline optimizations during training, with engineers keeping the final word on training decisions.
From 8% to 61%: Non-invasive Decoding’s Big Jump
The whole result comes down to two numbers: 61% word accuracy overall, versus roughly 8% for earlier non-invasive methods. The strongest participant reached 78%, and across all sentences, more than half came out of decoding with at most one word wrong. On natural sentences, no non-invasive system has ever come this close to a practical threshold before.
Two caveats keep the result honest. For one, 61% is an average across participants, and the spread is wide — decoding quality differs substantially between individuals. For another, the experiment covered typed sentences in a controlled setting, not open-ended conversation; a clinical communication tool would need to cope with unprompted, continuous speech intent. Even so, moving from single digits to a majority of words correct alters what kind of question this is: no longer whether non-invasive decoding can work, but how quickly it can be driven to where implanted devices already stand.
Public Code, Public Datasets, and a $5M Fund
Meta has published the training code for v1 and v2 alike; BCBL, the partner institute, has made the v1 dataset available; and a $5 million fund under the Digital Brain Project backs open neuroscience datasets. The v1 findings appeared earlier in Nature Neuroscience.
Brain2Qwerty does not stand alone, either. It belongs to Meta’s wider brain-modeling stack, together with the Tribev2 encoding model, the NeuralSet large-scale brain data pipeline, and the NeuralBench evaluation benchmark. Building in-house tooling for encoding, data, and evaluation at once says that Meta regards non-invasive neural decoding as a research track to run for years, not a single demonstration. The broader research community benefits as well: open release lowers the entry barrier in a field where data collection is the hardest step, so other groups can now train against — and benchmark against — published baselines rather than starting from nothing.