On June 29, 2026, Meta published Brain2Qwerty v2, a system that decodes sentences from non-invasive brain recordings in near real time: participants wearing a magnetoencephalography (MEG) device type, and the model converts their brain signals directly into the sentences they are entering. For people who have lost communication to brain lesions, that road has, until now, mostly run through invasive surgery.
Getting usable text out of brain signals has typically required implanted electrodes — stereotactic EEG or electrocorticography. Those neuroprostheses have genuinely restored communication for individual patients, but surgical cost and risk make them hard to scale. Non-invasive methods, meanwhile, have been stuck at single-digit accuracy: scalp electrodes pick up signals blurred by the skull, and the faint magnetic fields MEG measures are easily lost in noise. Brain2Qwerty v2 pushes that boundary a long way forward.
The Experiment: Nine Participants, 22,000 Sentences
Nine volunteers took part, each wearing an MEG device for roughly ten hours and typing a total of about 22,000 sentences. Compared with many brain-computer interface studies that work from just a few hours of data, that scale gives deep models enough training material. It also reflects a deliberate bet: rather than engineering better features for small datasets, Meta chose to collect enough data for a deep learning approach to take over.
More important is the scaling pattern the team observed: accuracy improves log-linearly with data volume. Meta’s inference is that closing the gap with invasive approaches may require nothing more than more data — no new surgical technique. That is the single most optimistic signal for the non-invasive route. Every additional hour of recording now buys a predictable, measurable gain, which turns progress into an engineering and data-collection problem instead of a waiting-for-a-breakthrough problem.
End-to-End Deep Learning plus Fine-tuned Language Models
The system skips the traditional hand-crafted feature pipeline and goes end to end: a deep neural network predicts text directly from raw brain signals. To handle noisy recordings, the team fine-tunes large language models on neural data, letting the semantic context an LLM provides patch noisy decodings into coherent sentences. That is the quiet architectural insight of the work: the language model acts as a prior over what plausible sentences look like, so the decoder does not have to recover every word from the signal alone — only enough for the prior to take over. Meta also notes that AI agents helped search for pipeline optimizations during training, with engineers keeping the final say on training decisions.
61% vs 8%: The Leap in Non-invasive Decoding
The results compress into two numbers: 61% word accuracy overall, against roughly 8% for prior non-invasive methods. The best participant reached 78%, with more than half of all sentences decoded with one word of error or fewer. It is the first non-invasive system to approach a practical threshold on natural sentences.
Two caveats frame the result honestly. The 61% figure is a mean across participants, and the spread is wide — decoding quality varies considerably from person to person. And the experiment covers typed sentences in a controlled setting, not open-ended conversation; a clinical communication tool would need to handle unprompted, continuous speech intent. Still, going from single digits to a majority of words correct changes the category of the question, from whether non-invasive decoding can work at all to how quickly it can be pushed to the level implanted devices already reach.
Open Code, Datasets, and a $5M Fund
Meta released the training code for both v1 and v2; partner institute BCBL released the v1 dataset; and a $5 million fund under the Digital Brain Project supports open neuroscience datasets. The v1 results were previously published in Nature Neuroscience.
Brain2Qwerty is not an isolated project, either. It sits inside Meta’s broader brain-modeling stack — the Tribev2 encoding model, the NeuralSet large-scale brain data pipeline, and the NeuralBench evaluation benchmark. Building tooling for encoding, data, and evaluation alike signals that Meta treats non-invasive neural decoding as a long-term research track, not a one-off demonstration. For the wider research community, the open release lowers the barrier to entry in a field where data collection is the hardest step: other groups can now train against, and benchmark against, published baselines instead of starting from zero.
Sources
- From Brain Waves to Words — Meta AI
- Accurate decoding of natural sentences from non-invasive brain recordings — Meta Research
- Brain2Qwerty training code — GitHub
- Brain2Qwerty — Nature Neuroscience
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.