Meta debuts non-invasive Brain2Qwerty v2 decoder

Meta published a research paper and blog describing Brain2Qwerty v2, a non-invasive brain-computer-interface pipeline that decodes typed sentences from magnetoencephalography (MEG) recordings. The team trained the model on about 22,000 sentences collected from nine volunteers, each contributing roughly 10 hours of MEG data while actively typing. Meta reports average word accuracy of 61% across participant-specific models, with the best individual participant reaching 78%, compared with about 8% for prior non-invasive methods. Meta has open-sourced both the training code and dataset, making the work directly reproducible by other BCI researchers.

The system feeds raw MEG signals into an end-to-end model, combining a convolutional encoder, a transformer, and a character-level language model, rather than relying on a manually staged signal-processing pipeline. The dataset captures typed-sentence production rather than attempted speech or imagined speech, which simplifies alignment between neural signals and characters. Meta's documentation also describes using AI agents to explore and select training configurations and hyperparameters.

The result illustrates how semantic priors from language models can raise tolerance to noisy neural signals, but small, participant-specific cohorts (nine people) also make it easier to show large relative gains that may not hold up on more diverse populations or portable, lower-fidelity sensors. Researchers building on this should treat the 61%/78% figures as lab-MEG, per-participant results rather than device-agnostic benchmarks.

Replication on larger and more diverse participant pools, cross-subject and cross-device transfer performance, real-time latency in less controlled settings, and progress on portable or lower-cost MEG alternatives. Also watch for peer-review outcomes on the arXiv preprint and any independent benchmarks run under a shared evaluation protocol.