Meta Brain2Qwerty v2 reaches 61% word accuracy on non-invasive MEG sentence decoding

Meta released Brain2Qwerty v2, an end-to-end deep learning pipeline that maps 10-hour MEG sessions directly to text. Nine volunteers produced 22,000 sentences while typing. The model bypasses hand-crafted event detection and instead fine-tunes large language models on neural embeddings, recovering coherent sentences from noisy recordings.

Prior non-invasive baselines reached 8% word accuracy. v2 improves this to 61% overall and 78% for the best participant, with more than half of decoded sentences containing one word error or fewer. Accuracy scales log-linearly with data volume, indicating that additional recordings alone can narrow the remaining gap to stereotactic or electrocorticography implants.

Open release of v1 and v2 training code plus the BCBL dataset enables external replication and scaling experiments. The work connects directly to accessibility needs for patients with brain lesions while supplying foundational models for perception encoding and large-scale neural benchmarks.

Next steps center on expanding participant hours and testing whether the same pipeline generalizes to silent speech or imagined typing without motor output.