Meta open-sources Brain2Qwerty v2, hits 61% MEG text accuracy

Meta's FAIR lab open-sourced Brain2Qwerty v2, a non-invasive system that decodes typed sentences from MEG brain recordings, and the accuracy numbers are the part that has people paying attention. The code and v1 dataset are on GitHub, with the v2 dataset under embargo pending journal publication.

The pipeline is a convolutional encoder that reads raw MEG signals, a transformer, and a character-level language model. On the v2 setup, as reported by MarkTechPost, the system was trained on around 22,000 sentences from nine volunteers, each recorded for 10 hours while typing inside an MEG scanner. Average word accuracy across those participants was 61%, the best participant reached 78%, and more than half of decoded sentences came out with one word error or fewer. Meta contrasts that with prior non-invasive approaches sitting near 8% word accuracy, which is a real jump if you take the comparison at face value.

Why this matters if you don't work in neuroscience: the credible non-invasive brain-to-text numbers have been low for a while, and a 61% average result is a different accessibility story, even though the MEG hardware doing the reading is a room-sized shielded scanner rather than anything wearable. For context, the earlier v1 study from 35 volunteers, being published in Nature Neuroscience, reported an average MEG character error rate of 32% against 67% for EEG, with the best participant reaching a 19% CER.

What the reporting doesn't give you is real-time latency in practice, how the model behaves outside constrained sentence typing, or how far generalization holds when you leave the training cohort. The v2 data embargo also means independent replication is on hold. Still, for anyone tracking the arc from implanted BCIs toward non-invasive interfaces, watching a 61% word-accurate MEG pipeline land as open source is worth flagging.