Overview
- Meta announced Brain2Qwerty v2 in late June 2026 and published training code for v1 and v2 while the full v2 dataset remains pending journal release.
- The system uses magnetoencephalography (MEG) — a helmet‑style, room‑scale scanner that measures tiny magnetic fields from brain activity — to record volunteers typing memorized sentences.
- Meta reports v2 was trained on about 22,000 sentences from nine volunteers (roughly 10 hours of MEG data each) and achieved an average word accuracy of 61 percent, with the best participant reaching 78 percent.
- Brain2Qwerty v2 replaces the previous keypress‑timing pipeline with an end‑to‑end AI stack that detects characters, aligns words, and uses a fine‑tuned large language model to turn noisy signals into sentences in real time.
- Major barriers remain: MEG hardware is large and costly, the experiments decoded planned typing not free thought, accuracy is still below everyday‑use or clinical standards, and Meta says more data and smaller wearable sensors are needed to close the gap with implanted BCIs.