Details
- Meta announced Brain2Qwerty v2, a non-invasive brain-to-text system that decodes full sentences from MEG recordings in real time.
- The research builds on Brain2Qwerty v1 and is part of Meta AI’s broader brain-modeling work, alongside Tribev2, NeuralSet, and NeuralBench.
- Brain2Qwerty v2 uses end-to-end deep learning on raw brain signals and fine-tunes large language models to improve language reconstruction.
- The model was trained on about 22,000 typed sentences from nine volunteers recorded for 10 hours each, and Meta says performance reached 61% word accuracy overall and 78% for its best participant.
- Meta is releasing the full training code for v1 and v2, while its partner BCBL is releasing the v1 dataset, aiming to accelerate open neuroscience research and future clinical communication tools.
Impact
The release strengthens the case for non-invasive brain-computer interfaces as a scalable alternative to surgical implants, especially for people with severe speech or motor impairments. If the reported scaling trend holds, the work could shape next-generation neuroscience R&D toward larger datasets, better foundation models, and broader open benchmarking over the next 12–24 months.
