Brain2Qwerty v2 hits 39% WER — can non-invasive BCI finally beat surgery?

meta released brain2qwerty v2, a model that achieves a 39% word error rate (wer) by decoding continuous brain signals from magnetoencephalography (meg) data without requiring invasive surgery. the system, trained on 10 times more data than its predecessor, uses a deep learning pipeline and a fine-tuned qwen3 language model to convert motor cortex activity into coherent sentences, achieving perfect decoding in 28% of attempts for the best participant. while the model significantly outperforms previous versions in semantic accuracy, its reliance on a language model causes it to hallucinate grammatically correct but factually incorrect sentences when signal data is weak, resulting in a higher character error rate than raw encoders. this research signals a shift toward using scalable software and larger datasets to solve the 'last mile' problem in non-invasive bci, potentially removing the need for high-risk surgical implants for speech restoration.