Meta AI Unveils Brain2Qwerty v2 for Non-Invasive Brain-to-Text

Meta AI introduced Brain2Qwerty v2 this week, a system capable of decoding natural sentences from non-invasive brain recordings in real time. This advanced pipeline reads magnetoencephalography (MEG) signals while a person types, reconstructing their typed output without requiring any surgical implants. Brain2Qwerty v2 represents a significant advancement over its predecessor, Brain2Qwerty v1, which was released in February 2025. Meta AI is also making the complete training code for both versions publicly available.

The Brain2Qwerty v2 pipeline integrates a convolutional encoder, a transformer model, and a character-level language model. This architecture allows it to achieve an average word accuracy of 61% (corresponding to a 39% word error rate), a substantial improvement from the 8% accuracy reported for previous non-invasive methods. In testing, the top-performing participant reached 78% word accuracy, with more than half of their sentences decoded with one word error or less. The system leverages character, word, and sentence-level representations, enabling it to correct local errors by utilizing broader contextual information.

Meta AI trained Brain2Qwerty v2 on approximately 22,000 sentences generated by nine volunteer participants. Each participant was recorded for 10 hours while actively typing, with their brain activity captured by a MEG device. MEG technology measures the magnetic fields produced by neuronal activity, offering high temporal resolution. The data used for training was collected in collaboration with Spain’s BCBL (Basque Center on Cognition, Brain and Language), and the data itself belongs to that research center.

This development is presented as research rather than a consumer product. The decoder is not intended for public use, and its performance was evaluated on a limited group of volunteers. Unlike earlier non-invasive systems that relied on manually designed pipelines to identify neural events, Brain2Qwerty v2 replaces this step with an end-to-end trainable neural network approach, further enhancing its decoding capabilities.