via MarkTechPost
Meta AI Unveils Brain2Qwerty v2: A Non-Invasive Brain-to-Text Pipeline Achieving 61% Word Accuracy
Meta AI has unveiled Brain2Qwerty v2, a cutting-edge end-to-end deep learning system that decodes typed sentences directly from non-invasive brain recordings in real time. Developed by Meta's Fundamental AI Research (FAIR) lab, this system leverages magnetoencephalography (MEG) signals to achieve a word accuracy of 61%, marking a significant milestone in brain-computer interface (BCI) technology without requiring surgical implants.
Published in June 2026, Brain2Qwerty v2 represents an evolution from its predecessor, offering improved decoding accuracy and efficiency. By processing high-temporal-resolution MEG data, the pipeline learns to map neural activity patterns to keystrokes and entire typed sentences. This breakthrough is particularly notable for its non-invasive approach, avoiding the risks associated with implanted electrodes while still delivering competitive performance.
The system is designed to decode natural language in real-world typing scenarios, with potential applications in assistive technology for individuals with speech or motor impairments. Meta has released the model as open source on GitHub (facebookresearch/brain2qwerty), inviting further research and development in non-invasive BCI systems.
As of 2026, this advancement aligns with broader trends in neural decoding, where deep learning models are increasingly able to interpret complex brain signals for communication. While still in research stages, Brain2Qwerty v2 sets a new benchmark for non-invasive brain-to-text pipelines, offering a glimpse into a future where thought-based typing may become a practical reality.