Meta AI releases Brain2Qwerty v2: non-invasive BCI decodes brain waves to text at 61% word accuracy

Meta's Fundamental AI Research (FAIR) team published Brain2Qwerty v2 on June 29, a non-invasive brain-computer interface system that converts brain activity directly into typed text, reaching 61% word accuracy overall and 78% for the top-performing participant — compared to roughly 8% for previous non-invasive methods.

What's new

Brain2Qwerty v2 is described by its authors as "the highest-performing end-to-end pipeline capable of real-time sentence decoding from non-invasive brain recordings." The system works with magnetoencephalography (MEG) devices, which measure the tiny magnetic fields produced by neural activity without requiring electrodes or surgery. Key metrics:

• 61% word accuracy across all nine participants
• 78% word accuracy for the best-performing participant
• Roughly 7.5× improvement over prior non-invasive methods (~8% baseline)
• Trained on approximately 22,000 sentences, with each participant completing 10 hours of MEG recording sessions while actively typing

The pipeline combines end-to-end deep learning for raw signal processing with fine-tuned large language models that use semantic context to resolve ambiguous neural signals — a design choice that drives accuracy above what signal processing alone can achieve. AI agents were also used to explore pipeline optimization.

Context

Brain-computer interfaces have historically required surgical implants to achieve meaningful typing speeds and accuracy. Neuralink and BrainGate have demonstrated results through implanted electrode arrays, but surgical invasiveness limits the population that can benefit and creates regulatory and adoption hurdles.

Non-invasive methods — EEG, MEG, fNIRS — have existed for decades but delivered poor BCI typing performance. The gap between implanted and non-invasive systems has been wide enough that most clinical BCI work has focused on surgical approaches for patients with severe motor impairments.

Brain2Qwerty's 61–78% range is the highest accuracy reported from a non-invasive MEG-based system. MEG requires large, specialized equipment and is not yet wearable, which limits deployment — but the research result establishes that non-invasive decoding at clinically meaningful accuracy is achievable.

Why it matters

The practical target for Brain2Qwerty is people who have lost the ability to communicate due to brain lesions, ALS, locked-in syndrome, or stroke — a population that currently has few high-quality non-invasive options. Meta frames this explicitly as research for "the millions of people who suffer from brain lesions that prevent them from communicating."

Beyond clinical use, the result shifts the BCI research question: if a non-surgical approach can reach 61% accuracy with MEG today, the path forward is whether portable or wearable brain-scanning hardware (e.g., OPM-MEG or next-generation EEG) can close the hardware gap. The AI architecture — combining neural signal models with LLMs for semantic grounding — is also likely to influence future work in academia and the growing BCI startup ecosystem.