Meta FAIR released Brain2Qwerty v2 on June 29, 2026 — a non-invasive brain-to-text decoder that reconstructs full sentences from magnetoencephalography (MEG) recordings while a person types. The participant wears nothing more invasive than a helmet of magnetic sensors. No surgery, no implants, no Neuralink-style brain chips.
The system achieves 61% average word accuracy across nine volunteers, with the best participant reaching 78%. At the character level, the word error rate (WER) is 39% on average, dropping to 22% for the top performer. These numbers may seem modest compared to the near-perfect transcription we expect from speech-to-text systems, but the context matters: this is raw brain activity, captured non-invasively, decoded in real time.
From Characters to Sentences
The v1 system, published in Nature Neuroscience on the same day last year, decoded individual characters — impressive but far from usable communication. v2 leaps to full sentence decoding, trained on roughly 22,000 sentences, a 10x increase in training data. The model processes continuous MEG recordings as a person types and reconstructs the intended words, capturing semantic meaning rather than just keystroke-level signals.
The architecture is end-to-end: raw MEG signals go in, sentences come out. The system was trained on data from volunteers typing inside an MEG scanner, with the model learning to map distributed neural activity patterns to corresponding text.
Open Science, Open Source
In an era where Meta has pivoted much of its frontier AI work toward the proprietary Muse Spark line, the Brain2Qwerty project represents a different philosophy. The code and dataset are fully open source on GitHub at facebookresearch/brain2qwerty. Meta also announced a $5 million fund for open neuroscience research, signaling that FAIR’s fundamental research arm remains committed to open science even as the commercial side shifts.
The release follows Meta’s broader neuroscience push. Just weeks earlier, the company published research on decoding visual representations from MEG signals, building toward a comprehensive non-invasive brain-computer interface stack.
The Non-Invasive Advantage
Brain2Qwerty v2’s approach stands in direct contrast to Neuralink and other implant-based brain-computer interfaces. Those systems achieve higher accuracy — Neuralink’s N1 implant has demonstrated thought-to-text with character error rates below 5% — but require brain surgery. Meta’s bet is that algorithmic improvements can close the gap without crossing the surgical threshold, making the technology accessible to millions rather than the select few willing to undergo invasive procedures.
The primary target population is people who have lost the ability to speak due to neurological conditions such as ALS, stroke, or locked-in syndrome. For these individuals, even 61% word accuracy represents a communication channel where none existed before.
What’s Next
The current system still requires an MEG scanner — a room-sized, multimillion-dollar machine that can only operate in magnetically shielded environments. Practical deployment will depend on either cheaper, more portable brain-sensing hardware (Meta has invested in wearable MEG technology) or algorithmic advances that work with lower-fidelity signals from EEG or fNIRS. Both paths are active areas of research at FAIR.
The roadmap outlined in the accompanying blog post points toward improving robustness across participants, reducing the calibration burden (currently the system requires per-participant training), and extending the approach to imagined speech — decoding what someone wants to say without requiring them to physically type.