Meta's Noninvasive Brain-Computer Interface Hits 61% Word Accuracy with Open-Source Release

Meta has unveiled Brain2Qwerty v2, an open-source noninvasive brain-computer interface that translates neural activity directly into written sentences. The system interprets electroencephalography (EEG) and magnetoencephalography (MEG) signals from the scalp to reconstruct text that a user is thinking about, reaching 61% word accuracy—a dramatic leap over the 8% baseline of competing noninvasive approaches.

How It Works#

The system operates through three distinct components. An encoder processes raw brain signals to generate character predictions. An aligner then groups those characters into coherent words. Finally, a large language model refines the output, leveraging linguistic context to produce natural-sounding text. A surprising side effect: the LLM can correct misspellings, effectively smoothing out errors that arise from noisy neural decoding.

During testing, participants were shown sentences and asked to mentally rehearse them before typing normally. MEG sensors—which measure magnetic fields—significantly outperformed EEG electrodes, achieving a 29% character error rate versus EEG's 65%.

Narrowing the Gap with Invasive Methods#

Invasive brain implants like electrocorticography have long held the accuracy advantage, but they demand surgery and are difficult to scale. Brain2Qwerty v2 cuts the performance gap roughly in half compared to its 2025 predecessor, suggesting that noninvasive approaches may eventually rival surgical alternatives if data volume continues to grow.

One key insight emerged from the research: improvements scale logarithmically with the size of the training dataset. This finding points to data as the main bottleneck, not fundamental limits in the underlying approach. As io.net co-founder Tory Green noted, the v2 leap stemmed almost entirely from ten times more training data rather than architectural innovation—a sign that the problem may yield to brute-force scaling faster than skeptics expect.

Open Science and Clinical Impact#

Meta has released both the model code and training data publicly. The work is part of Meta's Digital Brain Project, a broader initiative to open-source neuroscience research. The company has also pledged $5 million to fund open datasets and complementary tools like (for processing MEG and EEG signals) and (for benchmarking brain-activity models).