Meta recently open-sourced Brain2Qwerty v2, a noninvasive Brain–Computer Interface (BCI) that can decode sentences from thoughts using electroencephalography (EEG) or magnetoencephalography (MEG) signals from the brain. In evaluations, the system achieved a word accuracy rate 61% on average, compared to 8% for other non-invasive methods.
Brain2Qwerty uses a three-stage deep-learning model to predict characters from the brain signals. During data collection, participants were shown sentences and asked to remember them before typing. Meta found that the MEG signals performed better, with a character error rate (CER) average of 29% vs. EEG's 65%. Compared to the baseline EEGNet model, Brain2Qwerty had a 2.5x better CER. To help research on open models of the brain, Meta made both the model code and the training data available online. According to Meta,
We believe this research has the potential to make a real difference for the millions of people who suffer from brain lesions that prevent them from communicating...We also find that decoding accuracy improves log-linearly with data volume, suggesting that the remaining performance gap with surgical approaches could be further narrowed through data scaling alone...We do this in close collaboration with the community, through our recent $5 million fund to stimulate open datasets in our Digital Brain Project. Our hope is that this work, done in the open, advances neuroscience to identify, diagnose, and treat neurological disorders faster than in siloes.
Previous work in non-invasive techniques has been limited by the "noise complexity" in the brain signals. Invasive techniques such as electrocorticography (ECoG) are more reliable, but because they require surgery they are "difficult to scale," according to Meta. In 2025, Meta released Brain2Qwerty v1; the new model has a word error rate (WER) that is nearly twice as good, "significantly narrowing the gap" with the WER of invasive techniques.
Brain2Qwerty v2 contains three modules: an Encoder that takes brain signals as input and outputs character predictions; an Aligner that groups characters into words; and an LLM that generates the final output from the aligned data. One unexpected result of this architecture is that the system can correct "typographical" errors when the human users misspell words.
In a post on X about the new release, io.net co-founder Tory Green compared the performance of Brain2Qwerty v2 to v1: Seems like the jump from v1 to v2 came almost entirely from 10x more training data, not an architectural breakthrough. That's actually the more exciting result. It means the limiting factor right now is labeled data from people wearing MEG headsets, not the fundamental difficulty of the problem. Constraints like that have a way of getting solved faster than people expect.
The Brain2Qwerty v2 code is available on Github and the training data can be downloaded from Huggingface. Brain2Qwerty is part of Meta's Digital Brain project, which "[open-sources] the modeling of brain activity for science and medicine." Other Digital Brain artifacts include NeuralSet, a Python package for processing neural signals such as MEG and EEG; and NeuralBench, a "unified framework for benchmarking AI models of brain activity."