Meta has open-sourced Brain2Qwerty v2, a noninvasive brain-computer interface that decodes thoughts into sentences using EEG or MEG signals, achieving 61% word-level accuracy—far surpassing other noninvasive approaches. By releasing the model code and training data, Meta aims to advance open brain science and potentially restore communication for people with severe motor disabilities.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Meta open-sources Brain2Qwerty v2, a non-invasive brain-computer interface that decodes sentences from brain signals with 61% word accuracy, halving the word error rate of its predecessor. Performance gains come primarily from a tenfold increase in training data, not architectural breakthroughs. MEG signals significantly outperform EEG, but data collection remains a bottleneck.
Meta recently open-sourced Brain2Qwerty v2, a non-invasive brain-computer interface system that decodes brain signals collected via electroencephalography (EEG) or magnetoencephalography (MEG) into complete sentences. In evaluations, Brain2Qwerty v2 achieved an average word recognition accuracy of 61%, compared to just 8% for other non-invasive approaches.
Meta found that MEG signals significantly outperform EEG: average character error rates (CER) of 29% and 65%, respectively. Compared to the baseline model EEGNet, Brain2Qwerty's character error rate is reduced by about 2.5 times.
Brain2Qwerty v2 consists of three core modules: an Encoder that receives brain signals and predicts characters; an Aligner that combines characters into words; and an LLM that generates the final output based on the aligned results. This architecture also brings an unexpected benefit: even if users make spelling errors during input, the system can automatically correct these 'typo-like' issues.
Meta previously released Brain2Qwerty v1 in 2025, and the new model's word error rate (WER) is nearly halved compared to its predecessor, further narrowing the performance gap with invasive brain-computer interfaces.
Discussing the release on X, io.net co-founder Tory Green compared Brain2Qwerty v2 with the previous generation: 'It looks like the performance improvement from v1 to v2 comes almost entirely from a tenfold increase in training data, not from architectural breakthroughs. This result is actually very encouraging because it means the real bottleneck is the ability to collect and annotate MEG data, not that the problem itself is approaching a theoretical limit. Data scale constraints like this tend to be resolved faster than people expect.'
To advance open brain science, Meta has made the model code and training data fully public. Meta states: 'We believe this research has the potential to bring real help to millions of people who have lost the ability to communicate due to brain injuries. We also found that as data scale increases, decoding accuracy grows log-linearly. This means that simply by continuously expanding data scale, we can hope to further narrow the performance gap between non-invasive and implantable approaches.'
'Through the $5 million fund established by the Digital Brain Project, we are working closely with the research community to promote the development of open datasets. We hope that this open research can accelerate neuroscience and speed up the discovery, diagnosis, and treatment of neurological diseases, rather than confining related research to isolated systems.'
Meta notes that previous development of non-invasive brain-computer interfaces has been limited by the high noise and complexity of brain signals. In contrast, invasive approaches like electrocorticography (ECoG) can obtain more reliable signals but require surgical implantation, making them 'difficult to scale widely.' Brain2Qwerty v2's code has been released on GitHub, and training data is available for download from Hugging Face. Brain2Qwerty is part of Meta's Digital Brain project, which aims to advance brain activity modeling in an open-source manner for scientific research and medical applications. Other open-source outcomes of the Digital Brain project include NeuralSet—a Python toolkit for processing neural signals such as MEG and EEG—and NeuralBench—a unified benchmark framework for evaluating AI models of brain activity.
This article is based on InfoQ's coverage of Meta's open-source release of Brain2Qwerty v2, with information sourced from Meta's official statements and third-party commentary. Performance data (61% accuracy, 29% and 65% character error rates, etc.) come from Meta's evaluations; Tory Green's analysis represents personal opinion. All facts are derived from a single source and have not been independently verified.
Brain2Qwerty v2 demonstrates the great potential of non-invasive BCIs, but data bottlenecks remain a major challenge. Meta's open-source strategy may accelerate research, but whether performance improvements are sustainable depends on the acquisition of larger-scale, high-quality MEG data.
Primary report
Primary source