Meta's Brain2Qwerty v2 Turns Brain Waves into Text

Imagine words appearing where there is only silence. For millions who have lost the ability to speak because of neurodegenerative disease or severe brain injury, silence is not a temporary pause—it is daily reality. Meta says its new Brain2Qwerty v2 could change that.

Inside the lab where thoughts turned into letters

Researchers in San Sebastian asked nine healthy volunteers, aged 25 to 56, to type more than 2,500 sentences across ten sessions while wearing a magnetoencephalography device, or MEG. That machine detects the tiny magnetic fields produced by neurons firing. The recorded brain activity, paired with the typed text, produced the training data that taught the model to map neural signals to language.

The jump in performance was striking. Where the first Brain2Qwerty reached roughly 48 percent word-level accuracy, version two hit about 78 percent in its best runs. Practically speaking, that improvement means most decoded sentences contained at most a single incorrect word. The team also notes a simple truth: more training data equals better decoding.

How did they get there? The engineers borrowed pattern-recognition techniques popularized by large language models. The pipeline runs in three stages. First, a neural network converts raw MEG signals into tokens that resemble letters. Next, an aligner reassembles those tokens into candidate words. Finally, a large language model smooths the output into fluent, meaningful sentences.

There are clear benefits. Because Brain2Qwerty v2 uses noninvasive MEG rather than implanted chips, it offers a path to communication that avoids risky brain surgery. People with locked-in syndrome, ALS, and similar conditions could one day use this technology to express themselves without implants.

Meta is not keeping this to itself. The company has released the base code for both version one and version two as open source, inviting researchers worldwide to reproduce, validate, and extend the work. That openness could accelerate medical research and help teams scale the data and safeguards this field needs.

Still, hurdles remain. MEG systems are large and expensive, and real-world patients present messier signals than healthy volunteers. Privacy and consent questions are real and urgent when machines start decoding thoughts. Clinical trials, more diverse datasets, and careful ethical frameworks will determine whether this lab success becomes a reliable tool for patients.

For now, Brain2Qwerty v2 reads like a milestone: it shows noninvasive neural decoding moving from hopeful experiment toward practical possibility. The next steps will decide how quickly that possibility becomes a lifeline.