Meta's Brain2Qwerty v2 hits 61% word accuracy decoding text from brain scans — no surgery

Meta’s AI research group has released Brain2Qwerty v2, a system that turns non-invasive brain recordings into typed sentences in real time. Trained on roughly 22,000 sentences from nine volunteers — each wearing a magnetoencephalography (MEG) helmet for 10 hours while typing — the model reaches a 61% word accuracy rate, and 78% for its best-performing participant. That’s a large jump over the ~8% managed by earlier non-invasive methods, and it starts to approach the quality of surgical implant techniques like electrocorticography without cutting into anyone’s skull. Meta is open-sourcing the training code for both v1 and v2, while its partner BCBL is publishing the v1 dataset.

The technical shift is a move away from hand-built signal-processing pipelines toward end-to-end deep learning that decodes straight from raw MEG signals. Fine-tuning large language models on the neural data lets the system lean on semantic context to reconstruct coherent sentences from noisy input, and Meta says it used AI agents to search for pipeline optimizations before engineers picked the final configurations by hand. Notably, accuracy scales log-linearly with data volume, implying the remaining gap with implanted electrodes might close through more recordings alone rather than new algorithms.

The stated goal is restoring communication for people with brain lesions, since surgical neuroprostheses work but don’t scale. The release slots into a broader Meta push toward open brain-foundation models — including its Tribev2 perception encoder, NeuralSet data tooling, NeuralBench evaluation suite, and a $5 million fund for open neuroscience datasets. For a technical audience, the interesting tensions are the scaling claim and the obvious longer-term question a consumer-grade thought-decoder raises about the sensitivity of neural data.