Meta has released Brain2Qwerty v2, an AI system that decodes brain activity into text without the need for a surgical implant. Published on 29 June, the update follows last year's Brain2Qwerty v1 and is described by the company as its highest-performing end-to-end pipeline for real-time sentence decoding from non-invasive brain recordings.
The system was trained on around 22,000 typed sentences gathered from nine volunteers, each of whom wore a magnetoencephalography (MEG) device for ten-hour recording sessions. Meta's team moved away from hand-built systems for spotting neural signals, instead training the model end-to-end so it decodes text straight from raw brain data. The neural signals were combined with fine-tuned large language models, letting the system draw on sentence-level meaning to make sense of noisy inputs. Engineers also used AI agents to test possible improvements to the decoding pipeline, though final training choices were made by hand.
The result is a word accuracy rate of 61%, well ahead of the 8% typical of other non-invasive techniques. Its top-performing volunteer reached 78% accuracy, with over half of that participant's sentences decoded with no more than one wrong word. Meta says results keep improving steadily as more data is added, hinting that the remaining accuracy gap with invasive, surgery-based systems might eventually close through data alone.
Meta is releasing the full training code for both versions, while research partner the Basque Center on Cognition, Brain and Language is publishing the v1 dataset. The work fits into a wider push at Meta to build open brain-focused models, including Tribev2, NeuralSet and NeuralBench, supported by a $5 million fund for open datasets under its Digital Brain Project.
