Meta: Brain2Qwerty v2 — Non-Invasive Thought-to-Text Decoding at 61% Accuracy, Without Surgical Implants

Meta Research has released Brain2Qwerty v2, a system that converts brain signals recorded outside the body — without surgery — into typed text with an average word-level accuracy of 61 percent. The research represents the largest measured leap in non-invasive brain-computer interface (BCI) technology ever recorded in the literature.

How Brain2Qwerty v2 Decodes Brain Activity

Brain2Qwerty v2 relies on MEG — magnetoencephalography, a technology that captures extremely weak magnetic fields produced by the electrical activity of neurons. The MEG scanner is worn like a helmet outside the skull and requires no surgical intervention or implants in brain tissue.

The system combines end-to-end deep neural decoding with fine-tuning of a large language model on brain data. Each of nine participants typed sentences inside the MEG scanner over ten hours of recording, producing approximately 22,000 training sentences. The system’s accuracy grows log-linearly with the amount of collected data — more recordings, better results.

Results and Comparison With Existing Methods

Average word recognition accuracy is 61 percent. For the best participant, Brain2Qwerty v2 reached 78 percent accuracy — more than half of sentences were decoded with one error or none at all.

Comparison with the rest of the field speaks for itself: other non-invasive BCI methods achieve only 8 percent word-level accuracy. Brain2Qwerty v2 exceeds that result seven times and reaches a level comparable to invasive implants that require neurosurgery. The gap between 8 percent and 61 percent makes Brain2Qwerty v2 the first non-invasive system that realistically enters the realm of clinical application.

Does This Mean the End of Surgical Implants?

Invasive BCI implants — such as those developed by Neuralink — still achieve higher precision and greater real-time information throughput, but involve neurosurgical intervention with associated medical risks. Brain2Qwerty v2 does not replace implants, but opens an alternative for patients for whom surgery is excluded — for example, individuals with amyotrophic lateral sclerosis (ALS) in early stages, brainstem injuries, or paralytic conditions. Meta Research emphasizes that the research is intended for the millions of people with motor impairments who have difficulty or are unable to communicate.

Open-Source Release and Code Availability

Meta has publicly released the complete training code for Brain2Qwerty v1 and v2 on GitHub alongside the research publication. The v1 dataset, collected in collaboration with BCBL (Basque Center on Cognition, Brain, and Language), is available on HuggingFace.

The open release of code and data enables independent replication of results and further development by the research community working on non-invasive BCI systems — a rare step for research of this level of commercial and medical potential.