Meta Brain2Qwerty v2 reaches 61% sentence decoding without surgery

61% word accuracy is the number that makes Meta’s Brain2Qwerty v2 worth watching. In a June 29, 2026 research post, Meta describes an end-to-end pipeline that decodes typed sentences from magnetoencephalography, or MEG, without a surgical implant. The comparison point is sharp: Meta says other non-invasive methods reached 8% word accuracy, while v2 moves the result into a range that can be evaluated as a serious communication research path.

The training setup used roughly 22,000 sentences from nine volunteer participants. Each participant spent 10 hours wearing an MEG device while actively typing. Instead of building a hand-crafted pipeline to detect neural events, the system learns directly from raw brain signals and uses language-model fine-tuning to turn noisy neural activity into coherent text.

The strongest single-participant result reached 78% word accuracy. Meta says that, for this participant, more than half of all decoded sentences had one word error or less. That does not put non-invasive decoding on equal footing with invasive neuroprostheses, which can access cleaner signals through stereotactic electroencephalography or electrocorticography. It does change the engineering question: how much of the remaining gap comes from signal limits, and how much can be narrowed through more paired neural-language data?

Meta’s answer is cautiously data-heavy. The post says decoding accuracy improves log-linearly with data volume, implying that scale alone may close part of the distance to surgical approaches. The caveat is just as important: the reported study is based on nine volunteers in a controlled typing setup, not a broad patient population using the system in everyday communication.

The release also matters because Meta is putting more of the stack in the open. The company is releasing the full training code for Brain2Qwerty v1 and v2, while the Basque Center on Cognition, Brain, and Language is releasing the v1 dataset on Hugging Face. The next proof points are external replication, broader participant coverage, and whether the 61% and 78% results hold when the task moves beyond carefully recorded lab sessions.