Meta’s Brain2Qwerty v2 Hits 61% Word Accuracy Without Surgery

Meta announced Brain2Qwerty v2 on Monday, an AI system that turns magnetoencephalography (MEG) brain signals into typed sentences at 61% word accuracy without surgery. The result came from nine volunteers who each typed for 10 hours while wearing a helmet-like MEG scanner. Meta released the full training code alongside a companion paper in Nature Neuroscience.

Even at 61%, the error rate is high for everyday communication, and the MEG scanner requires a magnetically shielded room that locks the setup inside research labs. Meta says decoding accuracy rises log-linearly with more training data, with no plateau yet visible. The team says it will collect more recorded sentences and run more volunteer sessions. The full training code is being released alongside the paper.

From 8% to 61% in a Single Model

Most of the public attention in brain-computer interfaces goes to companies that put electrodes inside the skull. Meta’s quiet leap happened on the non-invasive side, where the published baseline sat at 8% word accuracy. The new model’s headline number is 61% across participants, with the best individual reading at 78%.

On that top score, more than half of every sentence the model decoded came back with one word error or fewer. Meta’s own write-up calls Brain2Qwerty v2 “the highest-performing end-to-end pipeline capable of real-time sentence decoding from non-invasive brain recordings” and says it “approaches levels of accuracy previously exclusive to techniques that require brain surgery.” The write-up is at the Brain2Qwerty v2 research write-up. Both claims are Meta’s, attached to a single study that has not yet been independently replicated.

The 8% baseline is Meta’s own comparison point, drawn from prior non-invasive work the company cites without naming each paper. The average score comes from nine volunteers who each typed for 10 hours wearing the MEG helmet. The top score came from the same set, at 78% word accuracy. Both numbers are the work of a single Meta study that has not yet been independently replicated. There is no consumer or clinical product attached to the release, and no regulatory filing either.

• 61% average word accuracy across nine volunteers
• 78% word accuracy for the best participant
• 22,000 typed sentences per participant (about 10x the v1 training set)
• 9 volunteers, each recorded for 10 hours
• Compared to roughly 8% for prior non-invasive decoding methods

How the Pipeline Reads Brain Signals

The setup starts with a sensor helmet. The signal coming out is raw.

Magnetoencephalography, or MEG, picks up the tiny magnetic fields produced by neurons firing inside the brain, and the scanner used in the study is bulky enough to fill a small room. Meta trained Brain2Qwerty v2 on roughly 22,000 sentences typed by each of nine volunteers while they wore the helmet. The recordings were fed into an end-to-end deep-learning model that decodes sentences straight from the raw brain signal without first detecting individual keystrokes. Fine-tuning large language models on neural data allows the system to leverage semantic context, bridging the gap between noisy brain recordings and coherent language. Meta also says it deployed AI agents to scan the space of training configurations, with engineers hand-picking the final setup.

The decoder itself is built from three hierarchical modules that handle letters, words, and sentences together, and v2 is trained on roughly 10x more data per participant than v1. v1 was accepted at Nature Neuroscience in 2025 and could only decode by aligning its predictions to the timing of individual keystrokes. v2 removes that constraint and generates sentences directly from continuous brain activity, allowing real-time decoding without per-keystroke alignment.

The architecture, training code, and decoded-sentence examples are on the Brain2Qwerty project page with decoded examples. The page walks through the three hierarchical modules and the LLM fine-tuning step. The v1 Nature Neuroscience paper is linked from the same page. Meta’s full training code for both versions is on GitHub.

Brain2Qwerty v2 recovers sentences coherently from noisy neural inputs, achieving a word accuracy rate of 61%, significantly improving upon the 8% word accuracy from other non-invasive methods.

Meta, in the Brain2Qwerty v2 research write-up.

Two Walls Still Stand Between Lab and Clinic

Meta names two challenges. The first is accuracy. The second is hardware.

On accuracy, the team’s own assessment is plain: “decoding performance is not yet good enough for everyday use.” A one-in-five error rate at the best-case 78% word accuracy is still high for daily communication, and Meta says more training data is the lever for closing the rest. The detailed scaling argument, including the log-linear curve and the absence of a visible plateau, is laid out at a published analysis of Meta’s non-invasive brain-computer interface research. On hardware, the MEG scanner used in the study is the size of a small room, requires magnetic shielding, and forces the user to sit still.

• Average error rate sits around 39% of words
• MEG helmet is large, magnetically shielded, and stationary
• Decoder currently needs around 10 hours of recorded typing per user
• Decoding is limited to overt typing, not imagined speech
• Real-time end-to-end decoding is new in v2 but not yet bedside-ready

Where Brain2Qwerty Sits in the BCI Race

The non-invasive side of brain-computer interface research includes several efforts running in parallel with the surgical programs. The Decrypt report on Brain2Qwerty v2 names Neuralink and Synchron, both pursuing implanted electrodes, and Merge Labs, backed by OpenAI chief Sam Altman. Merge Labs is also developing technology aimed at restoring communication for people with neurological disorders.

Outside the implant category, Neurable introduced AI-powered EEG headphones in September 2024 designed to monitor focus and cognitive fatigue. A year later, MIT spinout AlterEgo unveiled a wearable that converts silent neuromuscular signals from the face and throat into text and commands, positioning itself as a non-surgical alternative to invasive brain-computer interfaces. The Decrypt report frames both products as part of the broader non-invasive push. Brain2Qwerty v2 has the highest reported real-time sentence accuracy of any non-invasive decoder named in the report. Neurable and AlterEgo target different problems (focus tracking and silent-speech command decoding, respectively). Meta’s release lands inside a category that already has shipped products and active research projects.

Meta’s stack is end-to-end deep learning, an LLM fine-tuned on neural data, and 22,000 sentences of per-person training. The architecture’s progression from v1 to v2 is shown in the table below. Meta says scaling the training data could narrow the remaining gap with surgical neuroprostheses. The MEG hardware has not changed between versions.

What Meta Is Putting in the Open

Meta is committing $5 million to stimulate open datasets as part of its Digital Brain Project. The fund sits alongside Tribev2, a perception-encoding model, and NeuralSet and NeuralBench, two brain-data processing tools Meta lists as sister releases. The stated goal is to build “open foundational models of the brain.”

Brain2Qwerty v2 is the sentence-level decoder that fits into that programme. The paper behind the model is at the published Brain2Qwerty research paper. Meta’s partner BCBL is releasing the v1 dataset on Hugging Face. The v2 dataset stays embargoed until journal publication. BCBL, the Basque Center on Cognition, Brain and Language, is the research lab Meta credits as the data collection partner across both versions.

Frequently Asked Questions

What is Brain2Qwerty v2?

Brain2Qwerty v2 is a non-invasive AI system from Meta that decodes typed sentences directly from magnetoencephalography (MEG) brain recordings. It is the second version of the Brain2Qwerty architecture and the first in the series to generate sentences in real time from continuous brain activity. v1 could only decode by aligning its predictions to the timing of individual keystrokes.

How accurate is Brain2Qwerty v2?

Brain2Qwerty v2 reaches 61% average word accuracy across nine volunteers and 78% for the best participant. Meta reports that the top reader had more than half of all decoded sentences come back with one word error or fewer. Prior non-invasive decoders averaged about 8% word accuracy.

How is Brain2Qwerty v2 different from Neuralink?

Brain2Qwerty v2 is non-invasive and reads brain signals from a helmet-like MEG scanner outside the skull. Neuralink develops surgically implanted interfaces inside the skull. Synchron, also named in the Decrypt report, pursues implanted interfaces that require surgery. The two implant-focused companies and Brain2Qwerty v2 target the same goal from different sides of the skull.

Can Brain2Qwerty v2 be used outside a research lab?

Not yet. Meta has released the full training code for v1 and v2, and BCBL is releasing the v1 dataset, but the system requires a large, magnetically shielded MEG scanner and roughly 10 hours of recorded typing per user. The setup is currently limited to research labs.

What is the next step for the project?

Meta says decoding accuracy rises log-linearly with the volume of training data and shows no detectable plateau yet. The team is collecting more recorded sentences and more volunteer sessions, while MEG hardware makers continue to develop smaller and eventually wearable sensors.

How is the 61% number measured?

Meta reports word accuracy, the share of typed words the model reproduces correctly, rather than full-sentence accuracy. The 61% average is word-level across nine volunteers, and the 78% best-participant score is also word-level. Meta adds that the top reader had more than half of all decoded sentences come back with one word error or fewer. Meta’s own assessment is that “decoding performance is not yet good enough for everyday use.”