Meta Brain2Qwerty Explained: Brain-to-Text Without Surgery

Brain2Qwerty is Meta's research system for turning brain activity, recorded from outside the head, into the sentence a person is typing. It first appeared as a preprint in February 2025. The peer-reviewed version was published in Nature Neuroscience in June 2026, and a second version, released the same month, uses a large language model to decode whole sentences with a word error rate of 39%. This article explains what the system measures, how the decoder works, what the numbers mean, and what it cannot do yet.

What Brain2Qwerty actually does

The name gives the task away. Participants sit in a brain scanner and type sentences on a keyboard. The system never sees the keyboard: it sees only the brain signals, and it has to work out which keys were pressed.

That is a narrower task than "reading thoughts", and a deliberate one. Typing produces a clear, repeatable pattern of activity in the parts of the brain that plan and control movement, and each keystroke gives the researchers an exact label to train against. It is the non-invasive counterpart of the implanted systems that decode attempted speech or handwriting in people with paralysis, which we cover in how speech brain-computer interfaces work.

The work comes from Meta's FAIR lab in Paris, led by Jean-Rémi King, with the Basque Center on Cognition, Brain and Language (BCBL) in San Sebastián, where the recordings were made.

MEG and EEG in plain words

Brain2Qwerty was tested with two ways of recording the brain without surgery.

• EEG (electroencephalography) uses electrodes on the scalp to pick up the tiny voltages produced when many neurons fire together. It is cheap, portable and the basis of every consumer "brain" headset. The skull smears the signal, so it is noisy and blurry.
• MEG (magnetoencephalography) measures the magnetic fields produced by the same currents. Magnetic fields pass through the skull almost undistorted, so MEG sees much sharper activity with millisecond timing. The catch is the machine: a helmet of 306 supercooled sensors inside a magnetically shielded room, costing millions and requiring the person to sit still.

That difference turns out to matter more than any detail of the model.

Version 1: the 2025 preprint and the 2026 paper

In the first study, 35 healthy volunteers at BCBL saw short Spanish sentences, five to eight words long, flashed one word at a time. They memorised each sentence and then typed it without seeing the screen. Twenty did the task in MEG, twenty in EEG (five did both), with under an hour of recording each per method.

The decoder has three stages:

1. A convolutional network reads a half-second window of brain signal around each keystroke. It has a spatial-attention layer that learns which sensors matter and a small subject-specific layer that absorbs the differences between people's heads.
2. A transformer looks across the whole sentence, so the guess for one letter can use the letters around it, and outputs a probability for each of 29 characters.
3. A character-level language model, trained on Spanish Wikipedia, rescores those guesses with a beam search, the same technique speech recognisers use to prefer "recognise speech" over "wreck a nice beach".

Results from the published paper: with MEG, the character error rate (CER) averaged 29%, and the best participant reached 18%. With EEG it was 65%. In the February 2025 preprint the figures were 32% and 67%. Put differently, MEG recovered roughly seven characters in ten on average, and some sentences outside the training set were decoded perfectly. EEG recovered about one in three.

Version 2: whole sentences with a large language model

The second paper, released by Meta on 29 June 2026, changes three things.

• More data per person. Nine volunteers each spent about ten hours in the MEG scanner, typing roughly 22,000 English sentences in total that they had heard a few seconds earlier.
• No keystroke timing needed. Version 1 had to know when each key was pressed to cut the signal into windows. Version 2 reads the continuous signal and is trained with CTC (connectionist temporal classification), the same loss used to train speech recognisers that do not know in advance where each sound starts. It predicts the keystrokes and where the spaces fall on its own.
• A language model that reads the brain signal. The predicted text, plus a compressed "brain token" for each word, is fed to a Qwen3 language model (0.6 to 4 billion parameters, fine-tuned with LoRA), which writes the final sentence.

The headline number is a word error rate (WER) of 39% on average, or about 61% of words correct, and 78% for the best participant. For that person, 28% of test sentences came out perfectly and 47% had at most one word wrong. The simpler pipelines in the same paper did worse: the brain encoder alone reached 55% WER, and with a classic word model 43%.

Two details in the paper are worth knowing because they address the obvious suspicion, that the language model is just inventing plausible English:

• When the researchers removed the brain tokens and left only the decoded text, WER rose from 39% to 49%. The model is using the brain signal, not only guessing.
• At the character level, the LLM version is slightly worse than the encoder alone (31% versus 28% CER), because it sometimes rewrites a sentence into something fluent but different. That is the typical failure of any language-model decoder, and the same trade-off exists in speech recognition.

The paper also reports that error keeps falling steadily as training data grows, with no sign of levelling off at about 90 hours pooled across participants. More data is likely to mean better decoding.

What it cannot do yet

The limits are spelled out in the papers themselves, and they matter more than the headline.

• It is not real-time. Both versions work on complete sentences. Version 2's language model needs the whole sentence before it writes anything, so the authors describe its latency as "necessarily slow".
• Participants were typing. These were healthy people pressing real keys. The brain maps in version 2 show the signal comes mostly from motor cortex. Whether it works for someone who cannot move, and so cannot generate labelled training data by typing, is untested.
• The good results need MEG. EEG, the only technology that is wearable today, roughly doubled the error rate in the direct comparison. MEG needs a shielded room. The authors' hope rests on newer wearable MEG sensors (optically pumped magnetometers), and they show the decoder still works with a quarter to half of the sensors.
• People differ a lot. In version 2, the best and worst participants were far apart; the worst had only 4% of sentences fully correct.
• Implants are still far ahead. The version 2 paper itself notes that implanted typing BCIs reach word error rates below 2%.

Why it matters anyway

Brain2Qwerty moves non-invasive brain-to-text from "picking the right word out of a list" to open sentences with a majority of words correct. It also shows a recipe that transfers directly from speech recognition: an acoustic-style encoder trained with CTC, a language model to fix spelling, and an LLM to produce fluent text. If wearable MEG matures, the same recipe could one day serve people who cannot speak or type and do not want surgery.

Meta has released the code for both versions under a non-commercial licence, and the version 1 Spanish dataset is public on Hugging Face. The version 2 dataset is held back until the paper is accepted. Alongside the first paper, Meta donated $2.2 million to the Rothschild Foundation Hospital in Paris for this line of work.

For the broader picture, including fMRI decoders and why Meta's EMG wristband is not mind reading, see can AI read your thoughts. For the implanted systems that already let people with ALS talk, see brain-to-text in 2026.

Frequently asked questions

What is Meta Brain2Qwerty?

Brain2Qwerty is a research system from Meta FAIR and the Basque Center on Cognition, Brain and Language that decodes the sentence a person is typing from brain signals recorded outside the head with MEG or EEG. Version 1 was published in Nature Neuroscience in June 2026; version 2, released the same month, adds a large language model and decodes English sentences with a 39% word error rate.

How accurate is Brain2Qwerty?

With MEG, version 1 has a 29% average character error rate and 18% for the best participant; with EEG it is 65%. Version 2 has a 39% average word error rate, about 61% of words correct, and 78% of words correct for the best participant, who got 28% of sentences exactly right.

Can Brain2Qwerty read thoughts?

No. It decodes the brain activity that comes with typing on a real keyboard, mostly from motor areas, in healthy volunteers who cooperate and spend hours in the scanner to train it. It has not been shown to decode imagined words, inner speech, or thoughts a person does not intend to type.

Is Brain2Qwerty a product I can use?

No. It is research. The best results need a MEG scanner inside a magnetically shielded room, the decoder is trained per person, and it only works on complete sentences, not in real time. Meta has released the code for research under a non-commercial licence.

How does it compare with Neuralink and other brain implants?

Implanted speech and typing BCIs are far more accurate today, with word error rates in the low single digits for the best systems, and they work in real time for people with paralysis. They require brain surgery. Brain2Qwerty's value is that it needs no surgery at all.

Sources

• Lévy et al., "Noninvasive decoding of typed sentences from human brain activity", Nature Neuroscience (2026)
• Lévy et al., "Brain-to-Text Decoding: A Non-invasive Approach via Typing", arXiv:2502.17480 (2025)
• Zhang, Lévy et al., "Accurate Decoding of Natural Sentences from Non-Invasive Brain Recordings" (Brain2Qwerty v2, 2026)
• Meta AI blog: Brain2Qwerty v2
• Meta AI blog: Using AI to decode language from the brain (February 2025)
• Brain2Qwerty code on GitHub
• Zhang et al., "From Thought to Action: How a Hierarchy of Neural Dynamics Supports Language Production", arXiv:2502.07429 (2025)