Meta Brain2Qwerty v2: 61% Word Accuracy

What Is Meta Brain2Qwerty v2?

Brain2Qwerty v2 is a non-invasive brain-to-text decoder. It reads raw neural signals and outputs full sentences — no surgery required. Meta announced it on June 29, 2026. It builds on Brain2Qwerty v1, which was published the same day in Nature Neuroscience. (source)

The system uses magnetoencephalography (MEG). MEG measures the magnetic fields made by electrical currents in the brain. It produces clearer signals than electroencephalography (EEG), which reads electrical activity from the scalp. Brain2Qwerty v1 reached a character error rate of 29% with MEG. With EEG, that rate rose to 65%.

We think the clearest way to frame this: MEG is roughly twice as accurate as EEG for this task, based on the published v1 results.

How Was v2 Trained?

Meta trained v2 on about 22,000 sentences. Nine volunteers each wore an MEG device and spent around ten hours typing sentences. The system captured their brain activity the whole time.

The pipeline combines two things:

• End-to-end deep learning applied directly to raw MEG signals
• Fine-tuned large language models that turn noisy neural data into coherent text

The v1 study used 35 healthy volunteers. Researchers captured 1,000 snapshots of brain activity per second. That resolution let them track the exact moment thoughts became letters and words. (source)

What Accuracy Does Brain2Qwerty v2 Achieve?

Here are the key results from Meta's published data:

You might also like

Jeremy Strong Plays Zuckerberg in Social

JetBlue Drone Strike at JFK: What Happened

Sergey Brin Lost Big on NYC Rent-Stabilized Flats

SpaceX Deploys Dozens of Engineers to Grok AI

V1 focused on character-level decoding. V2 moves to full words and sentence-level meaning. Meta describes this shift as "enabling accuracy for overall communication." (source)

What Did the Research Reveal About the Brain?

The v1 study gave researchers a close look at how the brain builds language. MEG's high time resolution made this possible.

One researcher described it this way: "The neural activity preceding the production of each word is marked by the sequential rise and fall of context-, word-, syllable-, and letter-level representations."

In plain terms: the brain first processes a word's context. Then its meaning. Then its syllables. Then its individual letters — in that order.

The system also picked up motor signals tied to typing. Errors often involved letters that sit close together on a QWERTY keyboard. The AI might confuse "k" and "l," just as a typist might. The system tracked both cognitive intent and physical motor commands at the same time.

Who Is This Meant to Help?

Meta says the research targets people who have lost the ability to speak or move. The Nature Neuroscience paper names specific conditions: anarthria, amyotrophic lateral sclerosis, and severe paralysis. (source)

Current communication neuroprostheses require brain implants. Those carry real risks:

• Infection
• Brain hemorrhage
• Brain damage
• Device degradation over time

Brain2Qwerty avoids all of those. The Nature Neuroscience paper says the results "narrow the gap between invasive and noninvasive methods."

No product launch has been announced. Meta describes this as a research milestone only. The broader push toward AI model access across the industry in 2026 gives context for why open-release research like this matters.

What Code and Data Is Meta Releasing?

Meta released the full training code for both v1 and v2. Its partner, BCBL, released the v1 dataset. Researchers can use both to replicate and extend the work.

This open approach mirrors how other labs have handled AI licensing expansion — sharing artifacts to speed up third-party research rather than keeping them proprietary.

What Are the Current Limits?

The system is not ready for clinical use. Key gaps remain:

• No published data on exact latency
• No confirmed path to EEG viability at scale
• MEG machines are large, expensive, and lab-bound — not wearable
• The v1 study used healthy volunteers, not patients with communication impairments

The gap between 61% average word accuracy and reliable communication for a non-speaking patient is still wide. The top participant's 78% accuracy shows what's possible. But that's the best case, not the norm.

Teams building OpenAI hardware and other human-computer interaction tools are watching non-invasive BCI research closely. Labs funding AI expansion have also begun backing neuroscience-adjacent work.

Meta's confirmed next step is already done: the training code and v1 dataset are live, and the v1 paper is published in Nature Neuroscience as of June 29, 2026.