# Brain2Qwerty v2 — Meta non-invasive brain-to-text hits 61% word accuracy

> Meta's Brain2Qwerty v2 reads typed sentences off MEG brain scans at 61% word accuracy — without surgery.

- Org: Meta AI Research
- Date: 2026-06-29
- Category: paper, algorithm, repo
- Source: https://ai.meta.com/blog/brain2qwerty-brain-ai-human-communication/
- Canonical page: https://ai-tldr.dev/releases/meta-brain2qwerty-v2/

## In short

Brain2Qwerty v2 decodes typed sentences from MEG brain recordings at 61% word accuracy, up from 8% for prior non-invasive methods. Meta trained the model on 22,000 sentences from 9 volunteers, with code and the paper now public.

## What it is

Brain2Qwerty v2 is Meta AI's second-generation pipeline for decoding what a person is typing directly from non-invasive brain recordings. It pairs a magnetoencephalography (MEG) headset with a deep neural decoder, so the volunteer never has electrodes implanted and never has to speak aloud. The release covers both the trained model and the surrounding research code.

## How it works

Raw MEG signals flow into a convolutional encoder, then a transformer, then a character-level language model that Meta fine-tunes on the neural data so it can use linguistic context to fix noisy character predictions. Compared with Brain2Qwerty v1, the v2 pipeline drops hand-coded event detection in favor of fully end-to-end training and uses LLM-based AI agents to search the configuration space — final hyperparameters were picked by the engineers.

## Why it matters

Until now, only surgical brain-computer interfaces could decode typed sentences at workable accuracy; non-invasive methods sat near 8%. Brain2Qwerty v2 lifts that to 61% average (78% for the best participant), with accuracy improving log-linearly with more recording data. That puts no-surgery brain-to-text on the same trajectory as implanted BCIs, which matters for people who can't undergo neurosurgery.

## Who it's for

neuroscience and BCI researchers, accessibility engineers

## Try it

git clone https://github.com/facebookresearch/brain2qwerty

## Quick facts

- Maker: Meta AI Research (FAIR)
- Version: Brain2Qwerty v2
- Word accuracy: 61% average, 78% best participant
- Prior non-invasive baseline: 8%
- Training data: 22,000 sentences from 9 volunteers (10h MEG each)
- Code: github.com/facebookresearch/brain2qwerty
- Published: Nature Neuroscience + ai.meta.com blog

## FAQ

### How accurate is Brain2Qwerty v2 at decoding sentences?

Brain2Qwerty v2 reaches 61% average word accuracy across nine participants, with the best participant hitting 78% and more than half of that participant's sentences decoded with one word error or less. Meta reports word error rate improves log-linearly with training data, so the gap to surgical implants is expected to narrow as more MEG data is collected.

### Does Brain2Qwerty v2 require brain surgery?

Brain2Qwerty v2 is fully non-invasive: it reads brain activity through a magnetoencephalography (MEG) scanner, the same hospital-grade equipment used in clinical neuroscience, not implanted electrodes. Meta frames this as the first time non-invasive decoding approaches accuracies that had been exclusive to surgical brain-computer interfaces.

### How does Brain2Qwerty v2 work under the hood?

Brain2Qwerty v2 swaps the hand-tuned event detection in v1 for an end-to-end deep network: a convolutional encoder feeds a transformer that is decoded by a character-level language model fine-tuned on neural data. Meta also used AI agents to search the pipeline configuration space, with engineers picking the final setup.

### Where can I get the Brain2Qwerty v2 code and data?

Meta released the full training code for both v1 and v2 on GitHub at facebookresearch/brain2qwerty, alongside a peer-reviewed Nature Neuroscience paper and a research publication page on ai.meta.com. The earlier Spanish MEG/EEG corpus from the Basque Center on Cognition, Brain and Language is on Hugging Face under bcbl190626/SpanishBCBL.

### What's actually new in v2 versus Brain2Qwerty v1?

Brain2Qwerty v1, shared in early 2025, decoded brain activity into text but with a much higher error rate. The v2 paper replaces manual event detection with end-to-end deep learning, fine-tunes a large language model on the neural signal for semantic context, and adds AI-agent-assisted pipeline search — together pushing accuracy from 8% to 61%.

## Links

- [Meta AI Announcement](https://ai.meta.com/blog/brain2qwerty-brain-ai-human-communication/)
- [Research Publication](https://ai.meta.com/research/publications/accurate-decoding-of-natural-sentences-from-non-invasive-brain-recordings/)
- [Code (GitHub)](https://github.com/facebookresearch/brain2qwerty)
- [Brain2Qwerty v1 Blog](https://ai.meta.com/blog/brain-ai-research-human-communication/)