Meta AI has released Brain2Qwerty v2, a brain-to-text decoding system that reconstructs typed sentences from non-invasive magnetoencephalography (MEG) recordings with a 61% average word accuracy, representing a significant leap over the 8% ceiling of prior non-invasive methods. The system, an upgrade to the February 2025 Brain2Qwerty v1, decodes natural language in real time without requiring surgery or implants, and the full training code for both versions is now available on GitHub under a CC BY-NC 4.0 license.
What Makes Brain2Qwerty v2 Different
Brain2Qwerty v2 maps raw brain activity directly to characters, then to words and sentences, using an end-to-end deep learning pipeline. Meta trained the model on approximately 22,000 sentences from nine volunteer participants, each recorded for 10 hours while actively typing inside a MEG device. MEG measures the magnetic fields produced by neuronal activity at high temporal resolution, offering a cleaner signal than the electroencephalography (EEG) used in earlier attempts.
The critical breakthrough is the replacement of hand-crafted neural event detection with learned representations. The pipeline combines three core components: a convolutional encoder that reads raw MEG signals, a transformer that models longer-range temporal structure, and a character-level language model that constrains output toward plausible text. Fine-tuned large language models add semantic context, bridging noisy brain recordings and coherent language output.
This is research published for scientific and engineering use, not a consumer product. The decoder was tested on a small group of healthy volunteers in a controlled setting, and the data belongs to Spain’s BCBL (Basque Center on Cognition, Brain and Language), where the recordings were collected.
The Accuracy Numbers and What They Mean
Brain2Qwerty v2 achieves an average word accuracy of 61%, corresponding to a word error rate (WER) of 39%. For the best-performing participant, accuracy reaches 78%, with over half of all decoded sentences containing one word error or fewer. These figures are measured against the ground truth of what the participant actually typed.
For context, Meta reports that earlier non-invasive methods hovered around 8% word accuracy. The improvement is not incremental; it is an order-of-magnitude leap enabled by deep learning replacing fragile signal-processing pipelines. Accuracy scales log-linearly with the amount of training data, a finding with direct implications for any team building biosignal decoders. More recording hours predictably raise accuracy, suggesting that the gap with surgical implants may narrow through data volume alone.
How the Decoding Pipeline Works in Practice
The pipeline replaces the brittle, hand-crafted event detectors of earlier systems with end-to-end learning from raw MEG signals. The convolutional encoder learns features directly from the data. The transformer then models relationships across the entire signal. The character-level language model rejects sequences that form no real words, pushing the decoder toward sentences a human would plausibly type.
Meta describes three concrete engineering decisions that make this possible: deep learning replaces hand-crafted event detection; large language models are fine-tuned on neural data to extract semantic representations; and AI agents iteratively refined the decoding pipeline through automated code development, though final training configurations remained human-selected.
The published architecture sketch reflects these components: a convolutional encoder over raw MEG channels feeds into a transformer, which outputs to a character-level head. A downstream language model then refines the output. Working developers can clone the repository and inspect both v1 and v2 code directly.
Brain2Qwerty v1 vs v2: What Changed
Brain2Qwerty v1, released in February 2025, was measured at the character level and reported up to 80% character accuracy using MEG. Version 2 shifts to word-level evaluation, a more meaningful metric for practical communication. The new system operates in real time, decodes at the character, word, and sentence level, and relies exclusively on MEG recordings from a smaller cohort of nine participants who each contributed more data.
V1 also demonstrated that MEG decoding was at least twice as accurate as EEG, establishing the signal modality as the preferred path for non-invasive brain-to-text work.
Why Non-Invasive Brain Decoding Matters
The primary motivation is restoring communication for people with brain lesions that prevent speaking or moving. Invasive methods using stereotactic electroencephalography or electrocorticography already feed neural signals to AI decoders, but they require neurosurgery and are difficult to scale. A non-invasive decoder could widen access, allowing a patient to type sentences using only external recordings.
For researchers, the released code supports reproducible neuroscience—any lab with a MEG device and ethics approval can retrain the pipeline on its own dataset. For AI engineers, the convolutional-encoder-plus-transformer pattern transfers to other biosignal decoding tasks such as EEG-based motor imagery or sleep stage classification. For data scientists, the log-linear scaling result provides a planning tool: it quantifies how much new recording data may lift accuracy for a given application.
Interactive Pipeline Demo
An interactive simulation embedded in the original announcement illustrates the pipeline in action. Users can select a sentence, toggle between MEG and EEG recording modes, adjust training data volume, and watch as the model passes through four stages: raw signal acquisition, convolutional encoding, transformer processing, and language model correction. The simulation renders noisy character-level predictions, final decoded output, and live accuracy metrics that reflect the reported 61% average and 78% best-participant figures. This widget is a demonstration, not a real model inference, but it faithfully represents the architecture and scaling behavior described in the paper.
What the Release Code Includes
Meta has published the full training code for both Brain2Qwerty v1 and v2 on GitHub. The repository includes configuration files, data preprocessing scripts, model definitions, and training loops. Developers can clone the repository, inspect the v1 and v2 subdirectories, and understand how the convolutional encoder, transformer, and character-level language model are assembled. The code is released under the CC BY-NC 4.0 license, permitting academic and non-commercial use.
Who Should Pay Attention to This
This development lands at the intersection of neuroscience, AI, and accessible assistive technology. For research labs working on brain-computer interfaces, the code and methodology provide a reproducible baseline that outperforms prior non-invasive systems by a wide margin. For AI engineers interested in biosignal processing, the pipeline architecture offers a proven template. For anyone tracking the progress of neural decoding, the log-linear scaling result is a key datapoint—it frames how close non-invasive methods may be to clinical relevance.
The decoder is not ready for consumer use, clinical deployment, or unassisted communication by patients. But the trajectory is clear: deep learning, combined with high-quality MEG data, has moved non-invasive brain-to-text from a proof-of-concept curiosity to a quantitatively serious approach. Teams that invest in building MEG datasets and adapting this pipeline to their own signal modalities are positioned at the leading edge of a field that may reshape assistive communication within the decade.
