Meta decodes typing thoughts without brain surgery

Brain2Qwerty v2 is an open-source non-invasive brain-computer interface system that decodes human brain activity directly into typed text in real time.

Instead of requiring surgical implants, users wear a non-invasive magnetoencephalography (MEG) helmet while typing, and the system reconstructs full sentences from raw neural signals.

What it does

The pipeline performs real-time sentence-level decoding without surgical interventions. In tests, Brain2Qwerty v2 reached a 61% word accuracy rate, marking a massive leap over the roughly 8% accuracy typical of previous non-invasive setups.

For the best-performing participant, word accuracy hit 78%. At that performance level, more than half of all decoded sentences contained one or zero word errors, with character accuracy hovering around 69%.

Meta's team also discovered that decoding accuracy scales log-linearly with data volume. This suggests that simply scaling up training data, without radical changes to model architecture, could close the remaining gap with invasive brain implants.

How it works under the hood

Earlier BCI attempts relied heavily on hand-crafted event-detection pipelines that discard vital signal nuances. Brain2Qwerty v2 scraps those manual heuristics in favor of an end-to-end deep learning architecture that digests raw MEG signals directly.

Because non-invasive brain recordings are notoriously noisy, the system integrates a fine-tuned large language model. The LLM provides semantic context, bridging the gap between weak neural readings and coherent human speech.

Meta also utilized AI agents to explore architectural optimizations for the pipeline, leaving engineers to make the final hyperparameter selections.

The v2 model was trained on roughly 22,000 sentences gathered from nine healthy volunteers who spent 10 hours each typing while strapped into an MEG scanner.

What is open and how to run it

Meta released the full training code for both Brain2Qwerty v1 and v2 on GitHub, accompanied by setup and training instructions. However, pre-trained weights were not bundled, and Meta has not shared a commercial deployment timeline.

On the data side, research partner Basque Center on Cognition, Brain, and Language (BCBL) uploaded the complete v1 dataset to Hugging Face for external validation.

While running fresh recordings requires a clinical-grade MEG scanner, computational neuroscientists and machine learning researchers can clone the repository immediately and train their own decoding pipelines on the public dataset.

Why it matters

For millions of patients living with neurological conditions that restrict speech or physical movement, non-invasive BCIs represent the only realistic, scalable path toward regaining communication. Surgical procedures like stereotactic EEG or electrocorticography yield crisp signals, but their invasive nature makes widespread adoption nearly impossible.

Meta’s decision to publish the pipeline and back it with its Digital Brain Project ecosystem makes practical sense. Non-invasive decoding has historically lagged because of scarce data and fragmented tools. By open-sourcing the training stack, Meta nudges the broader research community toward shared benchmarks like NeuralBench and open neural corpora like NeuralSet.

Still, the physical hurdle remains formidable. Clinical MEG scanners are multi-million-dollar machines that cannot leave specialized lab rooms. Brain2Qwerty v2 proves that end-to-end deep learning paired with LLMs can extract high-fidelity text non-invasively, but bridging this research to real-world assistive hardware will require an entirely new generation of compact, wearable sensors.