Meta AI has unveiled Brain2Qwerty v2, an advanced AI system that decodes brain activity into text without requiring surgical implants. Building on last year's v1, the new version achieves a word accuracy rate of 61%, a dramatic improvement over the 8% accuracy of other non-invasive methods. For the best-performing participant, accuracy reached 78%, with over half of all sentences decoded with one or fewer word errors.
The system uses magnetoencephalography (MEG) to record brain signals while participants type. Meta trained the model on approximately 22,000 sentences from nine volunteers, each recorded for 10 hours. Unlike traditional pipelines that rely on hand-crafted neural event detection, Brain2Qwerty v2 employs end-to-end deep learning to decode directly from raw brain signals.
Fine-tuning large language models on neural data enables the system to leverage semantic context, bridging the gap between noisy brain recordings and coherent language. AI agents were also deployed to explore optimizations for the decoding pipeline, with final configurations selected manually by engineers.
Meta is releasing the full training code for both v1 and v2, and its partner, the Basque Center on Cognition, Brain, and Language (BCBL), is releasing the v1 dataset. The company hopes this open approach will accelerate neuroscience breakthroughs and help millions of people with brain lesions who cannot communicate.
The research suggests that decoding accuracy improves log-linearly with data volume, indicating that further scaling could narrow the gap with surgical approaches. This work is part of Meta's broader effort to build open foundational models of the brain, including Tribev2 for perception encoding, NeuralSet for processing brain data at scale, and NeuralBench for systematic evaluation.