The Large Brain Model

#Introduction

Large language models were trained on what humans write. They were never trained on what humans experience. The LBM is the first foundation model built on the second: 15 years of healthy human neural data, researcher-collected across 4,000+ institutions in 156 countries.

The published EEGM2 architecture outperforms every published EEG foundation model, including models with up to 110M parameters, with linear compute scaling that fits on consumer wearable silicon. It runs at 1–14M parameters, on-device. See arXiv:2502.17873.

#What the LBM is

The LBM is a general-purpose foundation model for cognitive state. It learns the patterns of what the brain does across cognitive load, attention, fatigue, emotion, motor intent and neurological state. It is the industry's first universal EEG foundation model: the intelligence layer every neurotech company can build on. By unifying fragmented neural data into a single system, the LBM does for brain signals what GPT did for language.

Emotiv-FM is the inference layer built on the LBM. It is what devices and partners actually run: an on-device model for latency-critical outputs and a cloud model for richer downstream applications. See Architecture.

#How to read these docs

• The corpus covers the training data, why it is different from every public EEG dataset, and what each property of the corpus produces in the model.
• The model covers capabilities, the two-stage architecture, the inference pipeline, the API surface and the peer-reviewed research.
• Access covers the four partner types, the application categories, what Dcode.AI provides, and the partnership pathway from scoping to production.

#Why now

AGI cannot be reached by scraping the internet alone. The frontier labs have exhausted high-quality text, and synthetic data shows diminishing returns. The race is on for a new modality, and cognitive signal is the only one that requires a physical sensor network to capture.

LLMs learn from text. VLMs learn from images. The emerging class of Biological Large Models learns from the body: physiological signal from wrist-worn sensors, and cognitive signal from the brain. The LBM is the cognitive BLM, the modality that heart rate monitors, sleep trackers and wrist-worn wearables cannot reach, however much data they accumulate. Attention, emotion, fatigue and intent require a different sensor, a different corpus and a different model.

1. The investment wave: record capital is flowing into the category

   More capital has flowed into neurotech over the past five quarters than in any comparable period in the field's history. BCI hardware, neural wearables and brain-computer interfaces are attracting category-defining investment. Every dollar going into device hardware builds demand for the intelligence layer above it. That layer is not yet built.

2. The benchmark: the field has its pre-ImageNet moment

   Meta FAIR published NeuralBench (May 2026): 36 EEG tasks, 14 architectures, 94 datasets. Foundation models already outperform traditional ML across most EEG tasks, but models are still small and datasets narrow. The unlock is scale and corpus diversity. The LBM will be submitted to NeuralBench when full-corpus training is complete.

3. The platform wave: platform health AI creates the market

   Every major AI platform now has a health product built on physiological wearable data. When heart rate, sleep and movement flow through platform AI, the differentiated layer becomes the one they cannot reach: what the brain is doing. The platform health AI wave does not compete with the LBM. It creates the market the LBM licenses into.

#The data fragmentation problem

The EEG neurotech industry is at an inflection point. Investment is at record levels, hardware form factors are maturing, and the first generation of neural wearables is reaching consumers. The constraint is no longer the hardware. It is the data layer beneath it, where every device speaks a different language, every dataset lives in isolation, and every team rebuilds the same preprocessing infrastructure from scratch.

#Inconsistent and noisy data

EEG devices vary wildly in their specifications, from 2-channel consumer wearables to 256-channel research systems. Each uses different sensor placements, input impedances, voltage resolutions and signal-to-noise ratios. This hardware heterogeneity means findings from one device rarely transfer to another, which fragments the knowledge base and limits scalability.

#Task-specific models

Current machine learning approaches in EEG are narrowly focused, with separate models for sleep staging, emotion detection, motor imagery and seizure prediction. Each model starts from scratch, unable to use representations learned from other tasks. This siloed approach wastes resources and prevents breakthroughs across applications.

#Inter- and intra-subject variability

Brain signals are inherently personal. Inter-subject differences such as cortical folding patterns, skull thickness and individual psychological profiles often overwhelm the neural signatures researchers seek to measure. Intra-subject variability from circadian rhythms, hormonal cycles, fatigue and ageing compounds the challenge.

#Stuck on data preparation

Researchers report spending 60–80% of their time on artifact removal, signal filtering and data alignment, tedious work that must be repeated for every new dataset and device. This leaves minimal time for discovery and development, a bottleneck that affects academia and industry alike.

#The Emotiv corpus

Every major EEG dataset was collected from patients in clinical settings, and those that were not suffer from the WEIRD problem: Western, Educated, Industrialised, Rich, Democratic populations only. Neither can train a general foundation model. Emotiv's corpus can.

Dcode.AI is built on the Emotiv corpus, the result of 15 years of sensor network deployment, 4,000+ institutional partnerships, and researcher-collected, event-marked data from day one. Every session is event-marked through EmotivPRO, a single software protocol across every institution and every country.

#At a glance

#Comparison with public corpora

The published EEG foundation models are trained on clinical or public-only corpora. The table lists the three most cited alongside the LBM's combined corpus.

* Emotiv LBM training hours (505,000+) reflect the total corpus including all major public EEG datasets plus 314,000+ hours of proprietary signal. The channel-minutes figure (245M+) reflects the same combined corpus.

#What corpus quality produces

Each property of the corpus maps to a property of the model, and each of those to something the device can do.

"The reason no general-purpose neural foundation model exists is not a model architecture problem. It is a data problem. The corpus required to train one has never existed — until now."Tan Le, CEO & Founder, Emotiv / Dcode.AI

#Capabilities

Five capabilities distinguish the LBM from every other published EEG model.

Cross-device cognitive transfer

Representations learned from one EEG hardware configuration generalise across different sensor layouts, channel counts and wearable form factors with minimal or no retraining. Because the model is trained across Emotiv's full device range, from 2-channel in-ear through 32-channel research-grade, it learns a unified representation of cognitive state that is not tied to any single electrode configuration. A partner's device, whether a 2-channel earbud, a 6-channel headband or a 16-channel AR headset, can pass its signal to Emotiv-FM and receive accurate cognitive state outputs without bespoke retraining.

Longitudinal baseline modelling

The LBM is trained on data from the same individuals observed over time, not just a population snapshot. This enables a qualitatively different kind of inference: not just "what is this person's cognitive state right now" but "how has this person's cognitive baseline shifted over the past six months". Intra-individual drift such as gradual attention decay, sleep-architecture changes and processing speed trends is only visible when the same brain is observed across weeks and months. Cross-sectional models cannot detect it. The LBM can.

Robust real-world inference

Emotiv-FM reconstructs full cognitive-state representations from incomplete or degraded channel sets, using surrounding signal context to fill gaps. The model degrades gracefully rather than failing when electrode contact is imperfect, a property that matters for any device worn in everyday life.

No per-user calibration

Inter- and intra-subject normalisation is learned from 150,000+ subjects across 156 countries. The model works accurately out of the box for a global consumer base, with no onboarding friction and no per-user setup.

On-device at frontier accuracy

The 1–14M parameter architecture runs inference directly on consumer wearable silicon without a cloud round-trip. For applications where latency, privacy or connectivity matter, on-device inference is not a compromise. It is the design intent.

#Architecture

Emotiv-FM is the inference layer built on the Large Brain Model. A two-stage architecture splits responsibility between the device and the cloud: on-device inference for latency-critical outputs, cloud inference for richer downstream applications.

#Choosing a surface

The scoping conversation produces a written recommendation on which surface a partner will consume. See Partnership pathway.

#Inference pipeline

Three stages separate the partner's hardware from the partner's application.

1. Any compatible device streams signal in

   From a 2-channel in-ear earbud to a 32-channel research cap. The model normalises across channel counts, electrode positions and form factors. The partner does not retrain for each new device.

2. Emotiv-FM produces standardised cognitive state outputs

   Device-specific noise, inter-subject variability and artifacts are suppressed. The output is a clean, model-ready representation of cognitive state (attention, load, fatigue, emotional valence, motor intent) calibrated against the full LBM corpus.

3. Partners build the application layer

   Dashboards, classifiers, adaptive interfaces, longitudinal cohort tools: anything that consumes cognitive state. Because the upstream representation is consistent, the application layer sees brain-state differences, not inter-subject differences or hardware noise.

#API surface

The usage-based API is one endpoint with standardised outputs. It is hardware-agnostic: Dcode.AI handles EEG preprocessing, artifact rejection and cross-device normalisation, so no domain expertise is required to get started.

#Standardised outputs

#Request shape

The shape below is illustrative of the endpoint's contract: a device descriptor and a window of signal in, a cognitive state and a confidence score out. The production specification, authentication and SDK bindings ship with API access.

# illustrative
POST /v1/decode
{
  "device": "earbud",        # any compatible form factor
  "channels": 2,
  "window_ms": 4000,
  "signal": [ ... ]
}

# response
{ "state": "attention", "score": 0.92 }

#Published research

The LBM is backed by research published in top-tier academic venues, demonstrating superior performance, generalisability and efficiency. Five foundation-model papers form the de facto standard the field cites.

#Access models

Dcode.AI serves four partner types. The intelligence layer is the same; the access model is calibrated to the partner's hardware, scale and use case.

#AI-powered neuro research

The LBM's training across thousands of experimental paradigms enables predicted neural response distributions for stimuli (packaging designs, content concepts, product formulations) without participant recruitment or hardware deployment. Results arrive in hours rather than weeks, at a fraction of conventional EEG study costs. Built for consumer insights, pharmaceutical research and advertising effectiveness. Enquire.

#Applications

The LBM is a general-purpose foundation model for cognitive state. The categories below are not aspirational. They are the categories of product that the corpus and the model architecture make tractable today.

#What Dcode.AI provides

Partners do not need to build the corpus, the model or the infrastructure. Dcode.AI brings the full stack, and a 15-year head start that cannot be replicated on a product timeline.

#Partnership pathway

A partnership with Dcode.AI is a structured pathway, not a one-off integration. The four stages below are how every serious partner conversation to date has been run.

1. Scoping conversation

   A short technical discussion to map the partner's hardware, target form factor, and the cognitive primitives the product requires. Output: a written scope and a recommendation on which model surface (on-device, cloud or both) the partner will consume.

2. Technical evaluation

   Access to the API and the on-device SDK against the partner's hardware and a representative data sample. Output: measured accuracy on the partner's target tasks, on the partner's silicon, not a benchmark in our environment.

3. Integration and co-development

   A scoped engineering engagement: SDK integration, model tuning where warranted, joint validation on the partner's user population. Co-development is the norm. The LBM gets sharper with every serious integration, and partners get a model calibrated to their deployment.

4. Production and ongoing model access

   Licensed access to the production model, with versioned releases, SLAs, and a model-improvement loop that flows back into the corpus. Partners ship on a foundation model that continues to improve underneath their product.

#Data ownership and privacy

The user owns their data. The architecture is built around on-device inference for privacy-critical workloads and a user-owned data model end to end.

• Partners do not need to take on the regulatory and reputational risk of routing raw neural signal through their own infrastructure.
• End users do not need to surrender their brain data to use the product.
• On-device primitives run with no raw neural data leaving the device; the cloud surface is used for richer downstream tasks.
• Private on-premises deployment is available for platform and enterprise partners with data-sovereignty requirements.

#Brainwear by Emotiv

Brainwear is the consumer product built on the LBM: in-ear EEG that measures cognitive state directly, not inferred from heart rate or movement. Every Brainwear session adds longitudinal real-world data to the training corpus: naturalistic, everyday neural signal in the modality with the least public dataset coverage in the field. brainwear.app

#Team

Neuroscientists, AI researchers and engineers committed to building the future of non-invasive brain decoding.

• Tan LeCEO

  Tan Le is CEO and Founder of Emotiv and Dcode.AI. Over 15 years she has built the largest non-invasive EEG dataset in the world — 14.4 million minutes of training-grade signal across 4,000+ institutions in 140+ countries — and the only corpus on which a true Large Brain Model can be trained. She sits on the World Economic Forum’s Global Future Council on Neurotechnology, has given three TED talks viewed by millions, and is the author of The NeuroGeneration. A former Young Australian of the Year and National Geographic Emerging Explorer, she is widely recognised as a leading voice on the responsible development of neurotechnology.

• Dr Geoff MackellarCTO

  CTO & Co-Founder, Geoff is the scientific architect of the Large Brain Model — responsible for the foundation model architecture, training methodology, and the physical sensor network that made the corpus possible. A PhD physicist with 15+ patents, he designed every generation of Emotiv’s sensor network, from the original EPOC to the MN8 in-ear EEG, building the signal acquisition and processing pipeline that underpins 197.5 million channel-minutes of training signal. His cross-disciplinary ownership — spanning foundation model architecture, signal processing, and applied neuroscience, the full stack from sensor to model output — is what makes the LBM’s corpus properties replicable at scale.

• Patrick ChuVP of Software Engineering

  With over 20 years of experience in neurotechnology, Patrick is a veteran in building and scaling complex tech solutions. As VP of Engineering, he leads the development of Dcode AI’s core platform, translating groundbreaking research into a functional, reliable, and scalable product. His extensive background in the field is instrumental in ensuring Dcode AI’s technology is poised to meet the demands of a rapidly evolving industry.

• Dr Navid FoumaniFoundation Model Architect & Lead

  The architect of Emotiv’s foundation model approach. Navid is lead author of EEG2Rep (KDD ’24 Audience Appreciation Award), the self-supervised representation learning framework that established best-in-class performance across six EEG tasks, and of EEG-X (arXiv:2511.08861), Emotiv’s device-agnostic, noise-robust transformer foundation model that handles varying channel configurations, recording lengths, and electrode combinations across heterogeneous hardware. As a NeurIPS reviewer in the EEG and brain foundation model track in 2026, a cohort that received over 100 submissions this cycle, Navid sits at the centre of the research community defining this category.

• Dr Jiazhen HongAI Research Scientist

  The architect of Emotiv’s edge and long-sequence EEG modeling. Jiazhen is lead author of EEGM2 (arXiv:2502.17873), the Mamba-2 architecture that runs at 4.5M parameters with linear compute scaling, enabling brain model inference on consumer hardware. He is also lead author of SAMBA (arXiv:2511.18571), a self-supervised Mamba-based U-shaped encoder-decoder that addresses one of the hardest open problems in EEG foundation modeling: long-sequence context across heterogeneous electrode configurations. SAMBA introduces Temporal Semantic Random Masking, Multi-Head Differential Mamba for redundancy suppression, and Spatial-Adaptive Input Embedding that generalises across devices, outperforming state-of-the-art across thirteen EEG datasets. Together, EEGM2 and SAMBA define Emotiv’s edge deployment and scalable architecture strategy.

• Dr Nam NguyenDirector of Corpus Curation

  The engineering authority behind the quality of the data the Large Brain Model trains on. Nam owns the curation standards that govern every step between raw EEG collection and the pre-training pipeline; artifact classification protocols, device-specific metadata schemas, and the consistency gates that determine what enters the corpus. His background is in signal processing under conditions where data quality has physical consequences. At Halliburton, he built acoustic signal processing systems for downhole leak detection in oil wells, developed ML classification pipelines for leak characterisation, and designed embedded firmware for drive-by-wire control of tractor tools operating thousands of feet underground. At BlackBerry, he developed audio localisation algorithms prototyped on DSP hardware. Earlier work spanned statistical software development for clinical research and pharmaceutical drug trials; environments where metadata integrity and pipeline consistency are non-negotiable. Twenty-one granted US patents. EEG curation at the scale of 197.5M channel-minutes across heterogeneous hardware requires exactly this discipline: systematic artifact classification, rigorous schema enforcement, and pipeline rigour built from years of signal processing where failure is not recoverable.

• Dr Nikolas WilliamsDirector of Neuroscience & Research

  EEG neuroscience and cognitive science authority with 15 years of academic and applied research experience. With a background and peer-reviewed publications across the areas of attention and memory, neurodegenerative disease, psychiatric neuroscience, and hardware validation, he leads the corpus curation and research methodology function at Emotiv and defines the protocols that define what the LBM trains on, how sessions are structured across institutions and devices, and what scientific standards the dataset must meet. As the senior scientific voice on neuroscience methodology, he is the bridge between Emotiv’s neuroscience rigor and its model development program, complementing Dr Geoff Mackellar’s signal-processing authority and Dr Navid Foumani’s foundation-model architecture work.

• Cuong NguyenSenior ML Ops Engineer

  Cuong ensures that Dcode AI’s powerful research models are ready for the real world. As a Senior ML Ops Engineer, he is an expert in bringing machine learning models from the lab into production. His work in building scalable and reliable systems for data management, model deployment, and monitoring is fundamental to Dcode AI’s product roadmap, making the universal foundation model accessible to developers and researchers globally.

• Quoc NguyenML Ops Engineer

  Quoc ensures that Dcode AI’s powerful research models are ready for the real world. As an ML Ops Engineer, his expertise is in bringing machine learning models from the lab into production. He is skilled in designing and developing scalable and reliable systems for model deployment, monitoring, and performance tracking, using tools like MetaFlow, VertexAI and TensorBoard. Quoc’s work is fundamental to Dcode AI’s mission of building a scalable and efficient foundation model, making it accessible to developers and researchers globally.

#Glossary

LBM

Large Brain Model. The foundation model trained on the Emotiv corpus.

Emotiv-FM

The inference layer built on the LBM: the on-device Mamba-2 U-Net and the cloud EEGX transformer.

BLM

Biological Large Model. The emerging class of models that learns from the body rather than from text or images. The LBM is the cognitive BLM.

EEG

Electroencephalography: electrical activity of the brain recorded from the scalp or the ear canal.

Channel-minutes

Minutes of recording multiplied by the number of channels recorded. The corpus holds 245M+.

Event-marked

Signal annotated with the stimulus or task event at the moment it occurred, which lets the model learn cognitive transitions rather than static states.

EmotivPRO

The single software protocol through which every session in the corpus was recorded and event-marked, across every institution and country.

WEIRD

Western, Educated, Industrialised, Rich, Democratic: the narrow population that most public datasets are drawn from.

10-05 / 10-10 / 10-20

Standard systems for electrode placement on the scalp. EEG-X embeds channels by their location in these systems.

Mamba-2

A state-space model architecture with linear compute scaling over sequence length, used on-device and in EEGM2 and SAMBA.

JEPA

Joint-Embedding Predictive Architecture, the teacher–student design EEG2Rep redesigned for EEG.

P300

An event-related potential used by speller BCIs; SpellerSSL aggregates it for calibration-light spelling.

NeuralBench

Meta FAIR's EEG benchmark (May 2026): 36 tasks, 14 architectures, 94 datasets.

#Get access

Be among the first to build on the Large Brain Model. Request API access for updates on the platform launch.

Or email api@dcode.ai.