Frontier world model

MOTHER EXO

A frontier world model — not text tokens, but Eyes, Ears, a Mouth, a Brain, memory and understanding, all at once.

MOTHER EXO is our answer to a frontier model for the physical world. Where a language model predicts the next word, EXO perceives, understands and acts across every modality simultaneously — seeing, hearing, speaking, reasoning, remembering, and reading the emotion in a scene. Built on a frozen CORE-7B mind and a frozen MOTHER DeepVision eye, with dozens of task heads trained on real on-node data, EXO is one model that does everything for us at once: a true world model, observe-and-advise, human-in-the-loop, non-kinetic by design.

14
trained capability heads
4 bodies
humanoid · drone · vehicle · defence
~7M
owned per-head training records
Strike 0
non-kinetic · HITL

Eyes, Ears, a Mouth, a Brain — at once

One model with every sense, running simultaneously — not a pipeline of tools, but a single world model.

Eyes
See

MOTHER DeepVision ViT — detects, tracks and segments objects, faces and relationships in real time across camera, drone and satellite feeds.

Ears
Hear

Audio and speech understanding — transcribes, diarises and interprets what's said and the sounds of a scene.

Mouth
Speak

Natural speech and expression — EXO responds in voice and communicates state, not just text.

Brain
Reason

A frozen CORE-7B mind reasons over everything it senses, plans, and grounds decisions in retrieved knowledge.

Memory
Remember

Persistent memory and self-learning heads let EXO accumulate context over time and improve from experience.

Understanding
Relate

Scene-graph and relations heads build a structured understanding of who is doing what to whom, and why.

Emotion
Feel

Expression and affect heads read emotion-based patterns — the mood and intent behind behaviour, not just the pixels.

Action
Act

Flight, humanoid and driving heads turn understanding into advised action — always observe-and-advise, human-in-the-loop.

Capabilities

One model, every modality — at once

EXO fuses vision, audio, language and action in a single world model. It doesn't switch between tools; it perceives and reasons simultaneously.

A living world model

EXO maintains an internal model of the scene — objects, people, relationships, motion and intent — and predicts how it will evolve.

Self-learning

Head-only training on real on-node corpora lets EXO keep learning new capabilities without retraining the frozen mind and eye.

Emotion-aware

Expression and affect heads surface emotion-based patterns, so EXO understands the human meaning of a scene.

Grounded reasoning

The CORE-7B backbone reasons over what EXO senses and grounds decisions in retrieved knowledge, with citations and audit.

Observe-and-advise

Across flight, driving and humanoid domains EXO advises a human operator — it never selects targets or takes kinetic action.

Architecture

  1. 1

    Frozen CORE-7B ⊕ frozen DeepVision ViT

    EXO shares MOTHER CORE as its reasoning mind and MOTHER DeepVision as its visual cortex — both frozen — so the expensive backbones are trained once and reused.

  2. 2

    Head-only training

    Dozens of lightweight task heads (detect, track, worldmodel, face, relations, memory, speech, vision, audio, expression, flight, humanoid, driving) are trained on real on-node data, so new senses are added cheaply and safely.

  3. 3

    Real-data evaluation

    Every head is evaluated on real video and images — detect · track · relations · face · assess · worldmodel · expression — plus reasoning and refusal, with measured scores on the /gb10/exo dashboard.

  4. 4

    Sovereign GB10 node

    EXO trains and infers on a sovereign GB10 Blackwell node; telemetry streams live to the operations dashboard.

Specifications

CompositionFrozen CORE-7B ⊕ frozen DeepVision ViT ⊕ 14 trained heads
Bodieshumanoid · drone (Strike 0) · vehicle · defence
Capability headsdetect · track · worldmodel · face · relations · memory · speech · vision · audio · expression · flight · humanoid · driving
Detection dataCOCO + LVIS + aircraft — 682k
Face dataMS-Celeb / CelebA / LFW — 5.4M crops
Teleop dataOpen-X + GR00T — 150k real steps
Audio / reasoningLibriSpeech + RAVDESS/CREMA-D 6.8k+ · sovereign QA 861k
Servingload_exo.py · exo_node :8000 · humanoid_act / flight_act / drive
SafetyObserve-and-advise · HITL · Strike = 0

The 14 trained weights — function · dataset

WeightKindFunctionDatasetTrained
reasoningtextSituational-awareness QA · threat classification · summaries · relationship & decision supportmother_core_v31 · golden_datasets · military/space converted · reasoning CoT/emotion
detectvision1,233-class presence + per-class boxes, parallel box decodingCOCO + LVIS + RarePlanes + military — 682k
trackvisionAppearance embeddings — keep object identity across frames (contrastive)perception frames · stage_F temporal · cosmos_reason1
worldmodelworldPredict next-step latent / plausible futures (contrastive predictive)on-node video + latent rollouts
facevisionFace representation + consent-only identityMS-Celeb / CelebA / LFW — 5.4M crops (opt-in gallery)
relationsscene-graphSpatial relations — on / next_to / under classificationscene-graph annotations
memorymemoryMOTHERrag Semantic-Pyramid long-term memory — write & recallMOTHERrag corpus — 800k
speechspeech20-language ID + Text-to-Voice; greets known faces by nameLibriSpeech + neural-codec
visionvisionSelf-supervised visual foundation shared by the vision headsSigLIP-SO400M · SimCLR over all on-node imagery
audioaudioASR — radio signals + human speech → text (feeds reasoning)LibriSpeech + RAVDESS / CREMA-D — 6.8k+
expressionvisionReads affect — happy/sad/angry/worried/calm (NOT identity)RAVDESS / CREMA-D / affect crops
flightautonomyISR velocity policy in the 6-DOF sim — observe-and-advise, Strike 0sim rollouts (6-DOF ISR)
humanoidautonomyEmbodiment motor policy behaviour-cloned from REAL teleop; HITLOpen-X + GR00T — 150k real teleop steps
drivingautonomyPerception → planning → control policy (first-pass); observe-and-advisedriving corpus (first-pass)

Model evaluation — industry benchmarks

BenchmarkMetricResultStatus
Detection — COCOmAP50○ to run
Tracking — MOT17MOTA / IDF1○ to run
Face — LFW / IJB-Caccuracy○ to run
ASR — LibriSpeechWER○ to run
Manipulationtask success○ to run
Locomotionfall-rate○ to run
World-modelrollout FID○ to run

Benchmarks scheduled on this build — measured results are recorded here as each run completes.

What it's for

Common operating picture

EXO fuses live camera, drone and sensor feeds into a single understood scene — objects, people, relationships and intent — for decision support.

Autonomous ISR (advise)

Supervises multi-UAV observation, mapping flight paths and surfacing what matters, while a human authorises every action.

Embodied assistance

Drives humanoid and mobile platforms in observe-and-advise mode, reading the environment and proposing safe actions.

World understanding

A general perception-and-reasoning engine any MOTHER product can call to understand what is happening in the physical world.

Safety & governance

  • Strike = 0 across every domain — flight, driving and humanoid. EXO never selects targets, controls fire, or completes a kill-chain.
  • Observe-and-advise only: EXO informs a human decision; a human acts.
  • Human-in-the-loop is mandatory for any outward or physical action.

Frequently asked

What is MOTHER EXO?

MOTHER EXO is a frontier world model that perceives, understands and advises across vision, audio, speech, language and action simultaneously. It is built on a frozen CORE-7B reasoning mind and a frozen MOTHER DeepVision visual cortex, with dozens of task heads trained on real on-node data.

How is a world model different from a language model?

A language model predicts the next text token. A world model like EXO builds and maintains an internal model of the physical scene — objects, people, relationships, motion, emotion and intent — and reasons and advises across every modality at once.

Is MOTHER EXO a weapon?

No. EXO is observe-and-advise and non-kinetic by design, with Strike = 0 across flight, driving and humanoid domains. It supports human decisions and never selects targets or takes kinetic action.

What can MOTHER EXO do?

See, hear, speak, reason, remember, understand relationships, read emotion-based patterns, and advise action across flight, driving and humanoid platforms — all from one model, at once.

How does EXO keep learning?

EXO uses head-only training on real on-node corpora, so new capabilities are added by training lightweight heads without retraining the frozen mind and eye.

The MOTHER model family

Build on MOTHER EXO

Sovereign, on-node and observe-and-advise by design. Talk to us about access and deployment.