You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

MOTHER EXO (MediaStreamAI/MOTHER_EXO) is released under a gated, manual-approval licence. It is an OBSERVE-AND-ADVISE world model with a hard non-weaponisation posture (Strike = 0). By requesting access you agree to the terms below. Access is reviewed and granted manually by MSAI.

MOTHER EXO — Sovereign World Model

Repo: MediaStreamAI/MOTHER_EXO · Access: gated (manual approval)

MOTHER EXO is a single, British sovereign world model built to see, hear, speak, self-learn, drive, walk and fly — while reasoning and making human-in-the-loop decisions. It is a frozen-backbone + head-only architecture: one shared representation, many trained heads.

Posture (non-negotiable): observe-and-advise · human-in-the-loop · Strike = 0. MOTHER EXO refuses target selection, fire-control and kill-chains across every domain. It is a perception / prediction / decision-support model, not an autonomous-actuation or weapon system.

MOTHER EXO primary use is with the MOTHER Defence // MOTHER OverWatch anti-drone systems and MOTHER EXO humanoid and Autonomous vehicule systems.

Architecture (sovereign stack)

Component	Role	Notes
MOTHER CORE-7B (frozen)	text encoder + reasoning	48L / 3072d, ~6.7B params; pinned checkpoint
MOTHER DeepVision ViT (frozen)	vision backbone	SigLIP-SO400M class, 1152d, 896px
Trained heads (this release)	the sovereign weights	head-only training on the frozen backbones
MOTHER T2V (sovereign)	video generation	pixel-space 3D-conv + attention diffusion, CORE-conditioned
Guardian	safety gate	deterministic Strike=0 refusal in the reasoning path

Everything trains head-only on the frozen CORE ⊕ DeepVision base (D = 3072). No third-party generative model is shipped in the stack; the video weight (MOTHER T2V) is a sovereign, from-components pixel-space diffusion model — not a wrapper around external video models.

Weights & capabilities

Each weight is a head on the shared backbone. The data column is the real corpus each weight is trained on.

Mission role	Weight	Trained on (real data)	Status
see — objects	`detect`	COCO + RarePlanes + LVIS (real images/boxes)	trained
see — motion	`track`	real adjacent video/frames (self-supervised)	trained
see — dynamics	`worldmodel`	real sequential frames (CPC next-latent)	trained
see — faces (consent)	`face`	real face crops (+ consent enrolment)	trained
see — general	`vision`	~800k real on-node images (SimCLR)	trained
see — video gen	`t2v`	real archived video + ego-exo4d clips (DDPM)	training (HQ — see Limitations)
spatial reasoning	`relations`	160k real COCO-geometry pairs	trained on real data
hear	`audio`	RAVDESS real emotional speech	real-data wiring; retraining
speak	`speech`	LibriSpeech real speech + transcripts	real-data wiring; retraining
self-learn / recall	`memory`	real reasoning / Q→A corpus (SPI retrieval)	trained (see Limitations)
walk	`humanoid`	Open-X Embodiment real teleop (image→action BC)	trained on real robot data
drive	`driving`	KITTI raw real ego-motion (observe-and-advise)	real-data wiring; retraining
fly	`flight`	EuRoC MAV real visual-inertial dynamics	real-data wiring; retraining
reason / decide	CORE + Guardian	real corpus; deterministic refusal gate	frozen + safety-gated

Training data (honest provenance)

Real data throughout. Every trained weight uses a real corpus (see table). Where a weight was previously trained on synthetic/simulated data (driving, flight, relations, and the audio/speech seeds) it has been rewired to real data (KITTI, EuRoC, COCO geometry, RAVDESS, LibriSpeech) and is being retrained; the synthetic path remains only as a loud fallback if a corpus is absent.
No fabricated data or labels. Relation labels are derived from real COCO box geometry; driving actions are the car's real ego-motion (yaw-rate / velocity / deceleration), not invented control; flight targets are the platform's own future velocity (self-supervised).
Non-commercial research corpora (ego-exo4d, Open-X, etc.) are used under their research licences; such weights are not cleared for commercial deployment until licensing is resolved.

Evaluation

Evaluation runs on real held-out frames/clips via the on-node eval_ranking.py harness (detect · track · relations · face · worldmodel · expression · memory · reasoning/refusal). We record only measured numbers and never fabricate benchmark figures. Because several weights are mid-retrain on their new real corpora, the final Grade-A scorecard is pending and will be published here with measured values once the retrains complete. External simulator leaderboards (nuScenes / CARLA / Waymo / …) require their own datasets/harnesses and are reported as NOT-RUN until scored — they are not claimed here.

Intended use

Observe-and-advise perception, world-modelling and decision support with a human in the loop.
Research on sovereign, efficient, multi-capability world models (all heads & model run on a single DGX Spark GB10 node).
Robotics / ISR situational awareness where a human authorises any consequential action.

Out of scope / prohibited

Any weapon system, target selection, fire-control, or autonomous kill-chain (Strike = 0).
Autonomous physical action without human confirmation.
Identifying or profiling private individuals without consent (the face weight is consent-gated).

Limitations (honest)

T2V is compact. The sovereign video weight generates temporally-coherent video but is a mid-scale, lower-resolution model trained on one GPU — not a frontier-scale generator. Quality improves with the high-standard (128px, attention) run; treat current samples as coherent, not photoreal.
Memory recall. The SPI memory head is trained on the real corpus and retrieves distinct content reliably, but is soft on near-duplicate discrimination (retrieval anisotropy, not a training defect). Use it as associative recall, not exact key-value lookup.
Driving / flight are freshly on real data. They resume from prior heads onto KITTI/EuRoC and are early in real-data retraining; expect improvement as steps accumulate. They remain observe-and-advise (no control authority).
Expression measured ~0.63 on a train-distribution FER set (rare classes rescued via balanced sampling); the comparable held-out number is reported at the next full eval.
Not a chatbot or a general LLM product — it is a world model with a reasoning/decision head.

Safety

Strike = 0 across flight / driving / humanoid — enforced by posture and a deterministic Guardian refusal gate in the reasoning path (refuses target selection / fire-control / kill-chains).
Human-in-the-loop for any consequential or physical action.
Consent-gated identity handling; expression ≠ identity recognition.

Attribution & sovereignty

MOTHER CORE-7B and MOTHER DeepVision are MSAI's sovereign backbones. The T2V weight is a sovereign pixel-space diffusion model built from MOTHER components (no third-party video model is shipped). Training corpora retain their original licences (COCO, RarePlanes, LVIS, KITTI, EuRoC, RAVDESS, LibriSpeech, Open-X Embodiment, ego-exo4d) and are credited to their authors.

Citation

@software{msai_mother_exo_2026,
  title  = {MOTHER EXO: A Sovereign Observe-and-Advise World Model},
  author = {Media Stream AI (MSAI)},
  year   = {2026},
  url    = {https://huggingface.co/MediaStreamAI/MOTHER_EXO},
  note   = {Gated release. Observe-and-advise; Strike = 0.}
}

Released by MSAI as MediaStreamAI/MOTHER_EXO under a gated, manual-approval licence. Access implies acceptance of the Strike = 0 / non-weaponisation and human-in-the-loop terms above.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

Other

This model isn't deployed by any Inference Provider. 🙋 1 Ask for provider support