README.md

---
license: other
license_name: msai-sovereign-gated
license_link: LICENSE
language:
  - en
library_name: transformers
pipeline_tag: other
tags:
  - world-model
  - multimodal
  - robotics
  - sovereign
  - observe-and-advise
  - detection
  - tracking
  - text-to-video
  - latent-diffusion
  - speech
  - reasoning
  - tool-calling
gated: manual
extra_gated_prompt: >-
  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.
extra_gated_fields:
  Full name: text
  Organisation: text
  Country: country
  Intended use (be specific): text
  I will use MOTHER EXO for observe-and-advise / research / defensive purposes only: checkbox
  I will NOT use it for target selection, fire-control, autonomous kill-chains, or any weapon system (Strike = 0): checkbox
  I will keep a human in the loop for any consequential or physical-world action: checkbox
  I will not use it to identify or profile private individuals without consent: checkbox
  I accept the gated licence terms: checkbox
---

# MOTHER EXO — Sovereign World Model

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

**MOTHER EXO** is a single sovereign **world model** that sees, hears, speaks, reasons, remembers
and acts across land, sea and air. It uses a **frozen-backbone + head-only** architecture: one
shared representation drives many trained heads and multiple robotic bodies (humanoid, drone,
vehicle) from a single inference.

> **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.

---

## Architecture (sovereign stack)

| Component | Role | Notes |
|---|---|---|
| **MOTHER CORE-7B** (frozen) | text encoder + reasoning + tool-calling | 48L / 3072d, 6.72B 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) | text-to-video generation | latent diffusion: 3D KL-VAE (118M) + flow-matching DiT (734M), CORE-conditioned |
| **Guardian** | safety gate | deterministic Strike = 0 refusal in the reasoning path |

Every weight trains **head-only** on the frozen CORE ⊕ DeepVision base (D = 3072). The video weight
(MOTHER T2V) is a sovereign latent video-diffusion stack built from MSAI components — a 3D KL
variational autoencoder and a latent flow-matching diffusion transformer conditioned on frozen CORE
text — not a wrapper around an external video model.

---

## Weights & capabilities

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

| Mission role | Weight | Trained on | Status |
|---|---|---|---|
| see — objects | `detect` | COCO + RarePlanes + LVIS | trained |
| see — motion | `track` | adjacent video/frames (self-supervised) | trained |
| see — dynamics | `worldmodel` | sequential frames (CPC next-latent) | trained |
| see — faces (consent) | `face` | face crops (+ consent enrolment) | trained |
| see — general | `vision` | ~800k on-node images (SimCLR) | trained |
| see — video gen | `t2v` | ego-exo4d + OpenVid captioned clips (~13k) | VAE training (256px); DiT pending |
| spatial reasoning | `relations` | 160k COCO-geometry pairs | trained |
| hear | `audio` | LibriSpeech + RAVDESS/CREMA-D | trained |
| speak | `speech` | LibriSpeech speech + transcripts | trained |
| self-learn / recall | `memory` | reasoning / Q→A corpus (SPI retrieval) | trained |
| walk | `humanoid` | Open-X Embodiment teleop (150k steps) | trained |
| drive | `driving` | KITTI raw ego-motion (150 drives) | trained; retraining on expanded corpus |
| fly | `flight` | EuRoC MAV visual-inertial dynamics | real-data fine-tune (staging) |
| fly — motor control | `flight_bc` | Blackbird rotor-command behaviour-cloning | roadmap (loader + trainer implemented) |
| reason / decide / tools | CORE + Guardian | frozen 6.72B backbone + refusal gate | frozen + safety-gated |

**Bodies:** humanoid, drone and vehicle are driven live from the shared model (land + air). A maritime
body (boats) is on the roadmap; no maritime head is trained in this release.

---

## Training data

- Every trained weight uses a real corpus (see table). Weights previously trained on
  simulated data (driving, flight, relations, audio/speech) have been moved to real data — KITTI raw,
  EuRoC MAV, COCO box geometry, LibriSpeech, RAVDESS/CREMA-D.
- Label provenance: relation labels are derived from real COCO box geometry; driving actions are the
  vehicle's recorded ego-motion (yaw-rate / forward velocity / longitudinal deceleration); flight
  targets are the platform's own future velocity (self-supervised).
- The T2V stack is trained on ~13k real captioned clips (ego-exo4d + OpenVid). The 3D VAE is trained
  first at 256px; the latent DiT trains on the frozen VAE latents once the VAE converges.
- Non-commercial research corpora (ego-exo4d, Open-X Embodiment, etc.) are used under their research
  licences; weights that depend on them are not cleared for commercial deployment until licensing is
  resolved.

## Evaluation

Metrics are recorded from measured runs on real held-out frames/clips via the on-node
`eval_ranking.py` harness (detect · track · relations · face · worldmodel · expression · memory ·
reasoning/refusal). Benchmarks that have not been run are marked NOT-RUN. Several weights are
mid-retrain on their real corpora, so the final scorecard is pending and will be published here with
measured values.

**Leaderboard harnesses (in-repo):**

| Board | Target | Adapter | State |
|---|---|---|---|
| VSI-Bench | visual-spatial world-model QA | `vsi_bench_eval.py` | ready; rank NOT-RUN |
| VBench | text-to-video generation | `vbench_eval.py` | pending trained DiT |
| BFCL | CORE agentic tool-calling | `bfcl_run.py` | ready; rank NOT-RUN |

External simulator leaderboards (nuScenes / CARLA / Waymo / HumanoidBench / Flightmare) require their
own datasets and harnesses and are reported as NOT-RUN until scored.

---

## Intended use

- Observe-and-advise perception, world-modelling and decision support with a human in the loop.
- Humanitarian and civil-protection response: aerial survey, access-route mapping, search-and-locate
  support, and coordinated movement of aid, with human authorisation for any consequential action.
- Research on sovereign, efficient, multi-capability world models (Built to 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

- **T2V scale.** MOTHER T2V is a ~0.85B latent generator (118M VAE + 734M DiT) trained on ~13k clips
  on a single GPU. It produces temporally-coherent video; it is not a frontier-scale generator, and
  the DiT stage is not yet trained.
- **Memory recall.** The SPI memory head retrieves distinct content reliably but is soft on
  near-duplicate discrimination (retrieval anisotropy). It functions as associative recall, not exact
  key-value lookup. The step-1500 checkpoint is the locked head.
- **Driving / flight.** Driving is trained on real KITTI ego-motion and is retraining on the expanded
  150-drive corpus. Flight is fine-tuning on EuRoC MAV dynamics. A Blackbird rotor-command control
  head (`flight_bc`) is implemented but not yet trained. All three are observe-and-advise, with no
  control authority.
- **Expression** measured ~0.63 on a train-distribution FER set (rare classes rescued via balanced
  sampling); the held-out figure is reported at the next full eval.
- This is a world model with a reasoning/decision head, not a general chatbot or LLM product.

## 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. MOTHER T2V is a sovereign latent
video-diffusion stack built from MSAI components (no third-party video model is shipped). Training
corpora retain their original licences (COCO, RarePlanes, LVIS, KITTI, EuRoC MAV, RAVDESS, CREMA-D,
LibriSpeech, Open-X Embodiment, ego-exo4d, OpenVid) and are credited to their authors.

## Citation

```bibtex
@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.*