The MOTHER AI
family.
7 sovereign models, 240B parameters, one lineage - the reasoning core others are built on, the cyber mind that defends, the vision tower everything sees through, the generative stack that turns a prompt into a programme, and the world model that puts it all in a body. Trained from scratch on owned data, on British infrastructure.
Every model in the family
Open a card for what a model is, what it was trained on and what it can do. Models published on Hugging Face link straight to their repository. The figures come from the published Art. 53 training-content summary and the model cards themselves.
MOTHER_CORE_V3 on Hugging Face- Parameters
- 6.72B
- Architecture
- MotherCoreModel · decoder-only
- Layers
- 48 · hidden 3072 · 24/6 GQA · SwiGLU
- Context
- 4096 · RoPE θ=10000 · RMSNorm
- Vocab
- 50258 SentencePiece · bf16
- Languages
- English · Welsh · Irish · Scottish Gaelic
- Next scale step
- CORE 70B · planned, 4× H200 TP
MOTHER CORE
A 6.72B decoder-only reasoning transformer (48L · 3072d) trained from scratch - deterministic at T=0, agentic, and the frozen backbone every other model in the family shares.
MOTHER CORE · CC
Defensive cyber cognition trained into CORE: live threat-intel fusion, exposed-asset discovery and SOC automation, air-gapped and defensive only.
MOTHER LLM
The from-scratch public language line, 13M to 1B - the expressive engine for narrative, code, policy and teaching where CORE is the deterministic one.
MOTHER DeepVision
The sovereign vision tower (27L · 1152d · 896px) driving detection, tracking, faces, expression, spatial relations and scene graphs.
MOTHER T2V
MOTHER’s own 3D video VAE plus a flow-matching Diffusion Transformer (~0.85B) - no Sora, Wan, LTX or SVD weights anywhere in the stack.
MOTHER IntuiTV
The IntuiTV Studio model: five trained heads and a six-agent fleet that direct, write, score, cast and stitch one prompt into a 720p programme with sound.
MOTHER EXO
The world model - frozen CORE mind, frozen DeepVision eye, 15 weights - proven driving a humanoid, two drones, a vehicle and a manipulator from one inference.
MOTHER memory
A three-tier agent memory - episodic, semantic and procedural - so the family carries context across sessions and missions.
Open-weight & auditable
Published under EU AI Act Article 53 with a copyright policy and training-content summary. No black boxes, no distillation, no borrowed weights.
One frozen brain, many senses
The family is not seven unrelated models - it is one architecture used seven ways. A frozen CORE mind and a frozen DeepVision eye are trained once and reused; lightweight task heads add new senses cheaply, and every head inherits the same auditable reasoning contract. Train a head, gain a capability, without ever destabilising the core.
- Frozen CORE backbone · 48L · 3072d
- Frozen DeepVision ViT · 896px
- Head-only training · atomic checkpoints
- Base weights never overwritten
- 15 published weights across the family
- One inference, five bodies (EXO)
Open source and commercially deployed
The CORE checkpoints are published open-weight on Hugging Face; the EXO world model and T2V generator are access-gated for controlled release. All of them run in production today - inside AUTM.ai, IntuiStudio, IntuiTV, MOTHER Defence and the MOTHER Robotics platform - on MSAI’s own sovereign compute.
- Open weights: MOTHER CORE V2 · V3
- Access-gated: MOTHER EXO · MOTHER EXO T2V
- Deployed in AUTM.ai, IntuiStudio and IntuiTV
- Serving on sovereign GB10 Blackwell nodes
- REST access with a free tier and enterprise plans
- On-prem and air-gapped deployment available
Sovereign by construction
Every model in the family trains and serves on MSAI infrastructure - our own data centres, our own power, our own control plane. Weights and data never leave your control: on-prem by default, with a full model card, checkpoint lineage and reproducible training steps behind every release.
Specialisations on the same core
Domain experts are fine-tuned on the deterministic CORE backbone, each a sovereign reasoning specialist tuned for its field while sharing one auditable core - and the public LLM line is tuned the same way for human-facing work.
- CORE-Reasoning · theorem proving
- CORE-Science · physics simulation
- CORE-Defence · threat assessment
- CORE-Legal · compliance
- LLM-Creative · narrative
- LLM-Policy · government writing
- models
- 7
- parameters
- 240B
- weights
- 15 published
- backbone
- MOTHER CORE 6.72B
- vision
- DeepVision ViT 896px
- generative
- T2V VAE + LatentDiT
- precision
- bf16
- serving
- GB10 Blackwell · on-prem
- licence
- MSAI open-weight
- compliance
- EU AI Act Art. 53
One family. One owner.
Every MOTHER model, its card, its weights and its training-content summary live at motherai.uk - the home of the sovereign model family.
Visit motherai.uk