Sovereign reasoning model

MOTHER CORE

A sovereign British reasoning model — the mind behind MOTHER.

MOTHER CORE is our sovereign large reasoning model (CORE-7B), trained and served entirely on UK infrastructure. It reasons over long context, grounds every answer in your own documents through retrieval, runs multi-step deep research, and calls tools — with governance, audit and refusal built in, not bolted on. CORE is the frozen backbone the rest of the MOTHER family is built on.

Request access All models

6.72B

parameters · custom arch

48L · 3072d

layers · hidden · GQA 24/6

77%

agentic 105-task · V3 measured

2.4M

training records · 9 groups

Capabilities

Long-context reasoning

Multi-step reasoning over large documents and conversations, with chain-of-thought kept internal and auditable.

Retrieval-grounded (RAG)

Answers are grounded in your own corpora via vector retrieval, with citations back to source — no hallucinated facts.

Deep research

Runs multi-step web + corpus research jobs that fan out, verify, and synthesise a sourced briefing on any topic.

Tool use & agents

Calls tools and orchestrates agent workflows using the Anthropic tool-use schema, with every action logged.

Refusal & governance

Safety and refusal are first-class: policy-gated outputs, human-in-the-loop on sensitive actions, full audit trail.

Sovereign & private

No foreign dependencies, no data egress — runs on UK/EU infrastructure under local law, on-prem or air-gapped.

Architecture

1
MotherCoreModel — 6.72B decoder-only
A custom decoder-only architecture: 48 layers, hidden width 3072, grouped-query attention (24/6 heads), SwiGLU MLPs, RoPE (θ=10000) and RMSNorm. 4096-token context, trained at 1536 sequence length in bf16. Tokenizer mother_hf (SentencePiece, vocab 50,258).
2
Trained from scratch on owned data
Full fine-tune with answer-only loss over 2,400,092 records across 9 balanced capability groups (A–I) on the NVIDIA GB10 DGX Spark node; cosine-LR warm-restart at 4e-6 (mother_train_v3.py). No third-party distillation — 100%-owned weights, sovereign-guarded checkpoints.
3
Measured, not projected
V2 (chunk 0600) scored 51/105 (49%); V3 (chunk 1550) scored 81/105 (77%) on the 105-task agentic benchmark, against a ≥80% gate. v4 is the next ranked run — its numbers are recorded only once measured.
4
Retrieval + reuse
Retrieval over Chroma + MongoDB grounds every turn with citations; the frozen CORE backbone is reused as the shared cognitive core across the MOTHER family (EXO, LLM, Code).

Specifications

Model	MOTHER CORE · MotherCoreModel (decoder-only)
Parameters	6.72B
Depth / width	48 layers · hidden 3072
Attention	Grouped-query (GQA) · 24 / 6 heads
Blocks	SwiGLU · RoPE θ=10000 · RMSNorm
Context	4096 (RoPE) · train seq 1536
Tokenizer	mother_hf · SentencePiece · vocab 50,258
Training	GB10 DGX Spark · full FT · cosine-LR 4e-6 · bf16
Corpus	2,400,092 records · 9 groups (A–I) · incl. 800k MOTHERrag
Measured	V2 51/105 → V3 81/105 (77%); v4 next
Hosting	UK sovereign · GB10, on-node · air-gapped tiers

Training corpus — 2,400,092 records across 9 capability groups (A–I)

Weight	Kind	Function	Dataset	Trained
Reasoning & agents	A–H	Vision-grounded QA · tool-use · multi-step agent traces · orchestration	balanced per-capability sampling	✓
Situational-awareness QA	reasoning	Answer questions over the live operating picture	sovereign QA — 861k	✓
Threat & decision	reasoning	Threat classification (INFO…SEVERE) + ranked decision support	military / space converted	✓
Memory (MOTHERrag)	memory	Long-term write & recall — semantic memory	MOTHERrag corpus — 800k	✓
Security & code (group I)	defensive-cyber	Detection engineering · mitigation · secure code · code-RAG	26k — security 12k · code 8k · rag 6k	✓

Model evaluation — industry benchmarks

Benchmark	Metric	Result	Status
MSAI Agentic (105-task)	success rate	77% · 81/105	◉ measured
MMLU	5-shot acc	—	○ to run
GSM8K	acc	—	○ to run
HumanEval	pass@1	—	○ to run
ARC-Challenge	acc	—	○ to run
HellaSwag	acc	—	○ to run
TruthfulQA	%	—	○ to run
MT-Bench	score	—	○ to run

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

What it's for

Sovereign enterprise assistant

A private reasoning assistant grounded in your policies, contracts and knowledge base — with citations and audit.

Deep research desk

Analysts commission multi-step research jobs and receive sourced, verifiable briefings instead of unattributed prose.

Agent orchestration

CORE plans and drives tool-using agents across your stack — Slack, Drive, GitHub, Postgres and more.

Safety & governance

Human-in-the-loop on any sensitive or outward-facing action.
Every request and tool call is logged for audit.
Refusal policy and content governance applied to all outputs.
No data egress — inference and storage stay on sovereign infrastructure.

Frequently asked

What is MOTHER CORE?

MOTHER CORE is Media Stream AI's sovereign reasoning model — a custom 6.72B decoder-only architecture (48 layers, 3072 hidden, GQA 24/6, SwiGLU/RoPE/RMSNorm) trained from scratch on 2.4M owned records on the GB10 node. It measured 77% (81/105) on its agentic benchmark at V3/chunk-1550, with retrieval-grounded answers, deep research and tool use.

Is MOTHER CORE sovereign?

Yes. CORE runs entirely on UK/EU infrastructure with no foreign technology dependencies and no data egress, available on-prem and air-gapped.

How does CORE avoid hallucination?

CORE grounds answers in your own documents through retrieval-augmented generation and cites the source passages it used, so claims are traceable.

What hardware does CORE run on?

CORE is served on NVIDIA GB10 Blackwell hardware on-node for low-latency sovereign inference.