In one sentence, what does NEXUS Core actually do?

It reads a request written in ordinary language, reasons through it in steps, and writes back a useful answer — a summary, a recommendation, a draft document or an instruction for another module to act on.

How does it 'know' anything?

It is trained on the licensed, cleaned library that SENTRA prepares. During training it learns the patterns of how language, procedures and technical documents fit together, so at run time it can predict the most reasonable continuation of a task rather than looking answers up in a database.

Why is it cheaper to run than a conventional large model?

It uses a mixture-of-experts design: only a small number of specialist sections of the network switch on for any given request. You pay for the fraction of the model that was actually needed, not all of it.

What stops it from inventing answers?

Answers are grounded in retrieved source material where possible, and every response passes through AEGIS, which checks for unsupported claims, unsafe content and policy breaches before the user sees it.

Do we have to send data to a third party to use it?

No. The same model can run in our hosted environment or entirely inside a customer's own infrastructure, with identical behaviour — which matters for regulated industries.

Why does this module carry the largest share of the investment?

Reasoning quality per dollar of compute determines whether every other module is commercially viable. Training runs, evaluation and the engineering team behind them are the most capital-intensive part of the programme.

All modules

M01 · In Build

NEXUS Core

Foundation model and inference engine

Investment

$3,400,000

Share of programme

23.8%

Timeline

Q3 2025 — Q2 2027

Delivery team

24 engineers / 6 researchers

62% complete

Sparse mixture-of-experts neural network visualisation for NEXUS Core — M01 · NEXUS Core — Foundation model and inference engine

In plain English

The brain of the platform. It reads a question, works through it step by step, and writes back an answer — fast enough and cheap enough to run all day inside a business.

What the software does

Answers questions and drafts documents using the language and jargon of the industry it was trained on.
Works through multi-step problems — reading a fault log, checking a manual, comparing the two and proposing an action.
Powers every other module, so an improvement here makes the whole platform smarter at once.
Runs either in the cloud or on a company's own servers, with identical behaviour.

How it works, step by step

01It learns from vetted materialThe model is trained on the cleaned, licence-checked library that SENTRA prepares, so what it knows can be traced back to a source.
02It only wakes the part it needsInstead of running the entire network for every word, it activates a small group of specialist sections. That is why it is quick and comparatively inexpensive to run.
03It answers through one fixed connectionAll the other modules talk to it the same way, so upgrading the model does not mean rebuilding the software around it.

A simple analogy

Think of a panel of specialists behind one desk: you ask a question once, and only the two or three experts who are relevant actually get up to answer it.

Why it matters

Reasoning quality per dollar of compute is the single biggest driver of whether an AI product is commercially viable at scale.

How NEXUS Core works

Inside the module

A question in ordinary language goes in; a reasoned, checked answer comes out.

Input

User request
A question, instruction or document written in plain language
Curated corpus
Licensed training material prepared by SENTRA
Context
Retrieved documents and conversation history

M01 pipeline · select a stage

1/4

The request is turned into something the model can reason over

The text is broken into tokens and combined with any retrieved context, so the model sees the question and the supporting material as one continuous problem.

TokenisationRetrieval-augmented context assemblyLong-context packing

Output

Answer
A grounded response with its supporting sources
Action
A structured instruction another module can execute
Trace
Model version, sources and checks, recorded for audit

Stage by stage, in detail

Curriculum intake

SENTRA hands NEXUS a sampled, deduplicated and licence-attested corpus with quality weights attached to every shard.

Three-stage training

General pretraining, domain adaptation on curated industrial corpora, then preference tuning against expert-labelled task suites.

Sparse routing

A learned router activates a small subset of experts per token, giving 70B-class quality at a fraction of dense compute cost.

Optimised serving

Speculative decoding, paged attention and quantised expert routing let the same weights serve cloud and constrained on-premise fleets.

Versioned contract

Downstream modules call one stable model contract, so capability upgrades propagate without integration work.

3.1x

Inference throughput

Versus the reference open serving stack at equal quality

91%

Cluster utilisation

Sustained training cluster utilisation after scheduler work

512k

Target context

Long-context extension under active development

Evaluation suites

Run continuously against every candidate checkpoint

Sparse MoE transformerPyTorch + custom CUDA kernelsSpeculative decodingPaged KV attentionINT8 / FP8 quantisationgRPC model contract

Questions answered

NEXUS Core FAQ — how the AI works, in plain terms

Common investor questions about what this module does, how it does it, and why it is funded as part of the programme.

Scope

NEXUS Core is the central intelligence of the Mintelligent platform. It is a sparse mixture-of-experts transformer trained in three stages — general pretraining, domain adaptation on curated industrial corpora, and preference tuning against expert-labelled task suites.

The paired inference engine is written for mixed CPU/GPU fleets with speculative decoding, paged attention and quantised expert routing, allowing the same weights to serve both datacentre and constrained on-premise deployments.

Every downstream module calls NEXUS through a single versioned model contract, so capability upgrades propagate across the platform without integration work.

Contracted deliverables

70B sparse MoE base model, three checkpoints per year
Inference runtime with speculative decoding and KV paging
Model contract SDK (Python, TypeScript, gRPC)
Continuous evaluation harness across 40 task suites

Achieved to date

13B pilot model trained and beating open baselines on internal reasoning suite
Inference runtime hitting 3.1x throughput over reference serving stack
Training cluster commissioned and running at 91% utilisation

Currently in production

70B production pretraining run
Long-context extension to 512k tokens
Expert-routing compression for edge targets

Where it is used

Multi-step industrial reasoning

Diagnostic and planning tasks where the model must chain evidence across dozens of documents and system readings.

Private model hosting

Enterprises running the same weights inside their own perimeter with no change to application code.

Downstream capability lift

Every other module inherits reasoning gains automatically through the shared model contract.

Platform dependencies

Consumes curated corpora from SENTRA
Perception inputs projected in by ORACLE
Release-gated by AEGIS evaluations
Distilled and distributed by VERTEX

Key risks and mitigations

Scaling run underperforms the pilot

Staged checkpoints with go/no-go gates before each compute tranche is released.

Compute cost inflation

Reserved capacity contracts plus spot bin-packing through the VERTEX fabric.