A post-LLM model class.

Verifiable, continual, hardware-flexible. By architecture.

meu is a continual learning system generating verifiable world models for specialised applications. Grounded in the mathematics of topos theory, meu models any system verifiably while learning the rules generating data from the data. For any domain, at any scale and capacity.

01

Verifiable

+
02

Continual

+
03

Hardware-flexible

+

The Architectural Shift

A map, not a search.
The shift shows up in every capability.

LLMs search across everything they have seen. meu holds a map of the relationships inside a system, a world model of the domain, and moves straight to the answer. Six properties follow from that shift, in one model class. No current AI model holds all six together.
Capability
meu
LLM
Verifiable
No black box. Verifiable from inside the model while it runs, by construction.
Interpretability reverse-engineered after the fact.
Steerable
Holds intent through every step of execution.
Prompts hope the chain stays on instruction.
Adaptable
Maps the tools and rebuilds the path as conditions change.
Fixed at training. Stack must reshape to the model.
Efficient
Right-sized to each task. Inverts the cost line.
Massive parallel pre-training. Inference at scale.
Continual
Learns while operating. No retrain to wait for.
Frozen on training day. Periodic retraining cycles.
Hardware-flexible
CPU when that's enough, GPU when it isn't.
GPU-bound at every layer.

The Capabilities

Six capabilities.
One model class.

The Intersection

All six must be present.
Not layered on later.
Not simulated through prompts.
Present in the architecture itself.

01

Verifiable

Ask meu why it answered, and trace every step from inside the model while it runs. Not interpretability reconstructed after the fact.

02

Steerable

meu holds intent through every step. Intervene, redirect, or inspect at any point, rather than hope the prompt holds.

03

Adaptable

Tell meu what you want, not how. It maps the tools, builds the execution path, and rebuilds it as conditions change.

04

Efficient

meu deploys only the model each task needs, working in parallel on the hardware you already own.

05

Continual

meu learns a new domain by operating in it, continually, with no retraining cycle to wait for.

06

Hardware-flexible

meu allocates compute where it is needed: CPU when that is enough, GPU when it is not. It runs inside the server estate you already own.

All six present in every deployment

Why Now

"The next decade of AI will be decided by who owns the architecture."

95%

Enterprise AI pilots deliver no measurable return. MIT NANDA, 2025

7%

EU AI Act fines up to 7% of global turnover. Enforcement August 2026

80–90%

Share of enterprise AI compute spent on inference. McKinsey, 2024

Trust

Auditability is becoming mandatory. The EU AI Act enters enforcement in August 2026. LLMs compensate with more tokens, scaffolding, and post-hoc checks. meu makes trust native to the architecture: white-box, verifiable, introspectable, and auditable by design.

Cost

LLMs carry cost twice: massive pretraining upfront, then broad execution on every workload. meu changes the cost logic: it learns while operating, moves directly to the relevant structure, spends fewer tokens, and runs right-sized on existing hardware.

Independence

With an LLM, your data improves someone else's model and you get the capabilities they choose to ship. meu makes capability owned: it learns on your data as it operates, compounds inside your environment, and improves with the work you do.

Proof

Internal proof.
Benchmarks this autumn.

Demonstrated

Already demonstrated internally, on CPU.

Real-time system construction: From a single prompt, meu built a verified program from formal logic and first principles.

Real-time resource allocation: meu read the live data of a system it had never seen and steered its allocation in real time.

Next · Autumn 2026

The internal benchmarking loop.

This autumn meu runs against ARC and a broader set of reasoning benchmarks, internally, on our own systems. Each run sharpens the core model and makes it faster at producing the next specialised model.

Product · How we ship it

Three forms.
One motion.

A model you can see into, steer mid-run, and host on hardware you own, that learns your domain while it operates. meu ships in three forms from one substrate, and they compound: every specialised model feeds the foundation, making it better and more efficient at producing the next one. The research runs on synthetic and available data, so the loop never depends on customers, and their data can stay inside their walls.

meu foundation

The model class itself. Semantic, verifiable, steerable, not bound to GPUs. Foundation and specialists improve one another by construction; the road to standalone.

Internal use 2026 · Standalone 2027

Your model

Trained on your data, in your environment, isolated. Built to read every form of data and map the whole system it lands in, down to each dependency and what a change affects. It learns by operating, not by retraining, and you steer every step.

2026 → 2027

Model engine

The platform layer. External builders and vertical products spin up their own specialists on the same substrate, widening where meu runs.

From 2029