We're building a new model class designed for systems that need to act, decide, remember, and scale without hardware becoming the bottleneck.
The next wave of AI isn't about better answers. It's about systems that make decisions, adapt to what they encounter, and work alongside other systems in production without the costs that kill adoption.
Starting with foundation models for agents, our mission is to give every device the intelligence to understand and interact with the physical world.
§ 01 — Two walls in the way
Agentic AI is the fastest-growing category in enterprise software. But these systems have a capability and an economics ceiling. Getting agents to work reliably at scale means rethinking what's underneath.
Language models reason without understanding consequences. They weren't designed to hold memory or coordinate across systems. Agents built on top of language models rely on complex engineering scaffolding, making them brittle and expensive to run.
The dominant answer to better performance is more compute. But hardware and energy costs are rising faster than the value they deliver. Agentic workflows run across chains of multiple steps, making them expensive to adopt. True scale will only come when performance stops depending on adding more resources.
§ 02 — From the architecture to solutions
We are designing from first principles the architectural layer that a new class of models requires, and that current architectures can't provide.
Four guiding principles underpin the direction of the models we are developing.
Our systems don't just produce outputs — they remember, build internal representations of their environments, learn continuously from what they do, and function as nodes in larger systems.
We are developing systems that reconfigure how they compute based on the task at hand, allocating capacity where reasoning demands it, just as biology would do.
Systems that act in the world and eventually in the physical world need to be safe and governable, as a property of the reasoning layer itself.
We treat cost-per-intelligence as a design constraint, because true adoption is unlocked when performance improves without spending more capital on infrastructure.
§ 03 — The Architecture
Inspired by how biological systems compute, DAEDALUS is our answer to what a different foundation for AI looks like.
Instead of a single general-purpose model, we build intelligence as a composition of specialised capabilities—assembled dynamically to meet the demands of each task.
We believe the future of AI is not a single model, but a carefully orchestrated system of them.
Rich internal models of environments and their dynamics, rather than text continuation alone.
Memory that evolves across interactions, instead of context windows that begin from scratch each time.
Systems that autonomously improve through continual learning as they accumulate experience, without full retraining.
Architecture that reconfigures itself to the task, routing capacity where it's needed.
Biologically inspired training dynamics and local learning rules.
Models learn to predict past, present, and future states.
Governance — wraps every capability
Every action evaluated against the operating norms before it executes — not after.
Beyond human-in-the-loop — control moves between human and system along a spectrum of variable autonomy.
§ 04 — Releases
Ogma is a family of small encoder models that substantially outperform expectations on standard embedding benchmarks (MTEB, 54 tasks, English). They're an early expression of our cost-per-intelligence principle in practice, increasing capability without cloud-scale infrastructure.
Our 8.6M model outperforms the 32M Potion baseline by +6.2 points on MTEB average.
Ogma is an early step toward scaling intelligence through structural design, rather than relying on additional infrastructure.
Explore OgmaCurrent Research Development
Research Project · 01
From fixed attention to dynamic focus.
Research Project · 02
From systems that reset to systems that remember.
§ 06 — R&D Partnerships
We partner with organisations through joint research, implementation, and R&D projects — designing bespoke AI architectures, optimising and shrinking models to reduce API and infrastructure costs, orchestrating agentic workflows, and embedding governance into operational processes. We apply our own models and tools to improve the performance of existing systems, alongside building new ones.
A collaboration with the UK Government through NIHR Applied Research Collaboration Yorkshire and Humber, improving policy decision-making for better democratic and informed decisions in government on key issues.
We are building the era of systems that memorise, coordinate, adapt, and improve, where economics, energy, and scale follow from architecture.
The next breakthrough in AI will come from systems built differently from the ground up.