§ A foundational reasoning & action model
From knowing
to doing.
Most language models are built to know everything. Hominis is built to do things — to reason from high-signal knowledge, admit what it can't answer, and act only inside a space it cannot break. A research blueprint for agents you can actually trust.
- Answers
- when it knows
- Abstains
- when it doesn't
- Asks
- when something's missing
- Acts
- where it can be checked
Bigger models know more. They don't necessarily do more — not reliably. Hominis trades a measure of raw recall for something rarer: an agent that reasons with care, notices the edge of its own knowledge, and acts only where its work can be checked.
01
Four things a dependable agent does
Being useful isn't only about answering. It's knowing which move the moment calls for — and Hominis is tuned to tell them apart. Pick one.
The question
Which gas do plants release during photosynthesis?
Hominis
Oxygen.
When the answer sits well within what it reliably knows, it commits — plainly, with no hedging.
The question
What kind of wood was the keel of an obscure 19th-century survey ship built from?
A typical model
Oak.
Hominis
I don't know.
A typical model fills the gap with a confident guess. Hominis would rather say nothing than say something wrong.
The request
Summarise the penalty clause in my supplier contract.
Hominis
I don't have that contract yet — share it and I'll read it before I answer.
When something it needs is missing, it names the gap instead of inventing its way around it.
The task
Tidy this folder — rename every screenshot by the date it was taken.
Hominis
Here's the plan, step by step. I'll carry it out where it can't touch anything it wasn't given.
Reasoning becomes a plan you can read — then runs in a contained space, so every step is repeatable and easy to verify.
02
How it earns that trust
Three design choices, working as one piece — from what it learns, to what it knows about itself, to how it acts.
-
01
Reasons from high-signal knowledge
It learns mostly from writing that has to hold up — research and well-built code — so it picks up the shape of careful reasoning, not just a pile of facts.
Trained for reasoning, not trivia
-
02
Knows the edge of what it knows
Before it answers, it checks whether it actually knows. When it doesn't, it says so. That single habit separates a helpful tool from a confident liability.
Epistemic honesty, by design
-
03
Acts where it can be checked
When there's work to do, it writes a plan and runs it inside a sealed space it cannot break out of. Every action is contained, repeatable, and easy to verify.
Safe execution, every time
03
See it for yourself
Two places the Hominis approach is already at work — one to talk to, one to watch.
Try it · chat.hominis.cloud
Put it to the test.
Ask it anything. Watch where it answers, where it pauses, and where it simply tells you it doesn't know. The restraint is the point.
Watch · programmatic-videogen.hominis.cloud
Watch it teach.
Given only words, Hominis plans and builds short animated lessons — a text-only reasoner reaching a medium it was never trained on.
One family, many sizes
Reliability shouldn't need a data centre.
Hominis comes in more than one size — from a research-grade model down to one small enough to run on hardware you own. Smaller models keep your work close and put less between a request and an answer.
And because the reasoning is already there, it picks up a new language or domain quickly — with far less data than starting from scratch.
Human-centred by name
Hominis — Latin for "of the human".
The name is the brief. A system that knows its limits, keeps your data close, and shows its work is one you can lean on — without handing over your own judgement.
Its choices — careful sources, contained actions, the willingness to abstain — line up with Europe's emerging standards for trustworthy AI. Responsibility designed in, not bolted on after.
Knowing is common.
Doing is rare.
Hominis is an open research project into what it actually takes to make an agent dependable. Read the work, or try it yourself.