Animas-R

The reasoning half of Animas — a language-model engine written in pure C. Training and inference, no Python, no framework, no runtime dependency. It runs where your data is, disconnected.

Schedule a Discovery Meeting

Reasoning

What it is

Our own model, and the engine that trains and serves it.

Animas-R is a language-model engine QBO owns end to end. It is one deployable artifact with no framework underneath it — no Python runtime, no CUDA-Python stack, no package tree to reconcile before the box will answer. That is what makes it auditable, small enough to ship, and able to run in an enclave that will never see a network.

Because it is ours, there is no cloud API to call and no keys to manage. The model answers on your hardware, inside your perimeter, whether or not the link is up.

Denied comms is not denied AI.

What it does

Trains and serves

The same engine does both. A model trained here saves in the inference format and serves identically — the two halves exchange weights rather than being bridged by a conversion step.

A modern architecture

A current-generation architecture, trained and served by the same engine on CPU and GPU alike — not last year’s model shape with a new label.

Provably correct

The GPU gives the same answer as the CPU, to the last decimal place, checked automatically on every change. What you train is exactly what you serve.

Fits the memory you have

An int8 inference path is what makes a large model fit a real budget, on hardware you can actually carry.

Scales across GPUs

One GPU or a rack of them, with no change to your code or your workflow — the same engine scales up as the job grows.

Tokenizes in the enclave

Your corpus is prepared and tokenized inside the airgap, verified character-for-character against the reference — nothing leaves to be processed.

Why it is built this way

No dependency to trust

Pure C with nothing under it. What you audit is what runs, and what runs is what ships.

Memory-clean, proven

Memory-clean across training, inference and generation — proven by the test suite on every change, not asserted in a datasheet.

Rack to vehicle

The same engine and the same behaviour on amd64 and arm64.

Sovereign by construction

Your corpus, your weights, your hardware. Nothing leaves the box to make the model work.

The model is ours. So is the box it runs on.

See it alongside Animas-V and the QBO API.

Request a demo→