Foundation-model lab · Vienna, Austria

We train foundation models on the physics, language, and imagery of your industry.

Then we hand over the whole build — the checkpoint, the data it was trained on, and the eval report — compressed to run self-hosted on a single datacenter GPU. No third-party APIs, no per-token bill, no weights you don't own.

Request a custom build See the open checkpoints

What ships to you

Checkpoint~20 GB, quantised — runs on one L40. Yours to keep.
Training setThe corpus, geometry, or labelled frames it learned from.
Eval reportBenchmarked against frontier models on public tasks and unseen regimes.
RunbookDeployment on your infrastructure — Kubernetes, air-gapped if needed.

100% self-hosted · no third-party APIs

One method, three domains

Pick the line that sounds like your job.

The same method — adapt a proven base, never train from scratch — applied to medical text, to fluid dynamics, and to a factory camera feed.

Language

Private domain LLMs

Reasoning models fine-tuned on your corpus, plus the chatbots and semantic-analysis products that run on them.

84% MedQA · ~20 GB · one L40

Explore Language

Physics

Flow surrogates

Operator models for CFD regimes a foundation model never saw — without erasing what it already knows.

1.8% wake error · solver-grade data included

Explore Physics

Vision

Detectors without labels

YOLO detectors adapted to your objects, cameras, and process — trained in simulation, shipped with the labelled set.

weld 0.97 · defect 0.91 · zero manual labels

Explore Vision

Evaluated in the open

Every claim on this page is a row in a table.

Published results · Empirisch Tech checkpoints
Build	Task	Result	Adapted from	Footprint
Thinking-LQ-1.0	MedQA — medical question answering	84.0%	DeepSeek-R1-Distill-Qwen-32B	~20 GB · 1× L40
GPT-4o	MedQA — reference point	≈88%	—	Hosted API
Flow-1.0	Wake error, unseen flow regime	1.8%	Poseidon-B	Single GPU
SynYOLO — weld	Weld detection, industrial camera	0.97	YOLOX	TensorRT · real-time
SynYOLO — defect	Surface defect detection	0.91	YOLOX	TensorRT · real-time
Simvera field	Mean localisation, live factory cameras	6.8 px	SynYOLO	100% synthetic training data

◆ Our builds. The GPT-4o row is the published comparison point — Thinking-LQ-1.0 lands within four points of it, self-hosted. Weights for our rows are on Hugging Face; reproduce them yourself.

How it works