About Empirisch Tech
We take a proven open checkpoint, adapt it to the corpus, geometry, or label set of a single industry, and hand the whole build over. You get the weights, the data behind them, and the evaluation that says where they hold and where they break. It runs on your own GPU, so there is no API in the middle and no per-token bill.
Why we exist
General capability is already paid for. What is missing is your corpus, your geometry, your label set, and an honest account of where the result fails.
Training a frontier model from scratch buys capability that open checkpoints already have. We start from one of those, spend the budget on domain adaptation and evaluation instead, and compress the result until it fits on a single datacenter GPU. That is a smaller claim than most of the market makes, and it is one we can put numbers behind.
What we work on
Language, physics, and vision are different fields with different failure modes, so each one is run as its own platform with its own customers. What they share is the bench, the same adaptation method, the same evaluation discipline, the same handover.
Private language systems for specialist domains. QLoRA fine-tuning on your corpus, 4-bit GPTQ packing, and replay weights that stop the base model forgetting what it knew.
GPU-native simulation for engineering. It covers compressible multiphase CFD, full-wave electromagnetics, and scale-resolving turbulence, with no licence fee between an engineer and a solver.
Industrial perception trained in simulation. We render the events you cannot film, train on that synthetic set, and deploy against live camera feeds. No collection campaign, no annotation team.
How a build runs
Medical text, fluid dynamics, or a factory camera feed. The subject changes, the sequence does not.
A frontier open checkpoint brings the general capability, so your data does not have to.
QLoRA and GPTQ for language, operator learning for physics, synthetic frames for vision.
Benchmarked against frontier models on public tasks and on regimes it has never seen.
Checkpoint, training set, eval report, and runbook, compressed to run self-hosted.
How the companies are set up
Empirisch Tech GmbH holds a participating interest in Empirical Systems. Adapting a checkpoint and operating it inside a regulated, air-gapped environment are different disciplines, and a build is only finished when both have signed off.
The lab. It carries the shared work that no single product could fund on its own, from applied research and evaluation to product incubation and the publishing of open weights and datasets.
The infrastructure side. It builds and operates the clusters, pipelines, and security posture that a model actually runs on, including on-premises and air-gapped installs for regulated customers.
Backed by the programmes we build on
Past engagements
The practice underneath
The delivery engineering behind every build we ship, available on its own if that is what you need.
Talk to us
A technical call goes straight to a founder, not to an account manager.
