Foundation-model lab

Foundation models for your industry

We train a foundation model on your industry and give you all of it. The model, the data behind it, and the eval report, running self-hosted on a single datacenter GPU. No APIs, no per-token bills, no weights you don't own.

Request a build 🤗 Open checkpoints

From the portfolio

Numerical AI

High-performance simulations on GPUs with AI

Cloud GPU physics with €0 licence fees. One account and shared credits. Sign up and run in minutes, and pay only for compute, from €1.

CMF Compressible multiphase CFD for shocks, cavitation, MHD, and reacting flows. 130+ validated examples.
CEM Palace full-wave 3D electromagnetics for resonators, antennas, S-parameters, and quantum hardware.
SRS PyFR scale-resolving turbulence for LES, DNS, and ILES on unstructured meshes.

AI Copilot · A100/H100 Visit Numerical AI →

Chaperone AI

Fine-tune and quantize, never train from scratch

We adapt a proven open checkpoint to your corpus with QLoRA fine-tuning and quantization, then ship the weights with the training data and the eval report.

Tune QLoRA on your corpus, geometry, or labels. Joint fine-tuning and replay-weights training mix the original checkpoint back in if the model forgets, as in PDEs foundational models.
Pack 4-bit GPTQ, about 60GB down to 20GB, self-hosted on a single L40.
Eval 84% MedQA on Thinking-LQ, within 4 points of GPT-4o. 30K+ Hugging Face downloads.

QLoRA · GPTQ · one L40 Visit Chaperone AI →

Simvera

Train in simulation deploy in the real

We render the events you cannot film, train the detector on that synthetic set, then run it on live camera feeds. No manual labels, no collection campaign.

Gen CARLA and NVIDIA Omniverse emit millions of frames with boxes, masks, and depth already attached.
Train SynYOLO learns from synthetic data only. Zero real photographs, zero annotation teams.
Deploy TensorRT on industrial RTSP or dashcams. The model has never seen the factory, and it still works.

Synthetic train · real-world deploy Visit Simvera →

Open research

Open weights and public datasets you can inspect

We publish sixteen open-weight models and two public datasets on Hugging Face so every published claim can be inspected in the open.

Phys Two models, Flow-1.0 and Palace-LoRA, plus the PyFR cylinder / Kelvin-Helmholtz dataset.
Vis Three YOLOX pylon detectors. COCO kept, cones added, in S, M, and L sizes.
LLM Eleven checkpoints, Thinking-LQ, quantized DeepSeek and Llama, Kiwi, and MedQA, plus the MedQA set.

See the register Hugging Face →

30K+

Hugging Face downloads

Open checkpoints published

130+

Validated simulation cases

Platforms in production

Third-party API calls

Portfolio register

Core platforms

Three specialist systems form the centre of the company. The application products below turn their capabilities into focused workflows.

02.1 Language systems

Chaperone AI

Private language systems for specialist domains. They provide useful reasoning while keeping data within the chosen boundary.

Visit Chaperone AI

02.2 Physics systems

Numerical AI

GPU-native simulation platforms for engineering. They support fluid dynamics, electromagnetics, and scale-resolving work.

Visit Numerical AI

02.3 Vision systems

Simvera

Industrial perception trained in simulation. It reduces manual labelling and transfers cleanly to real camera environments.

Visit Simvera

Focused applications

Capability, shaped into a task.

Focused tools support specific work. They share research and infrastructure across the portfolio.

A / 01

DeepResearch

Evidence-led research, document analysis, and report generation.

Language · research Explore

A / 02

PropensityAI

Grounded search and discovery with answer-level citations.

Language · search Explore

A / 03

AI 3D Generator

Text- and image-to-3D generation on a GPU-accelerated pipeline.

Geometry · creation Explore

The parent company

A common method.
Distinct products.

Empirisch Tech handles the difficult shared work. This includes applied research, evaluation, product incubation, and production infrastructure.

The shared foundation does not make every product the same. Language teams and simulation engineers need different interfaces. Factory cameras and research workflows need different evidence. Each platform is designed for its own field.

01Specialise the productOne clear job and audience per platform.
02Make evidence inspectableResults, limitations, and operating conditions belong together.
03Design for controlDeployment and ownership are product decisions, not afterthoughts.

Open research

Public checkpoints from the Chaperone line.

Named models from chaperoneai.net, plus quantized DeepSeek checkpoints on Hugging Face, published so the work can be inspected, not just described.

Language

Model	What it is	Published result	Adapted from	Weights
Chaperone-Thinking-LQ-1.0	Domain reasoning model, ~20 GB, one L40	84% MedQA	DeepSeek-R1-Distill-Qwen-32B	Hugging Face
DeepSeek-LLM-67B-Chat-gptq-8bit	8-bit GPTQ chat model, 67B	58.11% ARC-Challenge	DeepSeek-LLM-67B-Chat	Hugging Face
DeepSeek-R1-Distill-Llama-70B-gptq-4bit	4-bit GPTQ reasoning model, 70B	37.88% MMLU	DeepSeek-R1-Distill-Llama-70B	Hugging Face

Physics

Model	What it is	Published result	Adapted from	Weights
Chaperone-Flow-1.0	Flow operator for bluff-body wakes and mixing layers	1.8% wake error	Poseidon-B	Hugging Face
Qwen2.5-Coder-32B-Palace-LoRA	Writes and repairs Palace electromagnetics configs	3,296 schema pairs	Qwen2.5-Coder-32B	Hugging Face

Vision

Model	What it is	Published result	Adapted from	Weights
yolox-pylon-s	Edge-camera detector with COCO kept and cones added	cone AP 74.5 · COCO 41.6	YOLOX-S	Hugging Face
yolox-pylon-m	Fixed-camera detector with an accuracy and speed balance	cone AP 77.5 · COCO 46.8	YOLOX-M	Hugging Face
yolox-pylon-l	Server-side multi-stream detector	cone AP 78.6 · COCO 48.5	YOLOX-L	Hugging Face
yolox-pylon-xl	Maximum-accuracy size for offline analysis	in training	YOLOX-X	Vision page