JKU Linz Sepp Hochreiter Banking · regulated Use case Healthcare · HIPAA Use case University research Use case Industrial AI Use case Public sector Use case Pharma · GLP Use case JKU Linz Sepp Hochreiter Banking · regulated Use case Healthcare · HIPAA Use case University research Use case Industrial AI Use case Public sector Use case Pharma · GLP Use case
Stop renting AI

Cloud services are no longer required.

Eldric ships three clients, one toolset. They work against your own Eldric cluster — or, if you only have a laptop, they talk directly to whichever provider you choose. You decide which.

Built-in web chat

Open a browser, you have ChatGPT. Hosted by your own Eldric server, themed per tenant, with knowledge-base search, voice in / out, plugins, and signed share-links — all included.

Command-line client

Pipe-friendly for shell scripts, interactive for everything else. Talks to a cluster or directly to any backend you like. macOS today; Windows and Linux in development.

Universal native client

A native desktop GUI that talks to your Eldric cluster or straight to OpenAI, Anthropic, xAI, Groq, Together, HuggingFace, or a local Ollama. One tool, every backend. macOS today; Windows and Linux in development.

The same GUI that you use without a cluster — for solo work, with whatever API key you bring — unlocks multi-tenant projects, agentic RAG, training pipelines, and the matrix-memory architecture the moment you point it at an Eldric server. Two modes, one client, one workflow.

Who it is for

From AI factories to your own organization.

Eldric is the operating system for sovereign AI compute — and the AI server, the AI infrastructure, and the AI platform for every organization built around it. Pick the angle that matches your role.

Sovereign AI compute operators

National AI factories and member-state-funded compute capacity. Run AI workloads under your jurisdiction's data-residency, procurement, and security frameworks — without depending on a hyperscaler.

See the operator angle →

Member-state AI consortia & research institutions

Public research labs, national HPC fabrics, and consortium-led AI compute programmes. Multi-tenant, multi-discipline, multi-jurisdiction — managed centrally, controlled locally.

See the consortium angle →

You want AI in your organization

Private chat, agents, knowledge bases, training — running on your own hardware. The data never leaves your network. For banks, hospitals, universities, factories, public sector.

See it for enterprises →

You run a datacenter

Eldric is a complete AI infrastructure platform that sits on top of your compute, storage, and network. Multi-tenant, federated, plugin-driven, with HA built in.

See the datacenter angle →

You sell services to your customers

ISP, MSP, telco, regional cloud? Run Eldric on your hardware and offer Private AI to your enterprise customers as your own white-labelled service — they never see a hyperscaler.

See it for service providers →

What you do with it

Ten things, in plain words.

These are the things people actually use Eldric for. Each one is a single click in the chat shell.

Private chat

Talk to an AI like ChatGPT. The conversation stays in your database, on your network.

Knowledge base

Upload documents, ask questions across them. Sources are cited inline so you can check the answer.

Voice dictation & transcription

Speak instead of type. Meeting transcripts, dictation, accessibility input. Telephone-style AI calls are in development.

AI inbox

Drafts replies to email, SMS, WhatsApp, Signal, Teams. A human approves before anything sends.

Train your own

Fine-tune a model on your own data (LoRA, QLoRA, DPO). Your data never leaves the building.

Industrial AI on the line

Read PLCs, OPC-UA, Modbus, MQTT Sparkplug B. Predictive maintenance, alarm triage, OEE analytics, shift-handover notes. Edge-deployed, soft-real-time — ~50 ms tag-to-action on the demo cluster. The strongest production-tested workload today.

Science look-up

One hundred and forty integrations to NASA, CERN, PubMed, NCBI, GBIF, IAEA, and more. Ask in plain English.

Code helper

Repo search, ticket triage, doc generation, test writing — your own private Copilot, on your own GPUs.

Multi-team

Each team gets its own tenant: separate data, separate model access, separate billing. One server, clean walls.

OpenAI drop-in

Tools written against the OpenAI API work unchanged. Point them at your Eldric and go.

Where it runs

From a Raspberry Pi to supercomputers.

Eldric is the same binary on every target. What changes is how much of it you switch on — and the resource budget it gets to play with. A field engineer on a Pi gets classification and short completions. A hospital network running a cluster gets multi-tenant agentic RAG and federated training. A national-AI compute programme running across a thousand nodes gets the same kernel sized to the metal underneath. Same software, scaled to the hardware.

Raspberry Pi

The Pi 4 (8 GB) runs the kernel and a small model end-to-end — enough for classification, short completions, and IoT-style ambient AI on a sealed device.

Workstations

A developer laptop or engineering workstation runs the GUI client and a local Eldric in tandem. Cloud-free dev loop. Fly without WiFi.

Single servers

One tower, one GPU, one team. A small business or department gets the whole platform on a single box without a cluster manager.

Robots & vehicles

Mobile robots, AGVs, inspection drones, and in-vehicle stacks carry an embedded Eldric kernel for on-device perception, voice command, and recovery decisions when the radio link is out.

Enterprise clusters

Multi-node deployments across racks, datacenters, or building floors. Rolling upgrades, automatic failover, multi-controller HA, federated learning. The full stack.

Supercomputers & AI factories

HPC fabrics, EuroHPC-adjacent compute, national-AI capacity programmes. Thousands of nodes, the same kernel, the same module surface — just bigger resource budgets and federation across sites.

Other targets ship in the same RPM / DEB / PKG today: edge gateways, kiosks, retail-store back-rooms, ships, oil rigs, mobile clinics. Anywhere you can run a Linux box, you can run Eldric.

Our demo cluster, right now

What the numbers actually say.

Measured, not promised. The demo cluster runs one inference GPU (RTX 4070 Ti, 12 GB) plus one router GPU (RTX 2080, 8 GB). Smaller deployments scale linearly down; larger ones up.

793/s
Sustained chat requests at 32 concurrent connections
41 ms
Median chat round-trip under that load
42 ms
p99 latency under load — stays flat as concurrency rises
140+
Integrations to scientific databases & agencies
Knowledge-base search at concurrency four still hits a ~7 second p50 latency cliff. Multi-user concurrent search is the next engineering focus — we will publish numbers when it lands. We do not claim sub-second search at scale today. Honest scope · 2026-05-15
Who uses it

Twelve sectors, one server.

Each one links to a fuller page. The common thread: data that cannot leave the building.

HospitalsHIPAA-aware BanksAir-gap ready FactoriesOPC-UA / Modbus UniversitiesResearch IP stays put Law firmsPrivileged data InsuranceClaims + policy PharmaGLP / 21 CFR Part 11 Genomics labsBLAST, CRISPR Trading desksNews + filings Public sectorMultilingual Q&A AutomotiveR&D + plant RoboticsAGVs, drones, AMRs Space / observatoriesNASA, ESA, JWST
Pedigree

Built on the work that started long short-term memory.

Eldric integrates the xLSTM architecture from the lab of Sepp Hochreiter at JKU Linz — the inventor of LSTM, one of the foundations of modern deep learning. We ship the runtime; we contribute training and integration work; we are honest about which pre-trained weights are public and which are still in distillation.

— On Eldric and the xLSTM architecture

We are an independent platform that runs xLSTM models, classical transformers, GGUF inference, and most other backends you can name. We sit in Vienna and follow the European AI research community closely — every claim on this site you can verify in the public code or in published papers.

Compliance

Made for regulated work.

The science worker ships GLP and FDA 21 CFR Part 11 modes. The whole platform sits on GDPR-compatible infrastructure — by virtue of running in your building.

GDPR data stays on-prem GLP compliance mode FDA 21 CFR Part 11 e-records HIPAA-aware Ed25519 signed licenses Audit ledger hash-chained
Get started

Run it on what you already have.

Eldric installs on a Raspberry Pi 4 for a single-team trial, or across a multi-node cluster for a hospital network. The download is the same; the configuration changes.

Go to downloads Talk to us