A private LLM for your enterprise,hosted in Europe.Your data stays with you.
Open language models for your applications. We confirm availability, API access and operation for your requirements.
On-premise or at European locations: data flows, operators, access rights and contractual terms are agreed before launch.
What a private LLM means for your enterprise
A private LLM is a language model running on your own hardware or in a dedicated European data centre instead of at a US provider. Data flows and control over access depend on architecture, operations, contracts and the model licence.
Private does not necessarily mean your own server room
What matters is control, not where the metal sits. A model in a dedicated data centre under European law serves the same purpose as one in your basement, without you having to buy hardware and hire people to run it. Both count as a private LLM; the difference is effort and capital outlay.
Why an EU endpoint alone is not a private LLM
Many providers advertise a European access point. If the company behind it is subject to US law, cross-border access obligations need assessment, and cutting off access remains a decision outside your influence. The sovereignty section below sets out the four levels and shows exactly where that line runs.
What enterprises actually run on a private LLM
Document analysis on confidential contracts, knowledge search across internal repositories, minutes and summaries from meetings, code assistance on your own source, triage of customer enquiries. The common denominator is always the same: data a company will not, or may not, hand to someone else’s cloud.
A private LLM in action: see the performance live
A short look at sovereign operation: from prompt to answer — entirely on our infrastructure across five European sites.
Open language models: catalogue and reference pricing
The catalogue shows models and reference prices per million tokens. We confirm availability, model version, location and binding terms before activation. Dedicated and on-premise operation is sized for your requirements.
| Model | Class | Context | Locations | Input€ / 1M tokens | Output€ / 1M tokens |
|---|---|---|---|---|---|
| GLM 5.3★ Best seller | All-round | 1M | All 5 locations | 1.40 € | 4.40 € |
| Kimi K3 | All-round | 1M | Berlin · Zurich | 3.00 € | 15.00 € |
| DeepSeek V4 Pro | All-round | 1M | All 5 locations | 1.32 € | 3.96 € |
| Qwen3.5 397B | All-round | 256K | All 5 locations | 0.60 € | 3.60 € |
| Kimi K2.6 | All-round | 256K | Berlin · Zurich | 0.95 € | 4.00 € |
| GLM 5.2 | All-round | 1M | All 5 locations | 1.40 € | 4.40 € |
| GLM 5.1 | All-round | 200K | Frankfurt · Vienna | 1.40 € | 4.40 € |
| MiniMax M3 | All-round | 1M | Vienna · Zurich · Bolzano | 0.30 € | 1.20 € |
| Mistral Large | All-round | 256K | Berlin · Frankfurt · Zurich | 0.50 € | 1.50 € |
| DeepSeek V4 Flash | Efficiency | 1M | All 5 locations | 0.29 € | 0.35 € |
| gpt-oss-120b | Efficiency | 128K | Berlin · Vienna | 0.15 € | 0.60 € |
| Qwen3.6 27B | Efficiency | 256K | All 5 locations | 0.40 € | 2.70 € |
| Qwen3 Coder 480B | Code | 256K | All 5 locations | 0.40 € | 1.60 € |
| Kimi K2.7 Code | Code | 256K | Berlin · Zurich · Bolzano | 1.44 € | 5.18 € |
| Devstral Small | Code | 128K | Vienna · Bolzano | 0.10 € | 0.30 € |
Prices net, as of September 2026, referenced against European providers; billed by actual usage. More models on request — the catalogue grows continuously.
Dedicated or on-premise operation: fixed price based on sizing instead of token billing — talk to us.
Many models, one conductor: the router for your private LLM
Fugi is our orchestration layer on top of the model pool: a small coordinator model receives every request and decides for itself what to do — answer directly, or assemble a team of specialists, hand out sub-tasks, cross-check results and condense them into one answer.
Learned, not hard-wired
Classic routers follow hand-maintained if-then rules. Fugi’s coordinator is itself a trained model: its strategies — who takes what, when to verify, when one answer suffices — emerge from experience and improve with every training run.
The team beats the star
Complex tasks are decomposed and assigned to the strongest specialists — with a verification step before the answer. A federation of open models reaches frontier-level results without needing one giant model.
The pool is yours
New models dock on without retraining, deprecated ones drop out, and your rules apply at system level — such as “only models at DACH locations”. The whole federation runs on our infrastructure across five European sites.
Fugi orchestrates the model pool across all five sites — access comes via the waitlist.
Pick your model — and see the difference
Choose a model, set your monthly volume: the chart plots your monthly output cost against Claude Fable 5 at market API pricing — and how the savings grow with every token.
GLM 5.3 is, per output token, 11.4× cheaper
Comparison basis: Claude Fable 5 at Anthropic’s verified API list price ($10 input / $50 output per 1M tokens, as of September 2026 — the most expensive Anthropic model ever listed; USD taken 1:1 as EUR here). Calculated over output tokens, because that is where the cost lives: Fable 5 always thinks, and thinking tokens are billed as output. User equivalence: around 0.3M output tokens per employee and month — a typical mix of assistant requests, knowledge-base queries (RAG) and agentic software development. Prices net.
A private LLM on par with the US models: proven in the benchmark
Open frontier-class models are no longer a compromise. The head-to-head: GLM-5.3 (open, our best seller) versus Claude Opus 4.8 (US frontier) — measured, not claimed. The gap has closed: in the independent Intelligence Index GLM 5.3 ranks ahead of Opus 4.8, and Kimi K3, #3 worldwide, runs in our catalogue as well.
| GLM-5.3 · open | Opus 4.8 · US | |
|---|---|---|
| Terminal-Bench 2.1agentic | 88.2 | 85.0 |
| Intelligence IndexArtificial Analysis, independent | 60 | 57 |
| HLE with toolsexpert knowledge | 62.5 | 57.9 |
| Cost / 1M outputour price vs. US list price | €4.40 | $25.00 |
| Licence & access | open weights | proprietary · API |
| Operable in the EU | ✓ | ✗ |
5.7× cheaper than Opus 4.8 per output token — and ahead in all three performance comparisons: agentic terminal tasks, the independent Intelligence Index, expert knowledge with tools. The price advantage no longer costs any quality.
Z.ai model card · Artificial Analysis · as of September 2026
GLM 5.3 is at the top of the price table — API access subject to confirmed availability →
Integration example for an OpenAI-compatible API
These examples show how to connect an OpenAI-compatible API. Replace the example address and model identifier with the values confirmed for your access.
from openai import OpenAI
client = OpenAI(
base_url="https://your-endpoint.example/v1",
api_key="IHR_ADVISORI_KEY",
)
antwort = client.chat.completions.create(
model="glm-5.2",
messages=[{"role": "user", "content": "Hallo!"}],
)
print(antwort.choices[0].message.content)The address is an example. You receive the API address, model identifier and credentials after confirmed activation.
Location and hardware matched to your needs
The map illustrates location options in our model catalogue. Availability, hardware, operator, data flows and contractual terms are assessed and confirmed for your request. It is not a live capacity display.
Built by security people — not bolted on
ADVISORI is a security company. Model operations follow the same standards we apply when advising banks and insurers.
Retention by agreement
We agree which data is processed, logged and retained. Requirements for training, subprocessors and deletion are documented before launch.
Protection by configuration
Transport encryption, encryption at rest and access rights are defined and verified for your operating model.
Agree processing terms
Processing locations, responsibilities, technical measures and required data processing agreements are clarified before operation.
Control for your IT
API access, budgets and audit logging are configured for your requirements. The scope is confirmed in the offer.
Multimodal guardrails — your policies, enforced at the endpoint
PII protection, multimodal
Personal data is detected and masked in text, images and documents before it reaches the model — and, on request, in the responses as well.
Prompt-injection defence
Attacks via manipulated inputs, documents or web content are detected and blocked at the endpoint — before they can steer the model.
Company-own policies
Your guidelines as enforceable guardrails: topic blocklists, compliance requirements, department-specific approvals — centrally maintained, effective per team and API key.
Every decision in the audit log
What was checked, masked or blocked is documented traceably — per request, per rule, subject to the agreed logging configuration.
The illustration shows a possible configuration. Filters, logging and detection limits are agreed and tested for your use case. Also available as a standalone service in front of any model: Fugi Guardrails (page in German) →
Evidence and certificates are provided with their applicable scope. This list of topics is not a certification of the service.
What sovereign operation looks like
When Washington pulls the plug: the dependency risk for your private LLM
What was long a hypothetical risk has been documented reality since June 2026 — shutdown, access lists, overnight price changes.
The off switch has been used
No longer hypothetical: in June 2026, Claude Fable 5 was switched off worldwide for 19 days by US government directive — for all customers. No contract, no SLA protected against it.
Access is granted politically
GPT-5.6 launched exclusively for around 20 partners individually approved by Washington. Who gets to use frontier AI is increasingly decided by the US government — European companies are not at the table.
Price changes overnight
The same model that was just included in the subscription costs $10 / $50 per million tokens from July 8 — no grandfathering. Local models flip the equation: fixed infrastructure, plannable costs. The price calculator above shows the difference.
CLOUD Act & structural dependency
US authorities can access data held by US providers — regardless of which data centre it sits in. And according to the first UN AI report, 75% of the world’s AI computing power is located in the US.
GDPR & EU AI Act
Documentation duties, data minimisation, transparency. Easiest to satisfy when the model runs where your data already lives: with you.
- 12 Jun 2026US government orders the shutdown of Claude Fable 5 & Mythos 5 — worldwide, for all users. Source →
- 26 Jun 2026OpenAI launches GPT-5.6 for only ~20 government-vetted partners — access approved customer by customer. Source →
- 01 Jul 2026First UN AI report: 75% of global AI computing power is located in the US, 15% in China. Source →
- 08 Jul 2026Fable 5 back online after 19 days — throttled, and from July 8 only via usage credits ($10 / $50 per million tokens). Source →
- 16 Jul 2026The open world catches up: Moonshot unveils Kimi K3 — ranked #3 worldwide, ahead of Opus 4.8. The weights were released on 27 Jul. Source →
- 28 Aug 2026Z.ai publishes the GLM 5.3 weights: #2 of all open models in the Artificial Analysis Intelligence Index, ahead of Claude Opus 4.8. The model runs in our catalogue. Source →
Digital sovereignty: what the term actually demands
Digital sovereignty is usually debated as a stance and rarely written down as a checklist. For running a private LLM, digital sovereignty can be pinned down precisely: four questions, each with a verifiable answer. Who can switch off your access? Whose law governs your data? Who decides on model changes and pricing? And could you take over operations yourself if you had to?
EU data residency
The data physically sits in Europe, but the provider remains subject to US law. That satisfies storage-location requirements and nothing else, and has nothing to do with digital sovereignty yet: cross-border access obligations require assessment, and shutdown remains someone else’s decision. Operators, subprocessors and contracts also need assessment.
Location settled, control notEuropean operator
Operations and contract sit with a European company, while the model still comes from a US provider on that provider’s terms. A clear step up for digital sovereignty; technically the dependency remains. If the model is switched off or repriced, it hits you just the same.
Law settled, model notOpen weights in a European data centre
The model is open source or available with open weights and runs at a European operator. This reduces dependence on a model API: you hold the weights, and the version stays frozen until you change it. This is where a private LLM genuinely begins, where most enterprises should land, with provisioning agreed individually.
Assess remaining dependenciesFull self-operation
Hardware, model and operations entirely in house. Maximum digital sovereignty, and the only route for classified environments or strict network separation. The price is capital expenditure on hardware, your own specialists, and responsibility for availability and updates.
Fully sovereign, at a costThe decisive line runs between levels two and three: only with open weights does dependency on someone else’s release decision end. Everything below that is data protection, not digital sovereignty. Why a purely European access point does not cross that line is covered in the next section.
An “EU endpoint” is not sovereignty
Many providers advertise “EU data residency” — meaning a gateway in Europe that receives your requests and forwards them to US model providers. What matters is not where the request arrives, but where the model computes.
| US cloud APIOpenAI, Anthropic & co. directly | AI gateway“EU data residency” / EU endpoint | Local modelsADVISORI | |
|---|---|---|---|
| Where does the model compute? | In the US cloud | Still at the US model provider | On your premises or in our data centres (DE/AT/CH/IT) |
| Who sees your prompts? | The model provider | Gateway and model provider | Only you |
| CLOUD Act access possible? | Yes | Yes — at the model operator | Assess operator and legal framework |
| Cost model | API price per token | API price + markup | Subject to offer and capacity |
| Can third parties switch it off? | Yes | Yes | No — you run it |
Private LLM and the GDPR: what self-operation changes
Running a private LLM in your own legal jurisdiction does not turn the GDPR into a formality, but it moves the question from a hard-to-evidence level to an easily evidenced one. Three things change concretely.
The transfer disappears instead of being justified
Calling a US model means justifying a third-country transfer under the GDPR and safeguarding it continuously. If the model runs in a European data centre, the transfer simply does not happen. That is the cleaner route: processing that does not take place needs no justification.
Article 32 asks for effectiveness, not intent
Article 32 GDPR requires technical and organisational measures reflecting the state of the art that ensure confidentiality, integrity, availability and resilience, plus a process for regularly testing their effectiveness. Under self-operation the logs, access records and configuration are yours, and so is the evidence. With an external service you hold the provider’s certificate and little else.
The allocation of roles becomes unambiguous
With us, ADVISORI remains a processor under Article 28 GDPR, with a data processing agreement, a named data centre and no training on your data. Under full self-operation the processor disappears entirely. Either way the chain is short and documented in writing, rather than running through several sub-processors across shifting jurisdictions.
How to set up and document such an operation in detail, from role allocation to deletion policy, is covered in our guide on GDPR-compliant AI (in German). It does not replace an assessment of your specific use case, which is exactly what we have advised on for years.
Open frontier-class models — three operating paths, one accountable partner
We select the right open model for your use cases, build the operation and take responsibility for it — including hardening, updates and monitoring.
On-premise with you
Operation on your own hardware, in your own building. Maximum control — the right choice for classified material, R&D and strictly regulated data.
European data centre
Dedicated operation on NVIDIA B200 infrastructure at one of our five European sites — no hardware investment of your own, under the law of the respective location (DE/AT/CH/IT), with a clear SLA.
Hybrid with failover
Local models as the default, defined fallback paths for peak load — orchestrated via Synthara AI Studio, with rules you define.
Your model speaks your industry
Beyond the catalogue, we fine-tune open frontier-class models on your industry and your documents: domain language, forms, processes, regulation. Training runs on our B200 clusters across five European sites — the fine-tuned model belongs exclusively to you and runs as its own endpoint.
Banks & financial services
Credit files, MaRisk, regulator language — a model that knows your rulebook.
→ Join the waitlistInsurance
Policies, claim files, terms and conditions — trained on your vocabulary.
→ Join the waitlistHealthcare
Doctor’s letters, coding, study protocols — with full data sovereignty.
→ Join the waitlistLegal & compliance
Contracts, briefs, deadlines — precise in legal language.
→ Join the waitlistIndustry & manufacturing
Maintenance reports, standards, bills of materials — your operational knowledge in the model.
→ Join the waitlistPublic sector
Notices, procurement documents, administrative language — sovereign and on the record.
→ Join the waitlistClicking your industry adds it straight to your waitlist request — we follow up with a concrete proposal.
The API is complete. Synthara takes it to your business teams.
Your endpoint works on its own — your developers integrate directly via the API, no extra software needed. If you also want to bring AI to teams without developers, you can optionally put Synthara AI Studio on top of the same sovereign models: AI assistant, no-code studio, workflows and reporting — with automatic failover between your models.
Discover Synthara AI Studio →Request, agree, configure
Describe your needs
Send a non-binding request with your preferred model and use case. A request does not reserve capacity.
Non-binding requestAgree availability and terms
We assess model, location, capacity, pricing and data processing terms individually. Access and delivery dates require explicit confirmation.
After individual assessmentConfigure and test access
Once agreed, you receive credentials and configuration. Test your integration and the agreed data flows before production use.
After confirmed provisioningWhat decision-makers ask us most often
What is local AI?
Are open models really good enough for enterprise use?
Which models are available?
What sets you apart from providers advertising “EU data residency”?
Do we need our own GPU hardware?
What does it cost compared to our current API usage?
How quickly are we productive?
Who operates and maintains the models long term?
Is a data processing agreement (DPA) available?
Can we enforce our own policies (guardrails)?
What availability do you commit to?
Secure the next available slot
Tell us your model, location and requirements. We check availability and respond with next steps. Your enquiry is non-binding.
Request accessQuestions first? Write to us: info@advisori.de · ADVISORI FTC GmbH