[email protected]

Llama Hosting in Germany: Run Llama 3.3, Understand Llama 4 in the EU

You want to use a Llama model on your own infrastructure. We explain which Llama version a company in the EU can use under the licence, and run Llama 3.3 70B on a dedicated GPU server in Germany.

  • Operated in a German data centre
  • Operation under a DPA
  • vLLM and OpenAI-compatible API
Reviews
Cloud Wolke HerausforderungServer NachhaltigkeitIT Beratung Service Consulting SoforthilfeTimo Wevelsiep Robin ZinsExperten für Innovation Migration AWSHetzner Hosting zuverlässig

Customers from mid-sized business, industry and the public sector, in Germany and internationally

  • Stadtwerke Brühl
  • DGHO e.V.
  • solarwerkstatt berlin
  • Krüger Dirndl
  • ABCO Water Systems
  • Golem.de
  • EVADXB
  • nextGYM
  • AInergy
  • ml&s
  • Odiseo Solutions
  • Annota
  • ARGE
  • SweetConnect GmbH
  • Aphy AG
  • CORGOS
  • Rekorder
  • SolidProof
  • Yonju
  • Keymate
  • Paritel
  • Mr. Clipart
  • Millenium
  • Negosh
  • Führerscheinmacher
  • BTI
  • Deutsche Gesellschaft für Datenschutz
Read client reviews

Starting point

Why run Llama yourself

Llama is Meta's open-weight family. The latest generation, Llama 4 (Scout and Maverick, April 2025), consists of multimodal mixture-of-experts models. The Llama 4 Acceptable Use Policy does not grant the rights under Section 1(a) of the licence for these multimodal models to companies with their principal place of business in the EU. The only exception is end users of a product or service that incorporates the models. We therefore do not offer Llama 4. Llama 3.3 70B Instruct can be operated: a text-only model from December 2024 under the Llama 3.3 Community License. Since April 2025, Meta has not published new model weights in the meta-llama Hugging Face account. If you are looking for a current model, Qwen, gpt-oss, Gemma and Mistral offer families with newer checkpoints under Apache 2.0.

  • Data stays in Germany

    The model runs on a dedicated server in a German data centre. Under a data processing agreement, with no data path to a model API.

  • Predictable costs

    A fixed monthly price per server tier instead of billing per consumed token. More requests do not increase the invoice.

  • Model version under control

    Checkpoint and quantisation stay until you agree to a change. No silent model update by a provider.

Model facts

Llama checkpoints at a glance

Vendor specifications and weight file size per format, with source.

CheckpointReleasedParametersContext (vendor)Weights per format
meta-llama/Llama-3.3-70B-Instruct12/202470.6B (dense, text only)131,072 tokensBF16: 141.1 GB
FP8: 72.7 GB
meta-llama/Llama-4-Scout-17B-16E-Instruct04/2025109B total / 17B active (MoE, multimodal)10M tokensBF16: 217.3 GB
meta-llama/Llama-4-Maverick-17B-128E-Instruct04/2025400B total / 17B active (MoE, multimodal)1M tokensBF16: 803.2 GB
FP8: 416.8 GB
  • GB = 10⁹ bytes, sum of the weight files in the listed Hugging Face repository. Operation additionally needs memory for KV cache, runtime and image processing where applicable. As of October 2026.

Licence

Llama: licence and commercial use

Licence

Llama 3.3 Community License + Acceptable Use Policy (Licence text, Llama 3.3 Acceptable Use Policy)

  • Commercial use: Yes, for Llama 3.3 70B Instruct. Anyone whose products or services had more than 700 million monthly active users on the release date needs a separate licence from Meta.
  • Obligations and thresholds: Anyone making the model or a service with it available includes the licence and prominently displays “Built with Llama” on a website, user interface or in the documentation. The Acceptable Use Policy must be followed. Its EU clause concerns multimodal models; Llama 3.3 70B Instruct is a text-only model.
  • Checked: As of October 2026

Not offered

  • Llama 4 Scout and Llama 4 Maverick are multimodal models. The Llama 4 Acceptable Use Policy does not grant the rights under Section 1(a) of the Llama 4 Community License to individuals domiciled in, or companies with their principal place of business in, the EU. The exception only covers end users of a product or service that incorporates the models. As a company based in Germany, we therefore do not offer Llama 4.

Technical classification, not legal advice. The licence text of the deployed model version is authoritative. Open LLM licences compared

Server tier

Minimum recommended server tier for Llama

The recommendation follows from the weight size plus headroom for context and runtime.

ModelFormatWeightsMinimum recommendedNote
Llama 3.3 70BFP872.7 GBManaged GPU Server 96On 96 GB, around 23 GB remain for context and runtime after the weights. This limits context length and concurrent requests.
Llama 3.3 70BBF16141.1 GBManaged GPU Server 192–
Llama 4 ScoutBF16217.3 GBNone of our tiersNot offered: no rights under Section 1(a) for companies with their principal place of business in the EU.
Llama 4 MaverickFP8416.8 GBNone of our tiersNot offered: EU exclusion in the licence; the weights also exceed our largest tier of 384 GB.
  • Which tier suits your use case depends on context length and concurrent requests. We check this before the proposal.
  • Meta publishes Llama 3.3 70B in BF16 only. The FP8 size comes from Red Hat AI's checkpoint in compressed-tensors format, licensed under the Llama 3.3 Community License. NVIDIA's FP8 and NVFP4 checkpoints are additionally subject to the NVIDIA Open Model License.

Pricing

The Managed GPU Server tiers

Managed GPU Server 24

1 × RTX PRO 4000 Blackwell, 24 GB GPU memory

€699 excl. VAT / monthcancellable monthly€499 excl. VAT one-time setup

Relevant for this family

Managed GPU Server 96

1 × RTX PRO 6000 Blackwell Max-Q, 96 GB GPU memory

€2,199 excl. VAT / monthcancellable monthly€1,999 excl. VAT one-time setup

Relevant for this family

Managed GPU Server 192

2 × RTX PRO 6000 Blackwell Max-Q, 192 GB GPU memory

Price and term on request

Managed GPU Server 384

4 × RTX PRO 6000 Blackwell Max-Q, 384 GB GPU memory

Price and term on request

The Managed GPU Server 288 with three GPUs is intended for several models side by side, because vLLM only splits a model when the attention heads are divisible by the number of GPUs.

All prices are net and exclude statutory VAT. The offers are addressed to businesses.

Payment
  • SEPA direct debit
  • Visa
  • Mastercard
  • American Express
  • Apple Pay
  • Google Pay
  • PayPal
  • Bank transfer

vLLM

Running Llama with vLLM

Architecture, quantisation and distribution across several GPUs.

Architecture

Llama 3.3 70B uses LlamaForCausalLM, one of the longest-supported architectures in vLLM. No additional package is required for serving.

Quantisation

Meta publishes BF16 weights only. For the Managed GPU Server 96 we use an FP8 checkpoint in compressed-tensors format. Before the proposal we check the origin, licence and answer quality of the quantised checkpoint.

Tensor parallelism

With 64 attention heads and 8 KV heads, Llama 3.3 70B can be split across two or four GPUs. For BF16 at around 141 GB we recommend at least two RTX PRO 6000.

Context

Llama 3.3 supports 131,072 tokens. In FP8 on a single 96 GB GPU little memory remains for the KV cache, so the configured context length is shorter there. We set the value based on your requests and state it in the proposal.

Operated by WZ-IT

What operation includes

The Managed GPU Server is an operated model environment, not an empty server.

View Managed GPU Server
Model and API
  • Model and vLLM

    An agreed model in the agreed quantisation, served through a managed vLLM inference layer.

  • OpenAI-compatible API

    Applications and coding clients connect to the server via base URL and API key.

Interface and monitoring
  • Open WebUI

    A managed chat interface for teams using the model without their own application.

  • 24/7 monitoring

    Host, GPU, vLLM and Open WebUI are monitored proactively; incidents are handled according to the service level.

Maintenance and contract
  • Updates with CVE assessment

    Operating system, drivers, vLLM and Open WebUI are reviewed and updated in a controlled way. Model changes only by agreement.

  • DPA and documentation

    Data processing agreement, documented configuration and a personal point of contact.

Scope, prices and multi-GPU servers are on the product page.

Frequently asked questions about Llama hosting

Licence, server tier, data path and operation

Reviews & projects

What clients say about working with us

From a knowledge base with an AI assistant and local AI integration to data sovereignty and ongoing operations.

AI Platform

We were looking for a secure yet easy-to-use alternative to the large American providers and came across WZ-IT. WZ-IT advised us on choosing the right models and implemented the integration for us. Beforehand, we were able to try everything out calmly on a test instance. Today WZ-IT runs Open WebUI for us on our own domain, with a convenient Microsoft login and open-source models hosted by a German provider. Our team can now work productively with AI without resorting to external tools or shadow AI. Working together is straightforward, and questions are answered quickly and clearly. With WZ-IT we also have a partner at our side for further AI projects. (translated)
Carl-Frederic KornHead of E-Commerce & Online Marketing, Krüger Dirndl GmbH
See the AI platform

Knowledge Management & AI

The knowledge from many years in the solar industry lived in the head of our former managing director, and it still lives in the heads of our employees today. We wanted to secure it, not as a folder full of notes, but in a way that lets us ask questions in daily work and get reliable answers. WZ-IT advised us from the start and weighed different approaches and knowledge management systems against each other with us before we decided. What was built is a custom knowledge base with an AI assistant that brings together conversations, notes, documentation and manuals: from the platform through onboarding our team to support in operation. Conversations are transcribed and facts are extracted; we review and correct everything ourselves before it is stored, because we do not want to trust an AI blindly. The system runs in our own environment, and the data stays with us. What surprised us was how efficient it was: WZ-IT used existing open-source software and extended it specifically for our requirements, instead of selling an expensive custom build from scratch, as is often usual with IT service providers. Timo and Robin understood our actual problem, preserving knowledge during the generational handover, and implemented it step by step: from the proof of concept through setup and validation with our own data to productive operation, explaining everything clearly along the way. Next, we want to extend this to the knowledge of our employees, so that this does not depend on individual people either.
Joshua von AmelnManaging Director, solarwerkstatt berlin GmbH
See knowledge transfer

Read client reviews

International

Selected projects

WZ-IT designs, develops and operates infrastructure and software for clients in Germany and internationally. We deliver projects remotely and continue supporting them in ongoing operations after go-live.

Model assessment

Have your model and server tier assessed

Name the model, context length and concurrent requests. We check checkpoint, licence and server tier before the proposal.

  • You speak directly with Timo and Robin, who are technically responsible
  • A professional assessment of your situation
  • Concrete next steps for infrastructure, software or AI

The conversation is free and non-binding.

WZ-IT's advice on our Azure migration was technically sound and completely non-binding right from the intro call - we took away a great deal.
Jakob ÖschlbergerInno7 GmbH
Timo and Robin, founders of WZ-IT

Which server tier should be assessed?

The selection is non-binding. We confirm model fit, licence and term before a proposal.

We usually respond within one business day. Please do not send access credentials yet.

WZ-IT's advice on our Azure migration was technically sound and completely non-binding right from the intro call - we took away a great deal.
Jakob ÖschlbergerInno7 GmbH