Will "Soofi" become Germany's way to AI sovereignty?

Hey guys! When it comes to artificial intelligence, everyone always looks captivated into Silicon Valley to OpenAI (ChatGPT) or Anthropic (Claude). But in the background, there is still a fundamental question for Europe:

How independent are we when the US giants pull the plug tomorrow, multiply prices or throttle access?

Right here comes Soofi Germany's latest project for digital AI sovereignty. (Pronounced: Sovereign OpenSource Foundati(A) But it was hardly The first model in mid-June Soofi-S at the start, There was a lot of headwind in the tech scene: From "Top performance and milestone for Germany as a location" about “Waste of money and beautified performance tests” up to ‘Oh, yeah. So they Volkswagen’d it“For example, on X/Twitter everything was there for the first time.

We took a closer look at the background, deconstructed the technical jargon for you and felt a little bit on the tooth!

Key facts at a glance:

Funding amount Phase 1: ~25 million euros from the Federal Ministry of Economics (BMWE).
Model size Soofi-S: 31.6 billion parameters in total, but only 3.2 billion active per calculation step.
Training data: gigantic 26 trillion tokens (high quality German & English data).
Calculating power: 253,000 GPU hours on Nvidia's B200 chips.
Licence: Apache license (open source for code, weights and data).
Download: via Huggingface.co

What is Soofi and why does the taxpayer pay for it?

The Ministry of Economic Affairs has invested around 25 million euros to bring together a consortium of universities, research institutes (such as Fraunhofer and DFKI) and start-ups (Ellamind, Merantix Momentum).. Around 30 experts are working directly on the project.

The objective: One European Open Source Foundation Model build up. Compared to U.S. providers: Not only the code, but also the model weights and the training data are open. Companies should be able to take this basis and adapt it to their own, highly specialized needs. Just open source and without any sensitive company data landing on American servers!

Technique understandable: How Soofi-S really works

Soofi-S is not a huge, unwieldy mega model, but relies on a clever, modern architectural design.

Tech Talk Translated: Mixture of Experts (MoE) Imagine a huge office building. Instead of a single employee trying to know EVERYTHING perfectly, there are many specialized experts sitting in this office. When a question comes in (e.g. on chemistry), the gatekeeper (the so-called ‘router’) forwards the question exactly to the chemistry expert.

In the case of Soofi-S, this means: Of the 31.6 billion parameters (overall knowledge), only approx. 3.2 billion parameters activated. The idea behind it: This saves a lot of computing power, makes the answers lightning fast and keeps the server costs low!

The basis was Nvidia's model. Nemotron 3 Nano (which critics criticized), but the team didn't just copy the model. You have developed your own training and reinforcement learning concept and 26 trillion tokens fed, including a large chunk of high-quality German specialist and newspaper data (including a license from the Genios archive).

Sounds good at first, so why the shitstorm? Did Soofi cheat on the performance test?

As soon as Soofi-S was released, there were the first dampers. The performance test (benchmark) of Soofi-S had been manipulated!

The problem: Test Data Leakage When an AI is trained and accidentally takes the exam questions from the later performance test (e.g. the renowned GPQA Diamond Test) contained in the training material, this is like a student who gets exactly the solution sheet for memorizing before an exam. The model shines phenomenal in the test, but is not so smart in practice.

What does the doer say? Prof. Dr. Alexander Löser explained Interview with Golem, that an error while downloading an external dataset (TUDA/QA-base from Hugging Face) incorrectly paraphrased exam questions slipped into the training phase. The team made the mistake transparent., completely exclude the GPQA test from the benchmarks and correct the scientific report. Löser emphasizes that the more than 80 other tests were completely clean and the model continues to perform excellently.

The VDI podcast "Technology on the ear" is also recommended here

In the podcast episode 136 Host Christopher Baczyk discusses the question of whether Europe loses its sovereignty with this above-mentioned Professor Löser from the Berlin University of Applied Sciences. If you are wondering why he is the contact person for both Golem and the VDI: He worked on Soofi.

The strategy: "Internet of experts" instead of mega-monoliths

Why don't we just build our own ChatGPT with 1,000 billion parameters in Germany?? Löser brings it to the point pragmatically: For this, Europe simply lacks the billions in budgets and the necessary energy infrastructure for huge data centers!

Europe's recipe for success does not lie in gigantic all-rounder models, but in the Ensemble approach:

Specialized AI experts: A German industrial company adopts a basic Soofi model and trains it with its own expertise (e.g. battery research or mechanical engineering).
The ‘Internet of Experts’: Many specialized AIs network with each other. When a problem needs to be solved (e.g. the development of an e-car), AI experts in chemistry, regulation and mechanics work hand in hand.
Affordable & Safe: Such specialized mini-AIs run on normal company hardware or with European cloud providers.

The hardware hook: How "sovereign" are we really?

Here we must be honest: Soofi-S was in the Industrial AI Cloud by T-Systems in Munich (Link to white paper) on 1,024 Nvidia B200 GPUs trained. Around 19 million euros of the budget went directly into the calculation time.

This shows the biggest weakness of European digital policy: Even if the code and the data to 100 % are European, Without U.S. hardware from Nvidia, almost nothing is currently running.. Should the United States, under current or future government, Restricting the export of AI hardware to Europe, If we were in the rain. Own European chip projects such as Vsora or Dare Unfortunately, they are still at the beginning.

Since the project is currently still in beta stage and the test phase hosted by Telekom is running, the Final classification under licensing law so far open, Even if the Open Source AI Definition 1.0 is fully fulfilled, the use in prductive systems is still strongly discouraged.

Outlook: What's next?

From Q4 2026: The improved and finely tuned versions of Soofi-S (including instructing and reasoning models such as Isar and Rhine) should be completely free on Hugging Face. Legal notice:
Practical guide for companies: By the end of 2026, the team also wants to publish instructions on how small and medium-sized companies can prepare their data appropriately in order to build their very own Soofi experts.
Soofi Phase 2 (possibly from 2027): A second funding phase is planned with 30 to 40 million euros. This is intended to be an even larger MoE model with 300 to 800 billion parameters be built.

Try SOOFI-S Beta for yourself

Soofi-S locally with LM Studio for beginners (or vLLM if you are not afraid of command lines)

You want the German open source model Soofi-S Run on your own hardware? Not like the usual suspects, but completely without cloud, without subscription and with 100 % Data protection? With appropriate hardware and LM Studio (GGUF versions) or vLLM (BF8, EntQuant, …) can do this in just a few clicks/commands!

1. Hardware check (before you start)

Memory (RAM/VRAM): For the quantized version of Soofi-S, usually enough is enough for you 24 GB RAM (better: VRAM on your graphics card, otherwise your hoof will fall asleep).
Soofi S is a Mixture-of-Experts model with 31.6 billion parameters in total, of which only around 3.2 billion are activated per token. Nevertheless, we must all Weights are in memory.
Depending on the quantization you need according to the official GGUF page: Q4 _K _M ≈ 21 GB, Q5 _K _M ≈ 25-26 GB (recommended standard), Q8 _0 ≈ 32-34 GB. Plan for at least 32 GB of RAM (or VRAM/Unified Memory on Mac). Apple silicon Macs (M2/M3/M4 Pro/Max with sufficient Unified Memory) are well suited because of the MoE architecture, as are Windows/Linux PCs with powerful GPU + a lot of VRAM
Storage space: Hold also approx. 5 to 10 GB free NvME/SSD memory ready.

2. Prepare LM Studio

Download: If you haven't already, get LM Studio from lmstudio.ai and install it. Please update old versions and in addition in the settings (developer mode) the Update Runtime/Engine: Soofi S uses a custom architecture (‘nemotron _h _moe’) that older llama.cpp/LM-Studio versions did not yet know (error message ‘unknown model architecture’). This support was only retrofitted in the course of 2026, without current runtime the model does not load.
Launch the app: Open the application and switch to the left sidebar "Discover" tab (Lupe or abbreviation Ctrl + 2 / ⁇ + 2).

3. Find Soofi-S on Hugging Face & download

Use search bar: Enter above in the search Soofi-S GGUF one. (Alternatively: Copy directly the Hugging Face URL of the Soofi-S GGUF version to the search bar).
Choosing the correct quantization (GGUF file):
- Recommendation for getting started: Choose a version like Q4_K_M Then you may reach your destination with 24GB (V)RAM, albeit painfully slowly.
Start the download: Click next to your desired file ‘Download’. You can see the progress in the tab Downloads.

4. Load model & loose shadow

Select model: Switch to Chat tab (Speak bubble icon on the left).
Load into memory: Click on the drop-down bar at the top (‘Select a model to load’) and choose your downloaded Soofi-S GGUF model off.
Start Chat: As soon as the loading bar is green, Soofi-S is ready for use! Ask the AI questions in German or let it solve tasks, as I said, that now runs completely offline on your computer.

Pro tips for power users:

Offloading to GPU: If you have an NVIDIA/AMD graphics card or an Apple M chip, turn on the GPU Offloading on/high to tickle out maximum generation speed (tokens/second).
Local API server (Optional): Use the developer mode of LM Studio Local Server-Tab. LM Studio provides you with an OpenAI-compatible interface (http://localhost:1234/v1) ready to integrate Soofi-S into your own scripts, browser extensions or workflows!
All EntQuant variants are still vLLM-only, Docker + Linux (if applicable WSL2 on Windows) do not need LM Studio.

TL:DR

Is Soofi The immediate rescue of our AI future? No, the hardware dependency on the USA is still too big for that. But it is an extremely important and thoughtful step!

The approach of not stubbornly copying the US concept of gigantic, power-eating ‘monolithic AIs’, but a Networked ecosystem of specialized open source models It is a perfect fit for European industry. If Phase 2 delivers and the companies move in, Soofi could become exactly the niche specialist Europe needs!

Will "Soofi" become Germany's way to AI sovereignty or is it just a beautiful illusion?