bettersorted Logo
Server room with racks – symbolic image for self-hosted AI models
AI Consultancy

Open-source AI models for businesses: the key models at a glance (as of October 2026)

AuthorMuhamed Alahmed
Published on
Reading time7 min
In short

For in-house use in companies, open models with a clear license are particularly suitable: Mistral (France), Qwen (Alibaba), gpt-oss (OpenAI), DeepSeek and the Swiss Apertus. Llama 4 from Meta is effectively excluded for companies based in the EU because of a licensing clause. The decisive factors are the license, German language capabilities, size and hardware requirements.

The most important points at a glance

  • Open models can be run in-house – data does not leave the company.
  • Good options for Germany: Mistral, Qwen, gpt-oss, DeepSeek, Apertus (usually Apache 2.0 or MIT).
  • Llama 4 is not licensed for companies based in the EU.
  • Selection based on license, German language skills, size and hardware.

Whoever wants to run AI models within their own company instead of sending data to ChatGPT & Co. now has a good selection of open models (“open source” or, more precisely, “open weights”). For companies in Germany, as of October 2026, the main families are Mistral, Qwen, gpt-oss, DeepSeek and Apertus interesting. Llama 4 from Meta is, by contrast, practically unusable for companies headquartered in the EU because of a licensing clause. Which model is suitable depends on the license, German language capabilities, size, and existing hardware.

“Open Source” or “Open Weights” – what does that actually mean?

For most “open” AI models, the model weights released – i.e. the trained model that you can download and run on your own servers. The training data and the full training process usually remain confidential. Experts therefore speak of Open Weights. For everyday business use, the key point is: you can run the model yourself, your data does not leave your infrastructure, and you are not dependent on price changes or the shutdown of a cloud service.

The license is important: it determines whether you may use the model commercially, modify it, and integrate it into your own products. Licenses such as Apache 2.0 or MIT are very permissive. By contrast, the manufacturers’ own licenses may contain restrictions – including the exclusion of entire regions.

Team testing software on laptops

The main model families at a glance

Mistral (France)

The French company Mistral AI releases many models under Apache 2.0, including the large Mistral Large 3 (December 2025) and Mistral Small 4 (March 2026), which combines text, images, and multi-step reasoning in one model. Benefits for German companies: European provider, strong German language capabilities, clear licensing, and models in many sizes – from small variants for a single server to large models for data centers.

Qwen (Alibaba)

The Qwen models from China are among the most capable open models and are released in rapid succession (including Qwen3.6 in April and Qwen3.8 in August 2026). The freely downloadable variants are usually available under Apache 2.0; the largest “Max” models, by contrast, are not open. Qwen offers a particularly wide range of sizes, including compact models that run on a single graphics card.

gpt-oss (OpenAI)

With gpt-oss-20b and gpt-oss-120b OpenAI released open models for the first time in years in 2025 – under Apache 2.0. The smaller variant is intended to run on comparatively modest hardware, while the larger one is designed for high-performance servers. For many companies, this is a natural starting point because the behavior and style feel familiar.

DeepSeek (China)

DeepSeek publishes its models under the MIT license and was a pioneer in 2025 with its “reasoning” models. When self-hosted no data flows to China – unlike when using the DeepSeek app or API. Some companies still examine closely how the models behave on politically sensitive topics.

Apertus (Switzerland)

Apertus was released by the Swiss AI Initiative (EPFL, ETH Zurich, CSCS) in September 2025, in sizes of 8 and 70 billion parameters, under Apache 2.0. Its distinguishing feature: it is fully open – including documented training data and methods – and was developed with European law in mind. In terms of performance, the first version trails the leading models; for companies with high transparency requirements, it is nevertheless an interesting option.

Gemma (Google)

Google publishes compact open models with Gemma, designed especially for devices with limited memory. Gemma is available under a proprietary Google license with terms of use — not under Apache 2.0. Before using it, you should review the terms.

Llama 4 (Meta): Note the EU exclusion

Meta releases Llama under its own “Community License”. For Llama 4 excludes companies with the associated usage policyheadquarters in the EU from the rights to the multimodal models — and all Llama 4 models are multimodal. For German companies that want to run a model themselves, Llama 4 is therefore practically not an option. Older, text-only models such as Llama 3.3 are not affected by this, but technically they are no longer state of the art.

What companies should pay attention to when making their selection

  • License: Apache 2.0 or MIT are the most straightforward for companies. Read proprietary vendor licenses carefully.
  • German: Test with real texts from your everyday work – spelling, technical terms, tone.
  • Size and hardware: Larger models are better, but they require significantly more graphics memory. More on this: What hardware does an in-house AI need?.
  • Task: For summaries and document search, medium-sized models are often sufficient; for complex analyses, larger ones are needed.
  • Origin and trust: When self-hosted, no data flows to the vendor – but the model’s behavior on sensitive topics still needs to be checked.
  • Further development: Choose families that are updated regularly, and plan for migration to new versions.

How companies use open models in practice

Hardly any company uses a model “as is.” The typical setup combines it with a user interface such as Open WebUI, a runtime environment such as Ollama and a knowledge base built from its own documents via RAG. This creates an internal AI assistant that accesses your content without data leaving the company. Whether this is worthwhile compared with a cloud service is shown in the comparison Local AI vs. ChatGPT & Co..

For technically interested readers

  • Models are usually provided in the format GGUF (for Ollama/llama.cpp) or as Safetensors (for vLLM) via the Hugging Face platform.
  • Quantization (e.g. 4-bit) significantly reduces a model’s size with only a small loss in quality – often the key to running it on a graphics card.
  • MoE models (Mixture of Experts) have many parameters, but activate only part of them per response – they are faster than their total size suggests, but still require memory for all parameters.

bettersorted relies for its customers on open, transparent components – operated on servers in Germany and bundled in the automaisa Hub. We provide manufacturer-neutral advice, are a BAFA-registered consultant and an authorized INQA coach. The Consulting and guided implementation can be subsidized through INQA-Coaching at 80%; the appropriate path is shown by the Funding Check. For an initial, non-binding consultation: Contact.

Example: Model selection in a social welfare association

An illustrative scenario: A social welfare association wants an internal assistant for concepts, service instructions, and funding guidelines. Health and social data must not leave the organization. Three open-source medium-sized models are tested with 40 real-world questions from daily operations. The criteria are accuracy, clarity in German, and speed on the existing server. The result: a medium-sized model with an Apache-2.0 license, combined with good document search, delivers the best answers — a larger model would have required more expensive hardware without being noticeably better.

Checklist for your model decision

  1. License read and commercial use permitted in the EU?
  2. Tested with 20–50 real tasks in German?
  3. Does the model fit into the available graphics memory — with reserve for multiple users?
  4. Is it clear who is responsible for updates and model changes?
  5. Are data protection and usage rules documented (AI policy)?
  6. Is there a standardized interface so the model remains replaceable later?

Costs and funding

The costs for your own AI model consist of three parts: Setup (planning, installation, integration, testing), ongoing operation (servers or hosting, updates, monitoring, data backup) and Support for employees (training, rules, contacts). We do not quote fixed prices because the scope and starting point vary greatly. The software itself does not incur license costs for open-source solutions.

Funding is not for the technology, but for consulting and guided implementation: through INQA-Coachingthe federal government covers 80% of coaching costs nationwide (up to €11,520, vouchers until 30.06.2028). A preliminary analysis can be subsidized through the BAFA consulting grant – with 80% in the new federal states, Lüneburg and Trier, otherwise 50%, for applications submitted by 31.12.2026. For investments, some federal states offer their own programs – see Funding opportunities for companies in Germany.

Frequently asked questions

Which open-source AI model is best for companies?

There is no single best model for everyone. For German companies, Mistral, Qwen, and gpt-oss are often the first choice because of their Apache 2.0 license and good German language capabilities. Test two or three models with your own tasks.

May I use Llama 4 in Germany?

As a company headquartered in the EU, you do not receive rights to the multimodal Llama 4 models under the usage policy, and all Llama 4 models are multimodal. For self-hosting in Germany, Llama 4 is therefore effectively ruled out.

Are open-source models as good as ChatGPT?

The best open models are close to the leading commercial models for many tasks. For typical business tasks such as summarizing, drafting, and searching your own documents, they are usually sufficient.

What is the difference between Open Source and Open Weights?

With Open Weights, the model weights are freely available, but training data and methods often are not. Fully open models such as Apertus also disclose these. For self-hosting, Open Weights with a suitable license are sufficient.

As of October 2026. Models, versions, and licenses change quickly – before making a decision, check the current license with the provider.

Portrait of Muhamed Alahmed, founder of bettersorted
About the author

Muhamed Alahmed

With over 10 years’ experience in IT, I develop solutions that not only work from a technical perspective, but also create real added value and open up new possibilities.

More about bettersorted →

More articles

Newsletter

Stay up to date on AI topics

Short updates on AI automation, funding programs and new posts — no spam, unsubscribe anytime.

We use Brevo as our marketing platform. By submitting, you agree that your data will be transferred according to Brevo's privacy policy .

CallEmailContact form