What is an open-source LLM?
An open-source LLM is a large language model whose weights (the "brain" of the model) you can download and run yourself. You put it on your own hardware or with a host of your choice and pay the maker no fee per question. Well-known families are Mistral, Llama, Gemma, Qwen and DeepSeek.
Closed models such as GPT, Claude and Gemini are only available through the maker's app or API, paid per use and under the maker's terms.
One nuance about the name: for most of these models only the weights are released. The training data and training code stay with the maker. "Open weights" is therefore the more precise term. The license determines what you may do with the model, and it differs per family.
Why do businesses choose an open-source LLM?
Businesses choose an open-source LLM for four reasons: control over data, independence from a single vendor, predictable costs and the option to adapt the model.
- Data stays inside: you can run the model locally, so customer data and documents never leave your own environment
- Less dependency: a model you have downloaded keeps working, even if the maker changes prices or terms for new versions
- Predictable costs: you pay for hardware, power and maintenance, and no fee per question
- Adaptable: you can further train an open model on your own documents and jargon
Which open-source LLMs are there in 2026?
The five families you come across most often in 2026 are Mistral, Llama, Gemma, Qwen and DeepSeek. They differ in origin and in license. This overview reflects the situation in October 2026.
| Family | Maker | Country | License |
|---|---|---|---|
| Mistral | Mistral AI | France | Apache 2.0 for Mistral Large 3 and Ministral 3 |
| Llama | Meta | United States | Llama 4 Community License (Meta's own license) |
| Gemma | United States | Apache 2.0 since Gemma 4 | |
| Qwen | Alibaba Cloud | China | Apache 2.0 for the open models |
| DeepSeek | DeepSeek | China | MIT for DeepSeek V4 |
Mistral (France). Mistral AI is a French company. Mistral Large 3 and the smaller Ministral 3 models (3B, 8B and 14B) were released on 2 December 2025 under Apache 2.0. For organizations that prefer to buy from a European vendor, Mistral is the first candidate.
Llama (Meta, US). Llama comes under Meta's own license, the Llama Community License. Commercial use is allowed. Only parties with more than 700 million monthly users have to request a separate license from Meta.
Gemma (Google, US). Gemma 4 was released on 2 April 2026 in four sizes, from models for phones and laptops to models for workstations. According to Google, Gemma 4 is trained on more than 140 languages. It is the first Gemma generation under Apache 2.0.
Qwen (Alibaba Cloud, China). The Qwen family has sizes for every hardware profile. The Qwen team releases its open models under Apache 2.0.
DeepSeek (China). DeepSeek releases the V4 models under the MIT license, one of the most permissive licenses available.
For Qwen and DeepSeek, some organizations take the Chinese origin into account, for example because of procurement policy or client requirements. If you run the model locally, your data stays in your own environment either way.
Open or closed model: when do you choose which?
Choose an open model if your data has to stay inside your own environment or if you use the AI intensively every day. Choose a closed model if you need the highest quality for complex analysis or want to start quickly without your own hardware.
| Situation | Logical choice |
|---|---|
| Sensitive data (healthcare, legal, finance) | Open model, local |
| Occasional use without sensitive data | Closed model via the cloud |
| Intensive daily use | Run the numbers on an open model on your own hardware |
| Complex analysis where quality comes first | Closed top model |
| Staying independent of a single vendor | Open model |
Combining both is also possible: sensitive processes on an open model, generic work in the cloud. How to set up cloud AI carefully is covered in using AI in a GDPR-proof way.
How do you choose an open-source LLM for your business?
You choose an open-source LLM based on four questions: the task, the language, the hardware and the requirements of your sector.
1. What does the model need to do? Answering email and summarizing documents asks less of a model than legal analysis. 2. In which language? Always test performance in your own language with your own material. Benchmarks are almost always in English and say little about your quotes and customer emails. 3. On which hardware? Larger models need more memory and a heavier machine. Start with a mid-size model and scale up if the quality falls short. 4. What does your sector require? Check the license and your procurement policy. If you want a European vendor, you end up with Mistral.
This choice is reversible. A clean implementation separates the agent (your processes, integrations and knowledge) from the model underneath, so you can switch models later without starting over.
What are the risks of an open-source LLM?
An open-source LLM requires your own maintenance. Four points belong on your list:
- Updates: models and the software around them get new versions; agree on who keeps track of them
- Security: a local model is as secure as the server it runs on
- Quality control: test at go-live and at every model switch with your own real-world cases
- License: read the license of the exact model and the exact version you use
These are exactly the points we arrange in an implementation on your own infrastructure, including knowledge transfer so you stay in control yourself. That way an open model becomes the engine behind your AI automation.
Frequently asked questions
What is the difference between an open-source LLM and ChatGPT? You download an open-source LLM and run it on your own hardware. ChatGPT is a service from OpenAI that you use through an app or API, with your questions processed on OpenAI's servers.
Can I use an open-source LLM commercially? Under Apache 2.0 and MIT you can. As of October 2026 those licenses apply to Gemma 4, the open Qwen models, Mistral Large 3 and DeepSeek V4. Llama allows commercial use under Meta's own license. Always check the license of the exact version.
What is the best open-source model for Dutch? That differs per release and per task. Google reports that Gemma 4 is trained on more than 140 languages, and Mistral is a European maker. Test two or three models with your own Dutch documents and compare the results.
Can I train an open-source LLM on my own data? Yes. The light option connects your documents as a knowledge base, the heavy option is fine-tuning on your own texts. Start with the knowledge base: it is faster to set up and easier to update.
Do I need my own server for an open-source LLM? For business use with several users, yes. The smallest models run on a laptop, which is enough for testing. Read more about AI on your own server.
Getting started yourself?
Read how AI on your own server works in practice, look at our AI agent implementation or schedule a no-obligation call. Want to build your own AI skills first? That is possible with a 1-on-1 AI training of 3 hours, from €449 excl. VAT.
