What is an open-source LLM?
An open-source LLM is a large language model whose weights (the "brain" of the model) are freely downloadable and usable. You run it on your own hardware or at a host of your choice, without paying a US provider per question. Well-known examples are Mistral, Llama, Gemma, Qwen and DeepSeek.
The contrast with closed models (GPT, Claude, Gemini) is simple: those you only use via the maker's API or app, paying per use, under conditions the maker can change unilaterally.
Why open-source is interesting for business
- Data stays inside: you can run the model locally, so your customer data and documents never leave your environment
- No vendor lock-in: the model is yours; nobody can raise the price, change the terms or revoke your access
- Predictable costs: one-time hardware instead of ongoing token costs
- Adaptable: you can tune an open model on your own documents, tone of voice and jargon
Since the recent export restrictions on US frontier models, that independence is no longer a theoretical advantage. Businesses that built their core process on an external API discovered that access to AI has become a geopolitical variable.
The major families at a glance
Mistral (France). The European flagship. Strong multilingual performance, including Dutch, and as an EU company a logical choice for organizations that want to buy European for compliance reasons. Various sizes, from compact to large.
Llama (Meta, US). The best-known open family with the largest ecosystem. Lots of tooling, lots of available knowledge, broad hardware support. A reliable starting point for general business tasks.
Gemma (Google, US). Compact models that deliver a lot of quality on modest hardware. Interesting if you want to run on a lighter machine.
Qwen (Alibaba, China). Technically strong family with good multilingual performance and sizes for every hardware profile. For some organizations the origin plays a role in the decision; technically they belong to the top of open models.
DeepSeek (China). Got a lot of attention for strong reasoning performance at low training cost. Same consideration as Qwen: technically interesting, weigh the origin.
Open vs closed: when to choose what?
| Situation | Best choice |
|---|---|
| Sensitive data (healthcare, legal, finance) | Open-source, local |
| Incidental use, no sensitive data | Cloud is fine |
| Intensive daily use | Open-source wins on cost |
| Absolute top quality for complex analysis | Closed frontier model |
| Independence from external parties | Open-source |
How do you choose the right model?
Four questions determine the choice:
1. What does the model need to do? Answering mail and summarizing documents demands less than complex legal analysis. 2. In which language? Always test Dutch performance with your own material; differences between families are bigger here than English benchmarks suggest. 3. On which hardware? Larger models are better but require heavier (more expensive) machines. A mid-size model is often the best compromise. 4. Which requirements does your sector set? For some organizations a non-European origin is a dealbreaker, and then Mistral is the logical candidate.
The good news: this choice is reversible. A proper implementation separates the agent (your processes, connections and knowledge) from the model underneath, so you can switch later without starting over.
Risks and how to cover them
Open-source does not mean carefree. Three points of attention:
- Updates: models and runners get updates you need to keep up with; agree on who does that
- Security: a local model is as safe as the server it runs on
- Quality control: test at go-live and at every model switch with your own real-world cases
These are exactly the points we handle in an implementation on your own infrastructure, including knowledge transfer so you keep a grip on it yourself. That is how an open model becomes the engine behind your AI automation.
Frequently asked questions
Are open-source models good enough for serious work? Yes. For most business tasks (customer questions, document processing, reports, email) modern open models perform comparably to closed models of a generation earlier, and that is plenty. Only in the most complex reasoning tasks do you still notice a clear difference.
Can I use open-source models commercially? Usually yes, but licenses differ per family and sometimes per model size. Check the license before you build; with an implementation by us that check is part of the project.
What is the best open-source model for Dutch? That shifts per release. Mistral and Qwen traditionally score well on multilingual tasks, but always test with your own material. During implementation we test multiple models on your real-world cases.
Can I train an open-source model on my own data? Yes, from light (connecting your documents as a knowledge base) to heavy (fine-tuning on your tone of voice). For most businesses the light variant is sufficient and much cheaper.
Getting started?
Read how AI on your own server works in practice, check our OpenClaw implementation or book a free intro call. Want to upskill your team first? Our 1-on-1 AI training covers it.
