There is no single "best" open-source LLM. There is the one that fits your real constraints: the licence under which you are allowed to use it, the hardware you have to run it, its quality in your language, and the level of sovereignty you are aiming for. This guide gives the method to choose and a comparison of the main model families, current as of mid-2026. One caveat upfront: this landscape moves fast, a new model comes out almost every month, so reason about durable criteria rather than today's ranking.
Most models called "open source" are in fact open-weights models: the publisher releases the model weights, which you can download, run and fine-tune on your own machine, but not necessarily the training data or code. Fully "open source", with published training data, remains rare.
For a business, this nuance has two practical consequences, and they are the only ones that really matter: the licence (what you are legally allowed to do with the model) and the ability to run the weights locally (which keeps your data with you). The rest is a vocabulary debate.
The licence. It is the first filter, and the most overlooked. Two families of licences are genuinely permissive for commercial use: Apache 2.0 (Mistral, Qwen, Granite) and MIT (DeepSeek, Phi). You do what you want with them, including in production, with no royalty. Conversely, some so-called open licences are not open in the strict sense: Llama 4's community licence imposes a user cap and explicitly excludes companies domiciled in the European Union from its multimodal models, which is the entire Llama 4 range, and Google's Gemma licence has its own usage terms. These models remain usable, but read the terms before you build on them.
Size against your hardware. An LLM is chosen first by what your server can load. Open-source "frontier" models now rival proprietary ones, but the largest demand several 80 GB video-memory cards: that is not an SME server. The right approach is the opposite of the usual reflex: start from the model that fits your machine, quantised to 4 bits if needed, not from the highest benchmark score.
Language. If your use cases are in French, or any language other than English, quality in that language matters more than an English-language ranking. Mistral's models, trained in France, are strong in French; Qwen's cover a very wide range of languages.
Sovereignty. This is the point that gives the whole exercise its meaning. An open-weights model you run on your own infrastructure sends no data out. The same model called through a third party's API sends your requests to that third party. Open source is only sovereign if it runs on your premises.
| family | publisher | licence | sizes | best for |
|---|---|---|---|---|
| Mistral 3 / Ministral 3 | Mistral AI (FR) | Apache 2.0 | 3B to 14B, + Large 3 (MoE) | European SME: strong in French, clean licence, from edge to server |
| Qwen 3.x | Alibaba | Apache 2.0 | 0.6B to 235B (MoE) | versatility, wide choice of sizes, multilingual |
| Llama 4 | Meta | Community licence (EU excluded from multimodal) | MoE models, very long context | mature ecosystem and tooling, provided you read the licence |
| Gemma 4 | Gemma licence | edge (E2B/E4B) to 31B | small hardware, strong memory efficiency | |
| DeepSeek V4 / R1 | DeepSeek | MIT | distilled to giant MoE (V4: 1M context) | reasoning and code, good performance/cost ratio |
| Phi-4 | Microsoft | MIT | ~14B | small fast model, CPU and edge |
| Granite 4.0 | IBM | Apache 2.0 | small to medium | enterprise compliance, reduced memory footprint |
Version numbers change month to month; what to remember is the publisher and the licence, which are stable and decisive.
One reservation, drawn from our own practice: agentic use. As soon as the AI doesn't just answer but calls tools (searching your documents, writing a file, booking an appointment), the reliability of those calls becomes the criterion that outweighs all others. On that ground, our tests with Gungnir lead us to advise against the current Mistral models: perfectly capable in dialogue and writing, they fall short at tool handling. If your use case is agentic, test that specific point before deciding, whatever the model.
A word on open-source "frontier" models, the ones that rival the best proprietary models: they do exist, but they are mixture-of-experts architectures of several hundred billion parameters that demand serious hardware. For an SME, they belong to an occasionally rented server, not to a daily installation.
Choosing the model is half the question. The other half is where it runs. An excellent open-source model called through a remote API does not make you sovereign: your data leaves anyway. What makes the whole thing sovereign is running the model on your own infrastructure.
That is exactly the role of a model-agnostic platform like Gungnir: you plug in the open-source model of your choice, you switch when a better one comes out, and the application, the history and your knowledge base stay with you. The model becomes an interchangeable part, not a point of dependency. It is the same logic detailed in our guide to sovereign AI for SMEs, and it is what separates a real ChatGPT alternative without the cloud from a mere change of provider.
An open-weights model publishes the weights you can download, run and fine-tune, but rarely the training data and code. Fully "open source", with published data, remains rare. For a business, what matters in practice is not the label but the licence (what you are allowed to do) and the ability to run the model locally.
It depends entirely on the model size. A small model (around 3 to 14 billion parameters), quantised to 4 bits, runs on a modest server, a small GPU, sometimes on CPU. "Frontier" models of several hundred billion parameters, on the other hand, need several professional 80 GB cards. The rule: choose the model to fit the machine, not the other way round.
Not any longer in 2026. The best open-source models rival proprietary ones on reasoning, code and multilingual tasks. For business use in a company (answering from your documents, writing, summarising), a mid-size open-source model is more than enough, and it has the decisive advantage of being able to run on your premises.
Yes for most, but read the licence. Models under Apache 2.0 (Mistral, Qwen, Granite) and MIT (DeepSeek, Phi) allow commercial use with no particular condition. Google's Gemma licence allows it under its own terms. Llama 4's community licence, however, explicitly excludes companies domiciled in the European Union from its multimodal models (the entire range is multimodal); the text-only Llama 3 models carry no such exclusion. Always check the licence before building on a model.
No, but if your use cases are in French, quality in the language matters more than a good score on an English-language ranking. Mistral's models, trained in France, are a solid starting point. A highly multilingual model like Qwen is also a good option. The key is to test on your real use cases, in your language, before deciding.
