Best open-source LLMs in 2026: the licence-first guide

In practice there are three licence families to separate:

• MIT and Apache 2.0 — permissive, well known, understood by every legal department. Commercial use, modification and redistribution with no surprises. That covers gpt-oss-120B, Qwen3 235B and Mistral Small 3.2 (Apache 2.0), plus DeepSeek V3.2 and R1, GLM 4.6 and Apriel 1.5 (MIT).
• Proprietary community licences — Meta's Llama licence, for example. Commercially usable in many cases, but with specific conditions you have to read. It is not open source in the OSI sense, whatever the marketing says.
• "Other" — what Hugging Face displays when the licence matches no standard template. That is the case for Kimi K2 and Nemotron Super. It does not mean "forbidden"; it means "read the text before committing".

Method: every row in the table below comes from the public Hugging Face API, queried on 10 August 2026 — repository existence, licence declared in the model card, downloads over the last month. Re-check before deploying: licences are occasionally revised.

Download volume is not a quality measure — it is an adoption measure, and therefore an indirect measure of how much documentation, feedback and tooling you will find around the model. On a real project that often matters more than a benchmark point.

gpt-oss-120B dominates adoption with over four million monthly downloads, far ahead of everything else, under Apache 2.0. Mixture-of-experts architecture: 117 billion parameters in total but roughly 5 billion active per token, which makes it far cheaper to serve than its nominal size suggests.

Mistral Small 3.2 24B is the pragmatic European choice and the most downloaded after gpt-oss: 24 billion parameters under Apache 2.0, which means a model you can genuinely self-host without building a data centre. For an SME wanting sovereignty without an infrastructure budget, this is the default starting point.

DeepSeek V3.2 and R1 are both under MIT, the simplest licence there is. R1 targets reasoning, V3.2 is an experimental release aimed at reduced compute cost. Their licence is a serious argument: few models at this level are this free to use.

Qwen3 235B A22B and GLM 4.6 cover the high-end generalist tier, one under Apache 2.0 and one under MIT. Qwen is particularly strong multilingually, which matters for European use.

Apriel 1.5 15B Thinker is the smallest on the list and the least downloaded (222 last month) — so the least documented. But 15 billion parameters under MIT with a reasoning orientation is exactly the interesting profile for a constrained deployment. Worth testing, not worth deploying blind.

The three non-permissive ones. Llama 4 Scout (Meta licence) offers very long context and multimodality, with strong adoption. Kimi K2 Instruct (one trillion parameters, 32 billion active) is built for agents and tool calling. Nemotron Super 49B is optimised for NVIDIA hardware. All three are good models; all three require a licence read before commitment. This is not administrative trivia: it is the kind of thing that stops a project during legal validation, six weeks into development.

• Can your data leave the European Union? If not, the subject is not the model but the hosting — and self-hosting a reasonably sized model becomes the only serious option.
• What is the smallest model that solves your case? Start there. A 15-to-24-billion parameter model covers far more use cases than the parameter race suggests.
• Will your legal team accept the licence? Ask before development. MIT and Apache 2.0 pass without discussion; "other" requires a read.
• Who will maintain the deployment? A self-hosted model is a production system: updates, monitoring, scaling. It is a cost transfer, not a saving.

For the comparison with commercial models, see our analysis of Claude vs ChatGPT for the enterprise and our guide to Claude API pricing. For the frameworks that orchestrate these models, see the best open-source AI agent frameworks.