In June 2026, two announcements put open source AI models back at the centre of the conversation for SMEs: the release of GLM-5.2 under the MIT licence and the availability of Mistral OCR 4 for self-hosting. For a business leader, the issue is not raw performance but a simple question: who controls your data, and what does it really cost? Here is a neutral reading of what these models change.

Key takeaways

• An open source model (or "open-weight" model) is one whose files can be downloaded, hosted on your own infrastructure, fine-tuned and used commercially, without depending on a remote API.
• GLM-5.2, released by Z.ai (formerly Zhipu) on 17 June 2026 under the MIT licence, offers a 1 million token context window and beats GPT-5.5 on several coding benchmarks for roughly one sixth of the cost, according to VentureBeat.
• On the European side, Mistral Large 3 is distributed under the Apache 2.0 licence, and Mistral OCR 4 (24 June 2026) runs entirely on your own infrastructure: documents never leave the company.
• The key argument for an SME is not performance, it is data sovereignty and cost control: no data sent to a third party, a predictable bill.
• The downside: self-hosting demands technical skills and hardware. For most SMEs, the realistic approach stays hybrid (API for the standard, EU-hosted open source for the sensitive).

Open source, open weight: what are we talking about?

The vocabulary is confusing. A proprietary model (such as GPT-5.5 or Claude) is used through an API: you send your request to the vendor's servers, which return the answer. You own nothing.

An open-weight model is downloaded: the model files are public. You can run it on your own servers, adapt it to your business and use it commercially, depending on the licence. Strict "open source" also requires publishing the training data and code, which few models actually do. In practice, the licence is what matters for an SME.

Why 2026 changes the picture

For a long time, open models lagged well behind proprietary ones. The gap has narrowed. According to independent rankings reported by VentureBeat, GLM-5.2 stands as the best open-weight coding model, scoring 62.1 on SWE-bench Pro and 81.0 on Terminal-Bench 2.1, just behind Claude Opus 4.8. All at a far lower usage cost.

On the European side, Mistral continues its sovereignty strategy. Mistral Large 3, released in 2026 under the Apache 2.0 licence, is built on a Mixture-of-Experts architecture (675 billion parameters, of which around 41 billion are active at inference). A French-incorporated company under EU law, Mistral hosts its services in the Union and guarantees, for its Pro and Enterprise tiers, that conversations are not used to train models. On 24 June 2026, Mistral OCR 4 shipped in a version deployable on your own infrastructure: a concrete case where sensitive documents never leave the company.

What it means for an SME

Three benefits stand out, along with as many limits to keep in mind.

1. Data sovereignty. This is the main advantage. Data security is the number one concern of SMEs facing AI. A model hosted internally or with a European provider transmits no data outside the company. That is decisive for regulated sectors (health, legal, finance) and consistent with GDPR.

2. Cost control. APIs charge by volume: the more you use, the higher the bill. A self-hosted model has a fixed, predictable infrastructure cost. Beyond a certain volume, the equation becomes favourable.

3. Business adaptation. You can fine-tune an open model on your own documents to specialise it, something a closed API does not always allow.

How to decide: the grid

The right reflex is not "open versus proprietary" but "which option for which use". Here is a reading grid.

For most SMEs, the answer is hybrid: an API for everyday, non-sensitive uses, and an open model hosted in Europe for confidential data or high volumes.

A regulatory calendar that rewards caution

The context reinforces the value of sovereignty. The penalty provisions of the EU AI Act apply from 2 August 2026. Keeping control of your data and your processing becomes an argument for compliance as much as for security. Open models hosted in the EU fit naturally into this logic.

FAQ

Is an open source AI model free?

The download and the licence are often free (MIT, Apache 2.0), but use is not: you have to pay for hardware or a host to run the model. The free part is the right to use, not the operation.

Is an open model less capable than a proprietary one?

The gap narrowed sharply in 2026. According to VentureBeat, GLM-5.2 beats GPT-5.5 on several coding benchmarks for a sixth of the cost. On other tasks, proprietary models keep the edge. You have to test on your real use case.

Does open source mean GDPR compliance?

Not automatically. Compliance depends on hosting and data governance. An open model hosted in the EU, on controlled infrastructure, greatly eases compliance, because the data never leaves your perimeter.

Do you need a technical team to self-host a model?

For very large models, yes: suitable hardware and AI infrastructure skills are required. An alternative is to go through a European host that offers these models as a managed service, without running everything in house.

Conclusion

Open source AI models do not replace proprietary ones: they widen the choice. For an SME, their value comes down to two words: sovereignty and cost. The winning approach stays pragmatic and hybrid, to be calibrated to your data and your volumes. To frame this choice, see our guide on choosing the right AI model and our case studies on AI projects in SMEs.