Open Source AI Models: Why They Matter for Business
By Alexandre Saint-Jean
Audio version
Audio version produced by text-to-speech from the article. Our AI charter
In late July 2026, Moonshot AI released Kimi K3, the largest open-weight model ever made public. For a business, this is not a footnote: it confirms that open models (Kimi, Mistral, DeepSeek) are becoming a credible alternative to proprietary American APIs, with a real benefit in control and cost.
What just changed with Kimi K3's release?
Moonshot AI, a Chinese company, released Kimi K3 between 16 and 27 July 2026. The model has 2.8 trillion parameters, a record for an open-weight model, with a one-million-token context window (TechTimes, 24 July 2026).
Two points matter in particular for a business. First, the licence: Modified MIT, which permits commercial use with no royalty. Second, the official API's price: roughly $3 per million input tokens and $15 per million output tokens, competitive against equivalent proprietary models.
Why do open models matter for a business?
An open model (precisely, open-weight: the trained parameters are published) can be hosted on infrastructure of your choosing: a European cloud, an in-house server, or the originating vendor's API. That flexibility changes three concrete things for a company.
The first benefit is independence from any single proprietary API vendor. The second is the ability to choose where your data actually travels, a central question for any organisation that has already asked how to use AI without losing control of its data. The third is downward pressure on pricing: the more capable open models exist, the more inference costs fall across the whole market, proprietary vendors included.
Where do Mistral and DeepSeek stand in this?
Mistral, Europe's reference player, also regularly ships open models, with the added advantage of native European hosting and a vendor subject to EU law. That is the point that really sets Mistral apart for a business focused on legal sovereignty: the full picture of Mistral's offering (hosted API, certified European cloud, self-hosting) is covered in our dedicated article on sovereign LLMs.
DeepSeek, another Chinese vendor, keeps up the same pace of releasing capable open models. With no confirmed figures yet on its latest generation, the principle stands regardless: both China and Europe are publishing top-tier open models, which meaningfully widens the options beyond the proprietary American giants alone (OpenAI, Anthropic, Google).
Should you choose an open model over a proprietary API?
This is not really a question of raw performance, the best open models now compete with proprietary ones on most business use cases. It is a question of architecture and control.
Three profiles stand out. For everyday use with no sensitive data, a proprietary or open API both work fine, and the deciding factor becomes price and perceived quality. For a requirement to keep data resident in Europe, an open model hosted with a European cloud provider (or directly with Mistral) fits the need. For a strong legal sovereignty requirement, self-hosting an open model on infrastructure you control remains the only option that removes all dependency on a foreign jurisdiction.
What should you check before deploying an open model in your business?
Before connecting an open model to production data, three checks matter. First, the exact licence: not all of them permit unconditional commercial use, so read it rather than assume it. Second, the vendor's country of origin and the law it falls under: even when only the weights are used locally, the originating vendor and its ecosystem remain a factor worth documenting. Third, hosting capability: running a model with hundreds of billions of parameters needs GPU infrastructure and expertise that few businesses have in-house, which is exactly why hosted offers like Mistral or certified European clouds matter.
The battle between open models has only just started, but the practical takeaway for a business is already clear: choosing the model and choosing where to host it are two separate decisions, and it is the second one that actually determines your level of control over the data.
Frequently asked questions
- What is an open-weight AI model?
- It is a model whose weights (the trained parameters) are freely downloadable, often through platforms like Hugging Face. You can host it on your own server or a cloud of your choice, unlike a proprietary model only accessible through its vendor's API.
- Can a European business actually use Kimi K3?
- Technically, yes: the Modified MIT licence allows commercial use, and the model is available through an API priced at around $3 per million input tokens. But it is a Chinese model, which raises the same legal sovereignty question as with an American vendor: what law is the provider actually subject to.
- Is an open model automatically sovereign?
- No. Open means the weights are accessible, not that hosting sits in Europe or is out of reach of extraterritorial law. Sovereignty depends on who hosts the model and under what jurisdiction, not on the model's licence.
- Should you wait before choosing an open model for your business?
- No. The field is moving fast, but everyday uses (summarisation, extraction, an internal copilot) are already well covered by the open models available today. The sensible move is to scope your hosting and data-control needs first, then pick a model to match, rather than the other way round.