Models

What is open-source AI?

Open-source AI is AI that anyone may use for any purpose, study, modify and share. Under the Open Source Initiative's definition of October 2024, that takes three things: the weights, the full code used to train and run the model, and enough information about the training data for a skilled person to build a substantially equivalent system. Most models sold as open source publish only the first, and are better called open-weight. For a company, the practical gain is often the same either way: anyone can run the model, so you choose who runs it, and under which law.

Published 6 Oct 2026 · checked on 6 Oct 2026

The definition, in one paragraph

The reference is the Open Source AI Definition, version 1.0, announced by the Open Source Initiative on 28 October 2024 (announcement). It carries the four freedoms of open-source software over to AI: use the system “for any purpose and without having to ask for permission”, “study how the system works and inspect its components”, modify it “for any purpose, including to change its output”, and share it “with or without modifications”.

Freedoms on paper are not enough; you need the means to exercise them. So the definition asks for the form in which a skilled person would actually change the system:

Data information — “sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system”;
Code — “the complete source code used to train and run the system”;
Parameters — “the model parameters, such as weights or other configuration settings”.

Note what it does not ask: the training data itself. It asks for a listing of the public data used and where to obtain it, and of third-party data and where to obtain it, including for a fee.

Open source, open weights, closed

Measured against that definition, AI models fall into three groups:

Kind	What you get	Examples
Open source	Weights, the training and running code, information on the training data	Ai2's Olmo, which publishes its training data, code and weights
Open weights	The weights, under a licence; rarely the code or the data	Mistral Small 3.2, DeepSeek V4 Flash, Llama 3.3 70B
Closed	Access through the maker's own service only	GPT, Claude and Gemini models; on Evanesca, Mistral Large 4, in preview, run by its maker in its EU region

So “open source” in a model's marketing usually means open weights: you can download and run the model, but not rebuild it. What open-weight AI means covers what the weights themselves let you do, and open-weight models in 2026 which to pick for what.

Licences: where “open” gets its conditions

Two models can both be “open” and allow very different uses. Mistral Large 2, for instance, came out in July 2024 under the Mistral Research License, for research and non-commercial use only (announcement). The licences of the open-weight models in Evanesca's catalogue, and what each allows:

Licence	What it allows	Models on Evanesca
Apache-2.0	Permissive open-source licence: commercial use allowed	Qwen3.8 27B, Apertus v1.5 70B, Qwen3.6 35B, Gemma 4 26B, Mistral Small 4, Qwen3.5 397B, GPT-OSS 120B, Qwen3 Coder 30B, Qwen3 235B, Mistral Small 3.2, Pixtral 12B
MIT	Permissive open-source licence: commercial use allowed	DeepSeek V4.1 Flash, DeepSeek V4 Flash, GLM-5.2
GLM-5.3 License	A licence of its own, to read before any commercial use	GLM-5.3
Llama 3.3 Community License	Commercial use allowed; above 700 million monthly active users, a licence from Meta is required, and “Built with Llama” must be displayed (licence)	Llama 3.3 70B

Read the licence before the benchmark: a model you may not use commercially is no use to a business, whatever its scores.

What the AI Act says about open-source models

Since 2 August 2025, makers of general-purpose AI models have four duties: keep technical documentation for the authorities, give the providers who build on the model the information they need, follow a policy to comply with EU copyright law, and publish a sufficiently detailed summary of the content used for training (art. 53(1), Regulation (EU) 2024/1689).

Open source buys a partial exemption. A maker that releases a model under a free and open-source licence, weights included, is spared the first two duties — never the last two, and never for a model with systemic risk, presumed when training used more than 10²⁵ floating-point operations (art. 51(2) and 53(2)). Those models carry extra duties: evaluation and adversarial testing, risk mitigation, incident reporting, cybersecurity (art. 55).

These duties are the makers', not those of a service that runs their models. Evanesca does not develop, train or modify models; its AI Act page explains what it does owe you, such as telling you that you are talking to an AI.

What open models change for a company

You choose the operator, and with it the jurisdiction. The CLOUD Act follows the company that holds the data, not the server (18 U.S.C. § 2713). An open model can run at a French operator; a closed one only where its maker sells it.
You are not tied to one vendor's API. Several operators can serve the same weights, and most speak OpenAI's format.
You can test before you trust: read the model card, run your own prompts, compare. Evanesca's benchmarks show which of six checks each model has passed.
You can adapt the model — fine-tune it on your own data — if you host it yourself.

Open weights do not make data sovereign on their own: the same model, served by a US company, stays within reach of US orders. DeepSeek V4 Flash is the clearest case: its maker's app stores personal data in China, while Scaleway runs the same family of weights in Paris.

What open does not give you

A guarantee of quality. Closed models still lead some tests: in DeepSeek's own report, a closed model stays ahead of its newest Flash model on the hardest agentic benchmarks (DeepSeek V4.1 Flash).
Freedom from conditions. Licences can cap users, forbid commercial use or require attribution, as the table above shows.
Small hardware. At 8-bit precision, weights take about one gigabyte of GPU memory per billion parameters: Mistral Large 4, at 1000 billion parameters, needs about 1000 GB before any context.
A support contract with the maker. You rely on the model card, the community and whoever operates the model.
Control over who else uses it. Anyone can run the weights, including without the maker's safeguards.

None of this argues against open models; it argues for reading the card, the licence and the route before you rely on one.

Three ways to start

A service that runs open models for you. Evanesca's agent, chat and OpenAI-compatible API run open-weight models in Europe, most of them in France, with prices per million tokens from €0.168 for input, VAT included.
An operator directly, such as OVHcloud AI Endpoints or Scaleway Generative APIs, with your own contract.
Your own servers, with an inference engine such as vLLM or Ollama — see self-hosted AI gateways.

The models list every operator and place, and the docs show the API. Evanesca is not open yet; this site shows how it will work.

Frequently asked questions

Is open-weight the same as open source?

No. The Open Source Initiative's definition requires the weights, the complete code used to train and run the model, and enough information about the training data to rebuild an equivalent system. Most models called open source publish only the weights, under a licence: they are open-weight.

Can I use open-source AI models commercially?

It depends on the licence. Apache-2.0 and MIT allow commercial use. The Llama 3.3 Community License allows it but requires a licence from Meta above 700 million monthly active users, and Mistral Large 2 came out under a research-only, non-commercial licence. Read the licence of each model.

Does the AI Act exempt open-source models?

Partly. A maker that releases a model under a free and open-source licence, weights included, is spared two documentation duties, but still owes a copyright policy and a public summary of its training content; a model with systemic risk gets no exemption (art. 53(2)).

Where can I try open models without a GPU?

Through an API run by an operator. OVHcloud and Scaleway sell access in France, and Evanesca runs open-weight models from both in a chat with an Agent mode and an OpenAI-compatible API. Evanesca is not open yet; this site shows how it will work.