Cyberpediaby Brain
ENDEFR

Open-weights vs proprietary AI models

An open-weights model or an online service: what it changes for company data, and what employees should do or avoid.

In 30 seconds

Where it happens
In any AI tool, whether it is an online service or a model downloaded and installed on a workstation.
What changes
With an online service, your data goes to a host. With an open model, everything depends on where it runs, and on who installed it and keeps it updated.
The risk
An "open" model is still subject to a licence and to risks, and it does not protect data if it is used through a third-party service.
The reflex
Use only the AI tools authorised by Brain, install no downloaded model without approval, and enter no sensitive data in an unapproved service.
If it's already done
Tell Brain's security team without delay and say which data was entered or which model was installed.

What it is

An AI assistant is built on a language model. This model can be distributed as files that anyone can download and run on their own machine: this is called an open-weights, or "open", model. It can also stay with its publisher, which offers it only through an online service: this is called a proprietary model.

For an employee, the difference lies less in the quality of the answers than in a simple question: where does the data you enter go, and who is accountable for it? This page explains both cases and what Brain expects for each.

Straight answers

What is the difference between an open model and a proprietary model?

An open model is published, instead of staying behind a service. Two examples are documented in the sources cited here: LLaMA, presented in February 2023 by a team of researchers, who write "We release all our models to the research community"1, and Mistral 7B, released under the Apache 2.0 licence2. A proprietary model can only be reached through an online service: the user sends their text and receives an answer, without seeing the model.

Where does my data go?

With an online service, the text you enter leaves the company to be processed by the provider. The UK NCSC guidelines in fact cover systems that use models hosted by an organisation or external interfaces (APIs), and call for AI systems that work without revealing sensitive data to unauthorised third parties4. With an open model installed on a workstation or an internal server, processing can stay on site. Nothing in the models themselves requires this: everything depends on where they run.

Does an open model protect my data?

Not automatically. If an open model is used through a third-party service, what you enter goes through that service, as with a proprietary model. What matters is where the processing takes place, not whether the model is open.

Who updates the model, and who is accountable for it?

With an online service, the model stays with the provider. With a downloaded model, it is with whoever installed it, and nobody else looks after it. The NCSC guidelines place update management in the "operation and maintenance" phase of an AI system and supply chain security in the development phase4. On personal data, the CNIL raises the question of how responsibilities are shared when an AI system is supplied by a publisher and used by a third party5.

Can the provider reuse what I enter?

The CNIL states that an AI system provider can lawfully reuse data entrusted by a data controller if several conditions are met: authorisation from the controller, a compatibility test, information to and respect for the rights of individuals, and compliance of the new processing5. Hence the value of using a tool whose conditions Brain has checked.

1How it works1 min

A language model is trained on a large quantity of text, then used to predict a continuation of text in response to a prompt7. The LLaMA researchers showed that a state-of-the-art model could be trained using publicly available data only, and that LLaMA-13B outperformed GPT-3 (175B) on most benchmarks1. Mistral 7B, with 7 billion parameters, outperforms Llama 2 13B on all evaluated benchmarks2. The largest Llama 3 model has 405 billion parameters, and its authors find its quality comparable to that of GPT-4 on many tasks3. They released this model, along with Llama Guard 3, a model for checking inputs and outputs3.

On the company side, the decision therefore covers three points: where the model is hosted, who updates it, and who is accountable for an incident. The answer varies from one tool to another. This is why the list of authorised tools serves as the rule.

2What you do1 min
  • Use only the AI tools authorised by Brain, whether they are based on an open or a proprietary model.
  • Do not install a downloaded model on a workstation without approval from the security team.
  • Enter no sensitive data in an unapproved service: data about customers, colleagues or health, and company secrets.
  • For a trial, prefer fictitious data, meaning data with the same structure as the real thing but unrelated to any person. The CNIL mentions tests on fictitious data to determine the data needed by an AI system5.
  • If in doubt about a tool, ask before using it.
  • If you have already entered data in an unapproved service or installed a model, report it right away.
3Pitfalls2 min

"Open" does not mean without a licence

Mistral 7B is released under the Apache 2.0 licence2. An open model is therefore published with a licence, which sets its terms of use. The sources cited here do not detail the licences of the other models: do not assume that every use is allowed.

A downloaded model can be booby-trapped or biased

OWASP ranks among the ten 2025 risks for generative AI applications the supply chain of models, exposed to various vulnerabilities, and data and model poisoning6. It also ranks misinformation, this time as a core weakness of applications that rely on models6. A model file fetched from an arbitrary site has not been checked by Brain. The NCSC in fact lists supply chain security among the points to address from the development of an AI system onward4.

A model can have memorised personal data

The CNIL notes that many studies have shown that large language models tended to "memorise" elements of their training data: first name, surname, address, telephone number, bank card number5. A successful attack on the confidentiality of a model can amount to a data breach5. This is true of an open model as of a proprietary one.

A model that keeps learning

Some systems reuse the data entered in order to improve, which the NCSC calls continuous learning7. The CNIL points out that this mode of operation raises data protection questions, because the data then serves two distinct purposes5. What a service does with what is entered is among the conditions to know before authorising it.

A framework that moves

The NIST AI risk management framework, voluntary in use, was published on 26 January 2023, then supplemented on 26 July 2024 by a profile devoted to generative AI8. The NIST page states that version 1.0 is under revision8. Recommendations evolve: Brain's rules, for their part, remain those of the company.

4Real cases1 min

27 February 2023: LLaMA

A team of researchers presents LLaMA, a collection of models from 7 to 65 billion parameters, trained using publicly available datasets only, and releases all its models to the research community1.

10 October 2023: Mistral 7B

Mistral 7B v0.1, a model of 7 billion parameters, is released under the Apache 2.0 licence, with a version fine-tuned to follow instructions2.

31 July 2024: Llama 3

The Llama 3 family is presented, with a 405-billion-parameter model and a context window of up to 128,000 tokens. The pre-trained and post-trained versions of the largest model are released. The versions that add image, video and speech are not widely released, because they are still under development according to the authors3.

Fact sheet

Identifier
CP-268
Wing
4. Artificial intelligence
Brain skill
Authorized AI Tools & Shadow AI (authorized_tools)
Brain reflex
air-023
Diagram
G3 · System & Protocol Anatomy
Languages
de, en, fr
Sources
8 sources checked on 3 November 2026
Sources8 sources
  1. arXiv, “LLaMA: Open and Efficient Foundation Language Models”, 27 February 2023. https://arxiv.org/abs/2302.13971 (accessed 3 November 2026)
  2. arXiv, “Mistral 7B”, 10 October 2023. https://arxiv.org/abs/2310.06825 (accessed 3 November 2026)
  3. arXiv, “The Llama 3 Herd of Models”, 31 July 2024. https://arxiv.org/abs/2407.21783 (accessed 3 November 2026)
  4. NCSC (UK), “Guidelines for secure AI system development”, 27 November 2023. https://www.ncsc.gov.uk/collection/guidelines-secure-ai-system-development (accessed 3 November 2026)
  5. CNIL, “IA : comment être en conformité avec le RGPD ?”, 5 April 2022. https://www.cnil.fr/fr/intelligence-artificielle/ia-comment-etre-en-conformite-avec-le-rgpd (accessed 3 November 2026)
  6. OWASP, “LLMRisks Archive”. https://genai.owasp.org/llm-top-10/ (accessed 3 November 2026)
  7. NCSC (UK), “Principles for security of Machine learning ML”, 22 May 2024. https://www.ncsc.gov.uk/collection/machine-learning-principles (accessed 3 November 2026)
  8. NIST, “AI Risk Management Framework”. https://www.nist.gov/itl/ai-risk-management-framework (accessed 3 November 2026)

Cite this article

Brain, “Open-weights vs proprietary AI models”, Cyberpedia, 2026. https://startbrain.io/cyberpedia/en/articles/open-weights-vs-proprietary-ai-models/

BibTeX 
@misc{brain_cp268_en_2026,
  author = {Brain},
  title = {Open-weights vs proprietary AI models},
  year = {2026},
  howpublished = {Cyberpedia},
  url = {https://startbrain.io/cyberpedia/en/articles/open-weights-vs-proprietary-ai-models/},
  note = {CC BY 4.0}
}
CC BY 4.0 licence