Open Models: Genuine Openness or a Slogan?

Meta launched its LLaMA model in February 2023 with its code and gave its weights only to approved researchers, and the weights reached the public within about a week. Within weeks independent developers had built improvements in instruction tuning, in compressing the model, and in running it on modest hardware. That episode ignited a question still open: who owns the machine's mind, and who may adjust it?

A memo that sparked debate

In May 2023 the SemiAnalysis blog published a memo that it said came from a researcher inside Google, and whose authenticity it had checked; an anonymous person had posted it on a public discussion server. Its author argued that independent developers were advancing fast and at lower cost, and that his company had "no secret sauce". The blog stressed, however, that the memo was one employee's opinion, not the company's position, and that it did not itself agree with it. Let us take it as a witness to the state of the debate, not a verdict on the outcome.

What does "open" mean?

A model is often called "open" merely because its weights are published. The Open Source Initiative's definition for AI, version 1.0 of October 2024, requires four freedoms: to use, study, modify and share. It also requires sufficient information about the training data for a skilled person to build an equivalent system, the complete code for training and running, and the parameters. Weights are therefore one pillar of three, and it is fair to say "open-weight" where the other conditions are not met.

Is open catching up with closed?

The Stanford AI Index (2025) reports that the best closed model outperformed the best open-weight model by 8.0% in early January 2024 on the Chatbot Arena leaderboard, and that the gap narrowed to 1.7% by February 2025. The measure is users' preference among models' answers, one face of performance, and should not be generalised.

Openness is not transparency

The Foundation Model Transparency Index, in its December 2025 edition, assessed thirteen companies and found an average of 40 out of 100 against 58 the previous year, with the researchers cautioning that the indicators were revised, so the comparison needs care. Meta fell from 60 to 31 and Mistral from 55 to 18, while IBM scored 95. Announcing "openness" does not replace disclosing the data, the impact and the risks.

The value of open

The firmer economic evidence concerns open software generally, not models. Hoffmann, Nagle and Zhou, in a 2024 Harvard Business School paper, estimated that the most widely used open software was worth 4.15 billion dollars to rebuild once, and that the firms using it would have had to spend 8.8 trillion dollars to build its equivalent. They added that 5% of developers create 96% of this value. The figures carry a double lesson: openness saves everyone money, but those who bear the real burden are a few who deserve support.

And Arabic?

Openness does not replace work. Releasing weights lets researchers tune a model on Arabic texts, but that demands good data, computing power and skills, as the article "Arabic Artificial Intelligence: Do We Make or Consume? Sovereignty and Arabic Language Models" explained. The question of abuse remains: making a model available to all makes it available to those who would misuse it, and no magic solution exists. The balanced approach is neither to close the door out of fear nor to open it out of heedlessness, but to distribute capability together with responsibility.

How should we read the announcements?

• Are the data, or a description of them, available, or only the weights?
• What is the licence? Does it allow commercial use and modification?
• Is the complete training code available?
• Have the risks and evaluations been disclosed?

Whoever asks well chooses well, and can benefit from the open without being taken in by its name.