The 'Open-Washing' Reckoning: Why Llama and Mistral Aren't Open Source

I love running models locally, but we need to have a serious conversation about the corporate hijacking of the term “open source.” Meta and Mistral have spent the last year wrapping themselves in the righteous flag of open-source software, soaking up the community goodwill while keeping their most valuable assets firmly locked in a black box. The Open Source Initiative (OSI) finally had enough.

Before you read: We previously covered how running Llama 3 locally frees you from cloud API costs and privacy risks. But while the weights are free to download, the legal and philosophical reality of how these models are built is much darker.

Table of contents

• The Open-Washing Epidemic
• The Training Data Black Box
• Fun Facts / Did you know
• The Legal Battlefield
• Sources

The Open-Washing Epidemic

Mark Zuckerberg loves to talk about how Llama is an open-source triumph. It is a brilliant PR move that positions Meta as the benevolent rebel fighting against the closed ecosystems of OpenAI and Google.

But claiming Llama 3 is open source is like handing someone a fully baked cake, giving them the oven temperature settings, but refusing to tell them the ingredients. If you cannot study how it was made, modify the recipe, and recreate it from scratch, it is not open. You just got a free cake.

Meta’s licensing agreement basically says you can do whatever you want with the model, unless your app reaches 700 million monthly active users, in which case Zuck wants to have a chat. It is the classic open-source ethos of “freedom for all, unless you’re an actual threat to my ad revenue.”

The Training Data Black Box

For two years, the OSI—the stewards who literally maintain the Open Source Definition—have been debating what “open source” means in the age of neural networks. They just released the official Open Source AI Definition (OSAID), and it draws a hard line in the sand: to be open source, you must provide “sufficiently detailed information about the data used to train the system.”

Llama 3 was trained on a staggering 15 trillion tokens. Mistral Large was trained on an undisclosed massive dataset. We have absolutely no idea what those tokens are.

Frankly, the OSI is right to demand this. If a model exhibits a dangerous bias, a security flaw, or starts hallucinating malicious code, you cannot patch the root cause without seeing the training data. Without the data, developers are just blindly fine-tuning over the cracks, hoping the foundation doesn’t give way.

Fun Facts / Did you know

• The term “Open Source” was coined in 1998 by the Foresight Institute, largely to make the “Free Software” movement sound more corporate-friendly and palatable to business executives.
• Despite claiming the “open” mantle, Meta’s Llama 3 license strictly prohibits developers from using Llama’s outputs to train any competing language models.

The Legal Battlefield

Why are these companies so terrified of releasing their training data? It is not about protecting a secret sauce. It is entirely about copyright liability.

If Meta publicly lists the exact 15 trillion tokens they scraped from the web, they instantly hand a pristine, itemized invoice to the New York Times, Getty Images, and every published author on Earth. Hiding the data is a survival mechanism. They use “open weights” as a marketing shield to build an ecosystem, knowing full well that true open source would bankrupt them in court. It is a brilliant legal strategy: crowdsource your QA testing to thousands of independent developers while keeping the copyright lawyers entirely in the dark.

We need to stop letting tech giants redefine industry standards just because they give us free toys. Open weights are great for local development, but let’s call them what they are. Llama and Mistral aren’t open source; they are just highly capable freeware.

Sources

• The Open Source Initiative: Open Source AI Definition
• Meta Llama 3 License Agreement
• The Register: OSI tells AI industry to stop open-washing