On July 24, 2026, Nvidia, Microsoft, Meta, Mistral, IBM, Hugging Face and about twenty others published an open letter called “Open Weights and American AI Leadership”, asking Washington to avoid “premature restrictions” on open-weight models. Within a day the list of signatories had doubled to around fifty, with Google and even OpenAI quietly joining. Two big names stayed off the list: Anthropic and Amazon.
To understand why the entire industry felt the need to write to the government about a file format, you first need to understand what “open weight” means. Good news: it fits in one image, the one we’ll keep for the whole article. An AI model is a dish, and there are three ways a chef can sell it to you.
Three ways to sell a dish
When a lab “releases a model”, it actually chooses between three levels of openness. Same kitchen, three very different products.
1. The closed model: you eat at the restaurant
GPT-5, Claude, Gemini: you use them through an API or a chat window. You order, a plate arrives, and that’s it. You never enter the kitchen, you don’t take anything home, and you pay per plate (per token, to be precise). If the restaurant changes its menu, raises its prices or closes, you adapt. That’s a closed model.
2. The open-weight model: the dish to take home
Here, the chef hands you the finished dish in a takeaway box: a big file containing the weights, the billions of parameters the model learned during training. Download it and the model is yours to run: on your servers, sometimes even on your laptop.
What you do not get is the recipe: the training data, the training code, the tricks of the kitchen. You can reheat the dish, season it to your taste, even serve it in your own restaurant, but you couldn’t recreate it from scratch.
That’s an open-weight model. It’s where most of the “open” action is today: Llama (Meta), the Mistral models, Gemma (Google), gpt-oss (OpenAI), DeepSeek, Qwen (Alibaba), Kimi (Moonshot AI).
3. The open-source model: the full recipe book
The rarest level: weights plus the training code plus the data (or a faithful description of it), under a license that lets you do almost anything. With the recipe book, you could re-cook the dish yourself, adjust the ingredients, and understand exactly what went into it. Genuinely open-source models at a large scale are rare; OLMo, from the Allen Institute for AI, is the textbook example.
This distinction matters, because “open source” gets used as a loose synonym for “open weight” all the time, including in the letter’s own coverage. The Open Source Initiative keeps reminding everyone: a dish without its recipe is not open source. It’s open weight. Still very useful. Not the same thing.
What the takeaway box changes for you
Concretely, having the weights in your hands unlocks four things:
- It runs at home. On your infrastructure, or fully offline as a local model. Your data never leaves the building, a big deal for hospitals, banks, public administrations, or anyone who cares about sovereignty.
- No bill per plate. There’s no per-token invoice; you pay for the hardware instead of the restaurant. Careful though: serious hardware for serious models. “Free to download” is not “free to run”.
- You can season it. Fine-tuning to specialize it on your domain, quantization to shrink it onto smaller hardware. Try doing that through a restaurant window.
- Nobody can take it back. Once the file is on your disk, no price change or discontinued API can affect you. The model is frozen, yours, forever.
That last point cuts both ways, and it’s the heart of the political debate: a published weight can never be recalled. If a model turns out to help with something dangerous, there’s no un-publish button. Every safety concern about open weights flows from that.
One honest caveat: “open weight” is not a single legal status. Some models come under genuinely permissive licenses (Apache 2.0 for Mistral or gpt-oss), others under custom licenses with conditions (Llama’s community license, for example). Always read the label on the box.
The recap table
| Closed model | Open weight | Open source | |
|---|---|---|---|
| The image | You eat at the restaurant | You take the dish home | You get the recipe book |
| What you get | An API, nothing else | The weights (the trained model) | Weights + code + data |
| Where it runs | The vendor’s servers | Wherever you want | Wherever you want |
| You pay for | Each call (tokens) | Your own hardware | Your own hardware |
| Customization | Prompts only | Fine-tuning, quantization | Everything, even re-training |
| Can be revoked? | Yes, anytime | Never | Never |
| Examples | GPT-5, Claude, Gemini | Llama, Mistral, Gemma, gpt-oss, DeepSeek, Qwen, Kimi | OLMo |
So why did the giants write to Washington?
Now the news makes sense. Here’s the sequence, and it all happened in one week of July 2026.
The trigger. Chinese labs, which publish almost everything open weight, have caught up. Early July, Moonshot AI released Kimi K3, an open-weight model that beats the American flagships on some benchmarks. Worse for the mood in Washington: American officials accuse Chinese labs of building these models partly by distillation of American closed models. In our kitchen language: sending a taster to the restaurant thousands of times, noting every flavor, and reverse-engineering a very close dish at home. On July 21, the US Treasury Secretary put it bluntly: “open source is not open season on American IP”, with threats of sanctions against labs running what he called industrial-scale distillation.
The fear. In response, Washington started weighing restrictions that could go beyond punishing actual theft, and reach the technique of publishing or distilling open models in general. The White House denied that a specific executive order was drafted, but the debate was real enough for the industry to react within days.
The letter. Published July 24 with 25 founding signatories, it makes essentially three arguments:
- Don’t confuse the technique with the theft. Distillation is a standard, legitimate way to build smaller models from bigger ones; labs use it on their own models constantly. Punish actual IP misappropriation with targeted legal tools, not by banning the technique.
- Restricting open weights would backfire. Startups, researchers, hospitals and defense contractors all build on open-weight models. If American labs stop publishing them, the world doesn’t stop using open weights; it just switches to Chinese ones. The default dish of the global AI kitchen would be cooked in Beijing.
- Help the open ecosystem instead. More compute access for startups and researchers, shared training resources, so that open American models stay competitive.
The absent. Anthropic and Amazon didn’t sign. Anthropic has long argued the safety side of the irrevocability problem: you can fine-tune the safeguards out of an open model, and you can never recall it. It’s a coherent position, not just commercial reflex, though it’s fair to note that closed-model vendors also happen to benefit commercially from restrictions on open ones. Both readings can be true at once. And OpenAI signing while Anthropic abstains shows the industry line isn’t “open vs closed vendors”: it’s about who believes the risk outweighs the ecosystem.
Where it stands. As of today, nothing has passed: no executive order, no law. The letter is lobbying, brilliant, coordinated lobbying, and reading it as such is the honest way to read it.
In a nutshell
“Open weight” means the chef sells you the finished dish to take home: the model file is yours to run, tune and keep forever, but the recipe (data and code) stays secret. It’s the middle ground between the restaurant-only closed model and the true recipe-book open source, and it powers a huge share of real-world AI, from startups to hospital basements.
The July 2026 letter exists because that middle ground became geopolitics: Chinese open-weight models caught up, Washington considered restrictions, and almost the whole industry answered “don’t ban the takeaway box, punish the actual thieves”. Whether you find the letter convincing or the abstentions wiser, at least now the words mean something. And that part, really, is not rocket science.
A word you don’t know? The glossary defines every AI term, in plain language.
Sources: TechCrunch, CNBC, Tom’s Hardware, Forbes, TechRadar.
