What Are Open Weights? Why People Want GPT-4o Released

Open weights means a model’s trained parameters are downloadable so anyone can run that model on their own hardware without asking the company that built it for permission. It isn’t the same as open source: with open weights you get the finished model, while the training data and the code that made it usually stay private.

GPT-4o has been switched off inside ChatGPT since February and people are still posting about it every day. A Change.org petition to bring back the model or release its weights has passed 20,000 signatures. The campaign runs on 3 hashtags (#Keep4o, #OpenSource4o and #BringBack4o). One phrase keeps coming back: an “archival open weight release”.

OpenAI retired it and pointed at the replacements

On February 13, 2026, OpenAI retired GPT-4o from ChatGPT together with GPT-4.1, GPT-4.1 mini and o4-mini. Its own reason was usage: the “vast majority” of people had moved to GPT-5.2. Only 0.1% of users were still choosing GPT-4o on any given day.

GPT-4o

0.1%

It wasn’t the first time OpenAI had tried: the company deprecated 4o once before and brought it back during the GPT-5 release. A subset of Plus and Pro users preferred 4o’s “conversational style and warmth” for creative work. They said they needed longer to move on so that feedback went into GPT-5.1 and GPT-5.2, which added personality controls like Friendliness and warmth.

So OpenAI didn’t delete the model without a word. It wrote a post, gave a reason, and pointed at the replacements. The people who lost it read the post and disagreed with the reason.

Bring it back, or release it

The petition asks for the model back: it wants the 4-series running and affordable, with a “legacy interface” for people who can’t code. OpenAI is the only one who can hand it back. It has decided not to. (The company is called Open AI and it won’t open 4o. Nomsayin.)

The petition also asks for the weights. Its wording is “release the weights and architecture to the public under a non-commercial research license.” That request survives the refusal. A model whose weights are public can’t be switched off by the company that made it. Anyone can run it on their own machine.

That’s the ask behind “archival open weight release”. A later petition makes it sharper: open-source the final checkpoints of GPT-4o and GPT-5.1 under a permissive license such as Apache 2.0, the same license Aleph Alpha put on Kolibri’s weights.

One lab already does part of this. On November 4, 2025, Anthropic committed to preserving the weights of every model it has released. It also committed to interviewing each model before retiring it, then writing up what the model said. When it retired Claude Opus 3 on January 5, 2026, it kept the model available to paid users on claude.ai and by request on the API. It gave the model a place to publish articles, while the weights stayed private and the model didn’t disappear.

That’s a third option the petition rarely mentions: preserve it and keep serving it. It costs money, which is what Anthropic says makes retiring models necessary in the first place, but it doesn’t require giving anything away.

Usage counts measure the middle

The 0.1% number is accurate and OpenAI published it itself. A usage number counts how many people pick a thing when a default does most of the picking. It says nothing about how much the people who pick it care, or what breaks for them when the model goes.

GPT-4o was the model a lot of people were attached to for reasons that never show up in a usage table: a tone they liked writing with, or a way of responding that newer models didn’t reproduce. The petition isn’t from the 99.9% but from the tail.

I spend a lot of my time on that gap: I benchmarked an open-weight German model against Claude Sonnet 5.5 on long German writing, 136 blind comparisons judged by 2 other models. The open-weight model lost every one. That tells you which model is better on average and nothing about the person whose work runs on the model that lost.

A benchmark is the wrong instrument for asking whether that person should be upset. Benchmarks and usage counts measure the middle. A retirement decision lands on the edges.

Open weights isn’t the same as free

Open weights takes away the permission problem. Downloading the weights and running them yourself is still work.

I ran Aleph Alpha’s Kolibri myself when it launched this October, a 78 billion parameter open-weight model released under Apache 2.0. The weights were free to download. I still needed about 78 GB of graphics memory to load them, which meant 2 data-center GPUs. On launch day no hosted provider served it so I rented an endpoint at $5.50 an hour. A model the size of GPT-4o would be heavier still: the petition’s own estimate puts it at 300 to 500 billion parameters.

Open weights means nobody can switch the model off. It doesn’t mean it runs for free, on a laptop, forever.

The 3 terms people mix up:

	What you get	Can the maker switch it off?
Closed model	Access through an API or an app, on their servers	Yes
Open weights	The finished parameters, runnable on your own hardware	No, you can keep running it
Open source	The weights plus the training data, code and recipe	No, and you can rebuild it

OpenAI already does the middle column: it ships its own open-weight models, the gpt-oss family, downloadable from Hugging Face. It keeps an Open Models page for those. So this was never about whether OpenAI knows how to release weights. It’s about whether a retired model is worth preserving, and who gets to keep it.

Prepare for the model that disappears

The GPT-4o campaign is a loud version of a problem every team using AI has: you build on a model that can change or disappear under you. A couple of things help, whether or not the weights ever come out.

• Run your own evals so a model swap is something you can measure instead of guess. That’s what my AI evals workshop is about.
• Pin the API version you depend on and write down when it last changed.

OpenAI is one company deciding what it owes the people who built on a model it has now retired. Anthropic decided to keep the weights in a drawer and to keep Opus 3 running. That’s more than the GPT-4o campaign has gotten so far.

If you’re working out how much to depend on a model that might get retired under you, that’s the kind of problem I take on stage.