Why OpenAI Puts Its Free Models on Hugging Face

August, twenty twenty-five. OpenAI does the thing everyone said it never would. It publishes two models you can just... download. A hundred and twenty billion parameters, and twenty billion. Apache two point oh. Free to take, free to modify, free to sell what you build with them.

OpenAI? Giving away weights?

And here's the part I can't get past. The place they put the files... was Hugging Face. A competitor's shelf.

Hang on. Why would you hand your crown jewels to the other guy's storefront?

That's the whole episode. Keep it in your head, because the answer tells you who actually wins this decade.

Start me lower. What IS Hugging Face? I know the logo. The little smiley face.

A hosting platform for AI. Models, datasets, demos. Run by Clément Delangue, French, from La Bassée. At seventeen he was one of the biggest French sellers on eBay. Then Stanford. Then this. By the time he sat down with the Acquired podcast, the company was valued at four and a half billion dollars... and the investor list is Nvidia, Google, Amazon, Salesforce, Intel, AMD, Qualcomm, IBM.

That's basically every chip company on earth in one cap table.

Which tells you exactly what they're buying. Chipmakers want a world where models are files you download and run on hardware you already paid for. OpenAI's business is the opposite shape. You download nothing. You send a request to their servers and you pay per token.

Rent versus own.

Rent versus own. And Hugging Face has gone very long on own. Over a million open repositories. More than fifteen thousand companies. In April of twenty twenty-five they bought a French humanoid robotics startup, Pollen Robotics, and Delangue said the goal was open source AI robotics.

Open source robots. Nothing could possibly go sideways there.

Hold that thought. Because for years the knock on OpenAI was the name. Open in the title, closed in practice. No weights, no architecture, nothing you could hold in your hands. Then last August... they broke it. Both gpt-oss models, full chain-of-thought visible, commercial use allowed.

That's revenue they're setting on fire. Why?

Because distribution is the prize. If every student, every hobbyist, every bank running models behind its own firewall learns on YOUR architecture, your formats, your tooling... you own the layer underneath the money. OpenAI's own help pages say it flat out: gpt-oss isn't served through their API. No API pricing. It runs on vLLM, on Ollama, on llama dot cpp. They are handing it away.

And did anyone actually take it?

Four million downloads in a single month on the big model's Hugging Face page. And that's just the official copy. There's a whole cottage industry squeezing it down into versions that run on a laptop.

Okay, but then... this isn't a rivalry. OpenAI's open models live on Hugging Face's servers, Hugging Face gets the traffic. That sounds like a marriage.

It's both, and that's the honest answer. Nobody's suing anybody. They're competing for a position. Hugging Face wants to be the neutral commons where every lab's weights sit side by side... which quietly makes every lab interchangeable. OpenAI wants to be the destination. The brand. The thing you type into.

Interchangeable. That's the scary word if you're OpenAI.

And here's the part neither of them loves. Community trackers scraping the Hugging Face hub were showing, in late September of this year, that the top of the open-weight charts wasn't American at all. Alibaba's Qwen family, sitting at number one, millions of downloads a month.

So the open race is being run and won... somewhere else?

Careful here. Those leaderboards are community snapshots, not audited numbers, and a download measures curiosity as much as real use. But the direction holds month after month. Which reframes everything. It isn't OpenAI versus Hugging Face. It's whether Western open models stay relevant at all.

Europe tried, though. Didn't the Swiss put one out?

Apertus. Second of September, twenty twenty-five. ETH Zurich, EPFL, the Swiss national supercomputing centre. Apache two point oh, more than eighteen hundred languages, eight billion and seventy billion parameter versions, up on Hugging Face. Warm reception. Well behind the frontier, and nobody pretended otherwise.

Open and free, and a lap down.

That's the trade everyone's arguing about. And then there's the cost of being everyone's front door. Early twenty twenty-six, attackers hijacked the Hugging Face platform to push Android malware. The kind that takes over a phone completely.

Of course. If you're the app store for AI, somebody poisons a package.

And that's a vulnerability a closed lab simply doesn't have. When your model only ever runs on your own servers, nobody tampers with the download. So OpenAI gets to say: controlled access is safer. And Hugging Face gets to say: a technology this powerful can't be three companies' private property. Both are good arguments. That's why it's a real fight.

So where does it land?

Unresolved. As of today, the eighth of October, twenty twenty-six. But go back to the thing you couldn't get past. OpenAI put its open weights on a rival's shelf because that shelf is where developers already stand. That's the tell. The money's in the model... the POWER is in the place people go to get it.

And right now that place has a smiley face on it.

Which matters to you even if you never download a single file. Because the question underneath all of this is whether, ten years out, running a capable AI is something you do... or something you're permitted to do.

One's a tool. The other's a subscription.

And nobody's voting on which one we get. They're just uploading files.