OpenAI Gave Away a Free Model You Can Run at Home. The Open vs. Closed War Changed Forever

For years, OpenAI was the face of closed AI. You could use its models through ChatGPT or an API, but you could never download them. That changed on August 5, 2025, when the company released gpt-oss, two open-weight language models under the permissive Apache 2.0 license. They were OpenAI’s first open-weight language models since GPT-2 in 2019.

One of them, gpt-oss-20b, was designed to run on consumer hardware with about 16 GB of memory. The larger gpt-oss-120b was built to run on a single high-end 80 GB GPU. Anyone could download them, run them offline, fine-tune them and build products on them.

It was a striking reversal, and it tells you a lot about where the AI industry is heading.

“Open weights” vs. “open source”

First, a vocabulary check, because these terms get mixed up constantly.

• Closed models are accessed only through a company’s app or API. You never see the model itself.
• Open-weight models let you download the trained parameters, the “weights,” and run the model yourself. Licenses vary from very permissive, like Apache 2.0 or MIT, to restrictive, with limits on commercial use or company size.
• Fully open-source AI, by the Open Source Initiative’s Open Source AI Definition released in October 2024, requires more: enough information about the training data, plus the code used to train and run the model, for someone to meaningfully study and rebuild it.

Most models people call “open source,” including gpt-oss, are more precisely open-weight: the weights are free, but the full training data is not.

How open models went from hobby to front line

Open-weight AI used to trail far behind. Then a series of releases closed the gap:

• Meta’s Llama family made capable open-weight models widely available starting in 2023 and kick-started a huge ecosystem of fine-tunes and tools.
• Mistral, a French startup, showed that a small team could ship highly efficient open models.
• Alibaba’s Qwen models, many released under Apache 2.0, became popular with developers worldwide.
• DeepSeek-R1, released in January 2025 under the MIT license, delivered reasoning performance comparable to OpenAI’s o1 and showed open models could reach near the frontier.

By the time gpt-oss arrived, open models were no longer a curiosity. They were a competitive threat, and a strategic tool.

Why a closed lab would give a model away

OpenAI’s announcement framed gpt-oss around making AI widely accessible and letting developers run models on their own infrastructure. There are also clear strategic reasons any major lab would ship open weights:

• Developer mindshare. If builders standardize on your open models, they are more likely to use your paid ones too.
• Competition. Strong open models from other companies were winning developers who needed local or private deployment.
• Use cases closed APIs cannot serve: offline devices, sensitive data that cannot leave a company’s servers, and countries or industries with strict data rules.

Open vs. closed: how to choose

Neither side is simply better. They are tools for different jobs.

Choose an open-weight model when you need:

1. Privacy and control. Data never leaves your machine or your servers.
2. Predictable costs at scale. Once you own the hardware, heavy usage does not mean a growing API bill.
3. Customization. Fine-tune on your own data, style or domain.
4. Offline or edge use. Laptops, phones, factories, field equipment.
5. No vendor lock-in. The model cannot be deprecated out from under you.

Choose a closed model when you need:

1. The absolute frontier. The most capable models are usually still closed.
2. Zero infrastructure. No GPUs to buy, manage or scale.
3. Built-in features like web search, long-running agents, multimodal tools and enterprise compliance.
4. Speed to launch. An API key and a few lines of code.

Many teams now use both: a closed frontier model for the hardest tasks, and a small open model for high-volume, private or offline work.

The safety debate

Open weights come with a real tradeoff. Once a model is downloaded, the creator cannot take it back, monitor how it is used, or patch out a dangerous behavior. Critics worry about misuse; supporters argue openness enables independent safety research, competition and resilience against concentrated power. OpenAI said it ran safety evaluations on gpt-oss before release, including testing deliberately fine-tuned versions for risky capabilities. Expect this debate to intensify as open models get stronger.

How to try one tonight

If you have a reasonably modern computer, you can run an open model locally with free tools like Ollama or LM Studio. Download a small model, ask it questions with your Wi-Fi off, and you will understand the appeal immediately: a capable assistant that runs entirely on your own machine.

The bottom line

When the company that defined closed AI starts publishing open weights, the line between the two camps has moved. Open models are now a core part of every serious AI strategy, not a sideshow. For builders, that means more choice than ever. For everyone else, it means powerful AI that you can own, not just rent.

Sources

• OpenAI, “Introducing gpt-oss”, Aug. 5, 2025
• Open Source Initiative, The Open Source AI Definition 1.0, Oct. 2024
• DeepSeek-AI, DeepSeek-R1 on GitHub
• Ollama