OpenAI just did something it hasn’t done in years: it released open-source language models.
The last time this happened was with GPT-2 back in 2019. Now, we’ve got two new ones: gpt-oss-120b and gpt-oss-20b.
What’s different this time? They’re being released under the Apache 2.0 license, which basically means anyone can use, modify, and run them however they want. It’s a pretty unexpected move from a company that’s mostly kept things closed off lately.
What are Open-Weight models?
Open-weight models are AI systems where the actual “weights,” the core numbers the model learned during training, are made public. These weights are what drive the model’s predictions, responses, and overall behavior.
Unlike closed models like GPT-4 or Claude, which you can only access through APIs or web services, open-weight models give users real freedom. The users can:
- Download and run them on your own hardware, whether that’s a personal PC, an office server, or even a phone
- Fine-tune them with your own data to get better results for specific tasks
- Peek under the hood and play around with the model architecture itself
- Use them however you want, without needing to pay for access or rely on a third-party platform
That said, “open-weight” doesn’t mean fully open-source. The users get the model’s weights, but not always the training code or the full dataset used to build it.
So it’s a compromise as users get control and transparency, but companies still keep parts of the training process private, often for proprietary or safety reasons.
Why OpenAI’s move will change everything?
Until now, OpenAI’s most powerful models were locked behind paid APIs, only accessible to big companies or well-funded users. The company cited safety concerns, but the result was limited access for the wider developer and research community.
That’s changing. With the release of open-weight models, OpenAI is giving independent developers, researchers, and startups the tools to run cutting-edge AI locally, whether for privacy reasons, cost savings, or just the freedom to tinker.
These models can be fine-tuned for everything from legal advice and medical support to regional languages and specialized coding tasks.
No need to wait for permission or updates from a central team, the users can customize and deploy on their own terms.
It’s also a win for transparency. Open weights let researchers dig into how the models work, test for bias, and check for safety issues. That kind of visibility builds trust.
This move is also seen in response to growing competition from Meta, Mistral, and DeepSeek. But no matter the reason, the outcome is the same: more people, in more places, can now tap into powerful AI tech and build things that weren’t possible before.
How to access Open-Weight models?
OpenAI has made its latest models freely downloadable on Hugging Face and GitHub. They run on hardware with 16 – 80 GB of memory, and are fully customizable, ideal for tasks like reasoning, instruction-following, and code execution, and fine-tuned for specific industries or languages.
By releasing these models openly, OpenAI is shifting power from a handful of big tech firms to developers, startups, and researchers around the world.
That means faster innovation, stronger security testing, and more experimentation happening sooner. Researchers can now inspect how the models work, find bias, or build new features without waiting for someone else’s roadmap.
The post OpenAI to release Open-Weight models: what they are and why it changes everything appeared first on Invezz