OpenAI has dropped two new open-weight models, though they aren’t exactly open-source

OpenAI has finally started to live up to its name and has dropped two new “open-weight” AI models. This marks the first time in over half a decade since the AI startup’s last major dive into the open ecosystem with GPT-2. It also marks a strategy shift for the company, which has lately come to be known for its more proprietary, closed-off approach to development.

Models can run on a single GPU

The new models, dubbed gpt-oss-120b and gpt-oss-20b, are designed to be cheaper and more accessible options for developers and researchers. The larger 120b variant can run on a single Nvidia GPU, while the lighter 20b version is nimble enough to operate on a consumer laptop with 16GB of memory. Both are built for advanced reasoning and can even be used as a personal assistant on your PC to write text or search through files.

It’s important to know these are open-weight, not fully open-source. This means their parameters, the core elements refined during training, are public. However, the full source code remains under wraps. It’s unclear how much of that is accessible to developers as of writing.

ALSO READ: ChatGPT agent can plan and execute tasks by itself

This release seems to be a direct response to the likes of Meta, Mistral AI, and DeepSeek have been making waves with their own open models.

As usual, OpenAI says it has put its new releases through serious safety checks. The company filtered out harmful data related to chemical, biological, and nuclear threats and even simulated how bad actors might try to misuse them. Their conclusion is that the models are not easily pushed to a “high capability” threshold for malicious purposes.

How to get your hands on these AI models

Developers can now get their hands on the models through platforms like Hugging Face and GitHub under a permissive Apache 2.0 license. They’re also being made available by cloud providers, including Amazon and Microsoft, and can run locally via programmes like LM Studio and Ollama.

It’s also worth mentioning that developers can connect these open models to OpenAI’s more powerful closed models and let the local AI offload complex tasks it can’t handle, like image processing, to the cloud.