OpenAI has launched two new language models: gpt-oss-120b and gpt-oss-20b, making them freely available to the public. What makes this release stand out is that, for the first time in years, advanced AI models that can understand and respond to human instructions are now accessible to anyone, including developers, researchers, and hobbyists, without needing expensive infrastructure or paid access.
The company claims that the larger model, gpt-oss-120b, is designed to run efficiently on a single high-end graphics card (80 GB memory), while the smaller 20-billion parameter version can work on devices with just 16 GB of RAM. This means that even personal computers, workstations, or locally hosted servers can now run sophisticated AI tools, without relying on cloud-based platforms.
Why this matters
In practical terms, this move could allow organisations and individuals to build AI-driven tools for writing, analysing data, or even performing basic medical or scientific tasks, directly on their own systems. It also brings down the cost of AI development significantly.
Crucially, both models are being released under the Apache 2.0 licence, which means users can adapt, modify, or integrate the models into their own products or services, without needing to pay royalties or seek special permissions.
Smarter, cheaper AI that’s ready to work offline
Despite their open-access nature, these models are not basic. They have been trained using methods similar to those used for OpenAI’s proprietary models, like reinforcement learning and instruction tuning. The 120-billion parameter model reportedly performs nearly as well as OpenAI’s paid o4-mini model on tasks that require reasoning, understanding context, and working with tools like code or calculators.
The smaller 20b version also punches above its weight, handling mathematics, coding, and even health-related queries with accuracy that matches or surpasses some commercial models. Both are compatible with existing AI toolchains, and have been tested on tasks such as function calling, chain-of-thought reasoning, and tool use.
Given growing concerns about AI misuse, OpenAI says it has subjected these models to additional safety checks. This includes testing how the models behave when intentionally fine-tuned for dangerous topics like cybersecurity or synthetic biology. According to OpenAI, even under those conditions, the models did not reach concerning capability levels.
A public red teaming challenge is also being launched, with a $500,000 reward fund, encouraging researchers to find potential risks or misuse cases. Findings from this effort will feed into future safety work, and results will be shared publicly.
Running locally: From laptops to company servers
To make adoption easier, OpenAI has worked with platforms like Hugging Face, Microsoft, and vLLM to ensure the models are widely available. Developers can run them using tools like PyTorch or Apple’s Metal, and even Windows users can access them via Microsoft’s ONNX Runtime and AI Toolkit.
The models have also been made compatible with local inference platforms such as LM Studio, Ollama, and llama.cpp. This gives people the flexibility to run AI offline, important for privacy-conscious users, companies handling sensitive data, or developers in regions with poor internet connectivity.
Availability:
Both models can be downloaded now via Hugging Face and are available in a memory-efficient format. Toolkits, guides, and sample code for using the models are also being shared publicly.
For those curious to try the models without any setup, OpenAI has launched a browser-based open model playground, where anyone can explore what the models are capable of.