In the Last Exam of Humanity test, their scores were 19% and 17.3%. They also stand out in tasks such as tool invocation, code execution, and web search.
OpenAI introduced two open-source AI models that work without internet access and already outperform the successful DeepSeek R1
In the Last Exam of Humanity test, their scores were 19% and 17.3%. They also stand out in tasks such as tool invocation, code execution, and web search.
OpenAI introduced two open-source artificial intelligence models that work without needing to be connected to the internet. They're available for installation on computers, laptops, or cell phones.
This is a direct response to pressure from DeepSeek and its successful R1 model. With this move, OpenAI is once again competing strongly in the open-source field.
The models are called gpt-oss-120b and gpt-oss-20b. The first has greater capacity and is designed for PCs with NVIDIA GPU, while the second is lighter and even works on laptops with 16 GB of RAM.
Both can be installed locally and don't require an internet connection. This is an important step toward more private, offline AI.
They outperform the DeepSeek R1 model. In Codeforces benchmarks, they reach scores of 2622 and 2516 points. Although they don't match models like o3 or o4-mini, they offer excellent performance in open-source AI.
The models have a high level of hallucinations: they generate up to three times more incorrect answers than the o1 model. OpenAI hasn't identified the cause yet.
It's always recommended to verify sources if they're used in sensitive or analytical tasks.
With these launches, OpenAI aims to compete with the progress of open models created in China. Additionally, it breaks with its historic closed model, although it keeps training data secret.
Both models are licensed under Apache 2.0, which allows unrestricted use for research, customization, or commercial development.