OpenAI presents new AI models that can be run directly on your computer
Today, on 5 August 2025, OpenAI has released a new GPT-OSS (Open-Source Style) model, which is the first open model from the company since 2019 (for the first time since GPT-2). The model comes in two versions: for 20 billion parameters and for 120 million parameters.
GPT-OSS-120B is a huge 120 billion parameter model that can be run on a single NVIDIA GPU. OpenAI claims that it performs as well as the o4-mini in reasoning, programming, and text analysis tasks.
TheGPT-OSS-20B is a compact model (~20 billion parameters) optimised for GR with 16 GB of RAM or more.
Both models are available for free download via Hugging Face, Azure, AWS SageMaker/Bedrock platforms under the Apache 2.0 licence, with the possibility of commercial use and modification.
Nuances.
The models are open, but not fully open-source: training data and source code are not published.
OpenAI emphasises that GPT-OSS has undergone thorough security testing, including external experts and abuse simulations.
Bottom line.
OpenAI has made a significant turnaround in the industry: The GPT-OSS-120B and 20B are rebranded as open access to flagship AI models, no longer with API restrictions or subscriptions. Now companies and research teams can train their own large language models without restrictions. Finally, you can make full use of your CPU's neural computing acceleration module. At the same time, OpenAI preserves the work of previous trained models. In fact, the company is trying to shift the artificial intelligence race from the field of hardware solutions for individual requests (where it is difficult for it to compete with large corporations) to the quality of training and output of its models, in which OpenAI still holds a dominant position even with its more limited financial resources.
Source: www.theverge.com
