In an unexpected turn of events, OpenAI has released two new open-source models, gpt-oss-120b and the slightly more modest yet equally powerful gpt-oss-20b, which most users will find far more appealing because it needs only a 16 MB graphics card to run smoothly.
OpenAI is placing special emphasis on four aspects of this release:
- Permissive licence: Both models are distributed under the Apache 2.0 licence, which is highly permissive and even allows commercial deployment.
- Agent-ready and code-executing: They can be used to build AI agents and are capable of running Python code.
- Customisable reasoning: Anyone can set the reasoning level to low, medium or high.
- Full chain-of-thought: The models expose their entire chain of thought, making debugging easier and inspiring greater confidence in the outputs.
Everything appears to be in its favour, because these systems can be downloaded onto our computers and used even offline, locally, guaranteeing complete privacy for data and queries. The models can be tested on the web, and to use and download them locally (gpt-oss-20b requires at least 16 MB of video memory for smooth operation, while gpt-oss-120b needs 80 MB) the simplest route is probably to install the Ollama software and, once installed, choose the model you wish to run and download it locally as well.
Without doubt this launch has huge implications and opens the door for everyone to use it free of charge and even commercially. There is speculation that OpenAI—expected to announce chatGPT 5 today—will use its open-source move as a pretext to withdraw its standard free plan and start charging all users, arguing that the service can now be downloaded and used free of charge in this manner.
These liberated models are reasoning models, which is why they work with less video memory, but the service will be quite slow, perhaps suitable only for leaving it running in the background. According to OpenAI’s data, gpt-oss-120b is not far behind its most powerful models, such as OpenAI o3.
* Original article written in Spanish, translated with chatGPT and reviewed in English by Jorge Mediavilla.
Leave a Reply