top of page
Search

OpenAI’s new models: gpt-oss-120b and 20b now available as open-weight downloads


ree

OpenAI releases "open-weight" models for public use

On August 5, 2025, OpenAI announced two new artificial intelligence models, gpt-oss-120b and gpt-oss-20b, which are now available under an open-source license. This means anyone can download, use, and modify them. These models are designed to be flexible and competitive, while maintaining a high level of safety and performance.



gpt-oss-120b and gpt-oss-20b models: what they are and how they work

  1. gpt-oss-120b: This model has 120 billion parameters and is the more powerful of the two. It is very similar to advanced models like GPT-4o but designed to be more open and accessible.

  2. gpt-oss-20b: A lighter model with 20 billion parameters. It is less demanding in terms of hardware resources and can even run on high-end laptops.


Both models are designed to perform complex tasks such as reasoning and data analysis, but unlike previous OpenAI models that were closed and accessible only via paid API, these can be downloaded and used freely.



Benchmarks and performance: how they compare to other models

These new models have been tested on various benchmarks, such as MMLU (a general intelligence test), and have proven to be very competitive. For example, gpt-oss-120b performed very well on reasoning tasks, even surpassing more expensive models like GPT-4o in certain areas. Additionally, the models did not show signs of high security risks during OpenAI's "preparedness framework" tests, which are designed to ensure the AI does not become dangerous or harmful.



Where to download them and how to use them

The gpt-oss models are not accessible through ChatGPT, neither via the ChatGPT app nor through the chat.openai.com web interface. They are not part of the interactive model family that you can use with prompts and responses in a chat, such as GPT-4o or GPT-3.5. Instead, these models are intended for developers, companies, and researchers who want to integrate them into their own software, applications, or services.

Anyone who wants to use them must download or run them via technical environments such as:

  • Hugging Face: a platform where you can find the model files and sample scripts to start inference.

  • GitHub: source code, configuration files, and tools for installing and using locally.

  • Microsoft Azure AI, AWS SageMaker, and Databricks: cloud environments that allow you to run these models without manual installation, ideal for companies with cloud infrastructures.


The Apache 2.0 license also allows commercial use, so businesses can modify and integrate them into real products without paying royalties or facing legal restrictions.



Not ChatGPT: what they are for and who can use them

It is important to clarify that these models are not designed for general users or for use via chatbot interface. They are not a “new version of ChatGPT,” but downloadable models intended for those who want to run artificial intelligence models in a controlled environment, on their own hardware or cloud servers.

For example:

  • A researcher working on a medical diagnostics application can use gpt-oss to run the model without sending sensitive data to external servers.

  • A startup that wants an offline AI assistant can use gpt-oss-20b on machines with consumer GPUs.

  • A company can integrate these models into its products without relying on OpenAI’s APIs.


OpenAI has also included advanced debugging tools (such as Chain of Thought mode and structured JSON output), making these models suitable for professional scenarios and regulated environments, such as healthcare, legal, or industrial sectors.


______

FOLLOW US FOR MORE.


DATA STUDIOS

bottom of page