OpenAI Unveils First Open Source Models Since 2019: Introducing gpt-oss!

This Tuesday, August 5, 2025, OpenAI announced the release of two open source AI reasoning models from its “o” series. These models are available under the Apache 2.0 license, which is among the most permissive in terms of open source and free software. Two versions, gpt-oss-120B and gpt-oss-20B, are available for local download on various devices.

First open source models in nearly 6 years

For the first time since the launch of its GPT-2 model in 2019, Sam Altman’s company has released downloadable models on the Hugging Face platform, which can be used by anyone without the need for internet or the ChatGPT app. This marks a return to the company’s roots, reflecting its commitment to open AI models in principle.

OpenAI has introduced two model sizes. The first, gpt-oss, has 120 billion parameters and can operate on a single Nvidia GPU. The smaller model has 20 billion parameters and can run on a laptop or smartphone with at least 16 GB of RAM. These models can also offload complex tasks to more powerful cloud-hosted AI models when they reach their limits. For instance, if a user wants to generate an image, they can redirect the request to one of OpenAI’s more powerful proprietary models.

Are these models truly competitive?

Trained using “the most advanced pre-training and post-training techniques to date,” according to OpenAI, gpt-oss 120B and gpt-oss 20B aim to become the most powerful open source models on the market. To this end, OpenAI is also collaborating with long-standing partners such as Orange, AI Sweden, and Snowflake “to better understand the real-world applications of our open models, including hosting models on-site to ensure data security […] and enabling individual developers, large enterprises, or even government agencies to customize these cutting-edge models on their own infrastructure.”

But how effective are OpenAI’s new models really? The company has subjected them to academic evaluations to compare them with other AI models. The gpt-oss-120B performs better than o3-mini and rivals o4-mini in high-level coding tasks, general problem-solving, and tool invocation. It also excels in health and mathematics queries. Meanwhile, gpt-oss-20B performs comparably to o3-mini in coding and general problem-solving but outperforms it in health and mathematics questions.

Several delays in deployment due to security concerns

The launch of gpt-oss-120B and gpt-oss-20B was postponed several times in recent months for security reasons. OpenAI has also considered the risk that malicious actors could refine its gpt-oss models for more effective use in cyberattacks or the development of biological or chemical weapons, which may explain the delays. The highly anticipated release of GPT-5 this summer, however, could overshadow the concerns related to a perfectly reliable open source model.