OpenAI has announced the release of gpt-oss-120b and gpt-oss-20b, two open-source language models that represent a significant advancement in the reasoning capability of this type of model. Available under the flexible Apache 2.0 license, according to the company, they stand out for their performance, which surpasses that of other open models of similar size in reasoning tasks, their ability to utilize tools, and their optimization for efficient implementation on consumer hardware.
These are OpenAI’s first open models since GPT-2, released in 2020.
As Sam Altman explained during the presentation: “We are excited to make this model, the result of billions of euros in research, available to the world, so that AI can reach as many people as possible. With this launch, we trust that new lines of research will be enabled and new types of products will be created. (…) We are thrilled that the world can build upon an open AI infrastructure, created in the United States, based on democratic values, freely available and oriented toward collective benefit.”
OpenAI affirms that, given their size, these models represent significant progress both in reasoning capability and safety, and that their objective in releasing them is to accelerate cutting-edge research and enable safer and more transparent AI development. The idea is that these open models will also help reduce barriers for emerging markets, sectors with limited resources, and smaller organizations that may not have the budget or flexibility required to adopt proprietary models.
The gpt-oss-120b model nearly matches OpenAI’s o4-mini model in core reasoning tests and can operate efficiently on a single 80 GB GPU. For its part, gpt-oss-20b delivers results similar to those of OpenAI’s o3-mini on common benchmarks and requires only 16 GB of memory. Both models support a context length of up to 128k and were trained using a high-quality text dataset focused on STEM, programming, and general knowledge.
Both models have been trained with a combination of reinforcement learning and advanced techniques inspired by OpenAI’s internal models. In evaluations, gpt-oss-120b outperforms o3-mini and, in some cases, equals or exceeds o4-mini in areas such as competitive programming, general problem-solving, tool usage, and healthcare.
On the other hand, safety is a fundamental component of OpenAI’s approach to these releases. The models underwent comprehensive safety training and evaluation, and an adversarially fine-tuned version of gpt-oss-120b was tested to assess risks of malicious use. As part of this initiative, OpenAI will also organize a “Red Teaming” challenge with a prize of $500.000 for the community to identify new security vulnerabilities.
The files for both models are available for free download on Hugging Face.
Image: OpenAI
Your email address will not be published. Required fields are marked *
Δ