OpenAI Releases Open-Weight AI Models, Embracing Transparency and Customization

OpenAI has released two open and freely available AI models, GPT-oss-120b and GPT-oss-20b, which can mimic human reasoning and perform tasks such as writing code and searching online. The models are open-weight systems, similar to Meta's Llama, allowing developers to customize them. However, the training data remains undisclosed, falling short of a truly open-source model. The release follows China's DeepSeek and aligns with the US President's AI Action Plan, but critics argue that open-source software can be less secure.

OpenAI has released two new open and freely available AI models, GPT-oss-120b and GPT-oss-20b, designed to mimic human reasoning and perform tasks such as writing code and searching online. These models are open-weight systems, similar to Meta's Llama, allowing developers to customize them according to their needs. However, the training data remains undisclosed, which falls short of a truly open-source model [1].

The release of these models follows the success of China's DeepSeek and aligns with the US President's AI Action Plan, which encourages the development of open models to become global standards in business and academic research. OpenAI's Chief Executive Officer, Sam Altman, acknowledged the need to develop a different open-source strategy after the success of DeepSeek's open R1 system [1].

Both models are relatively compact and efficient. The more powerful GPT-oss-120b can run on a single 80 GB GPU, while the GPT-oss-20b can operate on a laptop with 16 GB of memory. They are available on the AI software hosting platform Hugging Face and can produce text in response to user prompts [1].

OpenAI has been working with early partners such as Orange SA and Snowflake Inc. to test and learn about real-world applications of these open models. The models have been trained using a mix of reinforcement learning and techniques informed by OpenAI's most advanced internal models, including o3 and other frontier systems. They have demonstrated strong performance on various benchmarks, including coding, competition math, health, and agentic tool use [2].

The release of these open models is part of OpenAI's ongoing effort to increase transparency and accessibility in AI development. However, critics have raised concerns about the security of open-source software, arguing that it can be less secure than proprietary models. OpenAI has acknowledged these concerns and has been conducting additional safety tests and reviews to address them [1].

In conclusion, OpenAI's release of GPT-oss-120b and GPT-oss-20b represents a significant step in the development of open-weight AI models. These models offer developers the opportunity to customize and run AI systems on their own hardware and services, potentially leading to increased innovation and collaboration in the AI field. However, the lack of disclosed training data and the ongoing debate about the security of open-source software will continue to be important considerations for developers and investors.

References:
[1] https://www.bloomberg.com/news/articles/2025-08-05/openai-releases-open-weight-models-after-deepseek-s-success
[2] https://openai.com/index/introducing-gpt-oss/