OpenAI Unveils First Open-Weight Models Since 2019 With Apache 2.0 License

OpenAI has finally released its first open-weight models since 2019, gpt-oss-120b and gpt-oss-20b, marking a significant shift in the company’s approach to AI development and deployment [1]. These models are available under the Apache 2.0 license and OpenAI’s gpt-oss usage policy, allowing users to download, modify, and deploy them freely without the need for a paid API [2]. The 117B parameter gpt-oss-120b is tailored for high-reasoning and production-level tasks, while the 21B parameter gpt-oss-20b is optimized to run on consumer-grade hardware, including devices with as little as 16GB of memory [4]. Both models support multi-step reasoning and tool usage and can be deployed locally or in the cloud without API dependencies [5].

The move is widely seen as a strategic response to the growing influence of open-source AI models. OpenAI has emphasized that the models are not fully open-source—users do not gain access to the training data or full training code—but the availability of model weights represents a middle ground between transparency and control [2]. According to OpenAI, the models were trained using a mix of reinforcement learning techniques and run on Nvidia H100 GPUs, with compatibility across a range of hardware, including Nvidia Blackwell and RTX GPUs, as well as chips from AMD, Cerebras, and Groq [5]. This broad compatibility aims to make the models accessible to a wider range of developers and enterprises.

In a strategic partnership, Amazon, Microsoft, and Baseten have integrated the models into their cloud platforms, allowing customers direct access to them. Amazon will offer the models via AWS Bedrock and Sagemaker, marking the first time AWS has hosted OpenAI models [5]. Meanwhile, Microsoft is incorporating gpt-oss-20b into Windows, providing optimized versions for developers [11]. These partnerships highlight the collaborative ecosystem now emerging around open-weight models and reflect the broader industry shift toward open-access AI.

OpenAI has also emphasized the safety aspects of the release. The company filtered out sensitive data such as chemical, biological, radiological, and nuclear material during pre-training and tested the models for potential misuse scenarios, including malicious fine-tuning. According to internal assessments, the models did not reach the company’s defined threshold for high capability in harmful tasks [2]. Greg Brockman, OpenAI’s president, stated the company is excited to see the ecosystem evolve and contribute to pushing the frontier of AI innovation [2]. Sam Altman also highlighted the significance of the release, calling it a “big triumph of technology” and expressing pride in the team’s efforts [2].

Analysts have noted that while the release meets the expectations of the open-source community, some limitations remain. Rowan Curran of Gartner pointed out that the models being primarily trained in English may limit their utility in multilingual settings [7]. Lian Jye Su of Omdia described the move as long-awaited and a positive step toward greater accessibility in AI development [6]. Meanwhile, David Nicholson of the Futurum Group highlighted that the Apache 2.0 license is more commercially friendly than other open-source licenses, which could drive enterprise adoption [9].

Despite the company’s efforts, some AI researchers have called for even greater openness. Hannaneh Hajishirzi of the University of Washington and AI2 suggested that true openness would require the full release of training data and detailed methodologies [12]. OpenAI has yet to provide that level of transparency, leaving the debate on open-source AI development ongoing.

The gpt-oss models are now available for testing via OpenAI’s open model playground and through platforms like Hugging Face and Together AI [3]. With their release, OpenAI has taken a significant step toward bridging the gap between proprietary and open-source AI, positioning itself within a growing industry trend [5]. As more companies and developers engage with the models, the impact of this shift on the broader AI landscape will likely become clearer in the coming months.

Source:

[1] OpenAI. (2025, July 31). Introducing gpt-oss. https://openai.com/index/introducing-gpt-oss/

[2] OpenAI. (2025, July 31). gpt-oss-120b & gpt-oss-20b Model Card. https://openai.com/index/gpt-oss-model-card/

[3] Northflank. (2025, August 4). Run OpenAI's new GPT-OSS (open-source) model on .... https://northflank.com/blog/self-host-openai-gpt-oss-120b-open-source-chatgpt

[4] TechTarget. (2025, August 4). Belatedly, OpenAI releases two new open weight models. https://www.techtarget.com/searchenterpriseai/news/366628493/Belatedly-OpenAI-releases-two-new-open-weight-models