OpenAI releases first open weight models since 2019

A return to open weights

OpenAI has officially released its first open weight models in five years. The company launched gpt-oss-120b and gpt-oss-20b on Tuesday, making them available under the Apache 2.0 license. This move comes after years of criticism from the AI community regarding the company's shift away from open-source contributions.

Open weight models are versions of AI where the training parameters are publicly available for others to use. This is distinct from fully open source models, which also include the underlying coding sources and the specific training data used. By releasing these weights, OpenAI is allowing developers to run and customize the models more freely than before.

The release is seen as a strategic response to the growing popularity of open models. Other major players like Google, Alibaba, and IBM have recently released similar models. As TechTarget reported, analyst Lian Jye Su noted that the move is a long-overdue contribution to the open source community after a period of absence.

Hardware and performance

The two new models are designed to serve different hardware needs. The gpt-oss-120b model is intended for data centers and is comparable in performance to OpenAI's o4-mini reasoning model. In contrast, the gpt-oss-20b model is a smaller version similar to o3-mini, specifically optimized for edge devices.

OpenAI optimized these models to run efficiently across the Nvidia hardware stack. This includes compatibility with Nvidia Blackwell and RTX GPUs. The 20b model is particularly accessible because it can run on devices with as little as 16 GB of memory, making it viable for consumer-grade hardware.

To achieve these results, OpenAI used a mix of reinforcement learning techniques. These techniques were informed by their other advanced models, such as o3. The models also utilize a mixture of experts, which is a standard technique used to improve the efficiency and capabilities of large language models.

Commercial and strategic impact

The choice of the Apache 2.0 license is a significant detail for enterprise users. Analysts suggest that this license is less restrictive than others and is well-suited for commercial activity. This allows companies to use the models for profit-generating applications with more confidence in the legal framework.

This release also positions OpenAI as a competitor to Meta's Llama models. Meta recently indicated it might stop releasing its most advanced superintelligence models as open weights. By providing high-quality open weights, OpenAI may be attempting to undermine Meta's influence in the commercial AI space.

The move is also intended to attract companies that want to fine-tune their own models. Fine-tuning is the process of taking a pre-trained model and training it further on a specific dataset to perform a particular task. By providing the weights, OpenAI makes it easier for these companies to build specialized tools without relying solely on closed APIs.

Limitations and future steps

Despite the excitement, some experts point out that the models have specific limitations. Because they were primarily trained on English text, they may struggle with performance in other languages. This remains a hurdle for global developers who require high-quality multilingual support out of the box.

Some researchers argue that OpenAI could go further to be truly open. They suggest that the company should release the full training data and the complete recipe of what worked and what did not during the training process. This would make the models fully open source, rather than just open weight.

The timing of this release is particularly notable following a major shift in the broader AI landscape. Earlier this year, the Chinese vendor DeepSeek made a significant impact with its R1 model, which helped fuel the current popularity of open source software. Many other generative AI providers have since capitalized on this momentum by releasing their own models to keep pace with the community.

Expert perspectives and analysis

Industry experts have noted that the release addresses a specific need for commercial flexibility. David Nicholson of The Futurum Group pointed out that while some models are only open for academic use, the Apache license allows for profit-making activities. He noted that commercial use is a different story than academic use.

While the community is excited, some researchers still want more transparency from OpenAI. Hannaneh Hajishirzi, a professor at the University of Washington and senior director of AI at Ai2, mentioned that "The quality of their model seems good," but she urged the company to release the full training recipe. This report on the new models was compiled by a reporter at Informa TechTarget who covers AI software and systems.

Furthermore, this move follows a major announcement from Meta CEO Mark Zuckerberg on July 30. He revealed that the social media giant will not be making its anticipated superintelligence models open. This shift in Meta's strategy makes OpenAI's decision to provide open weights even more significant for developers looking for high-quality, accessible alternatives.