OpenAI Returns to Its Roots: First Open-Weight Language Models Released Since GPT-2

OpenAI has recently unveiled two new large language models—gpt-oss-120b and gpt-oss-20b—developed under an open-weight framework. This marks the company’s first non-commercial language model release since GPT-2 in 2019, signaling a renewed commitment to its founding vision of ensuring that the fruits of artificial intelligence benefit all of humanity.

Unlike traditional open-source models, the gpt-oss series does not disclose its training datasets or underlying code. Instead, it provides access to the post-training weight parameters—numerical values the model assigns during the learning process. While the absence of training transparency precludes full traceability, these open weights still allow developers to perform inference, fine-tuning, and build derivative applications atop the models.

According to Benjamin C. Lee, a professor of computer science at the University of Pennsylvania, open-weight models represent a “middle ground between proprietary black-box models and fully open-source alternatives.” They offer developers the ability to harness powerful language capabilities without bearing the prohibitive costs of training such models from scratch.

Of the two models released, gpt-oss-120b boasts approximately 120 billion parameters and delivers performance on par with OpenAI’s commercial-grade o3 model. It is intended for deployment on high-end hardware—ideally systems equipped with 80GB of GPU memory. By contrast, gpt-oss-20b is significantly more lightweight, requiring only a 16GB memory machine, making it ideal for local computing tasks such as offline coding assistance or text generation.

Both models are released under the Apache 2.0 open-source license, granting developers the freedom to modify, deploy, and commercialize their applications—thus lowering the barrier to AI adoption in both enterprise and academic settings.

Functionally, the gpt-oss series supports advanced capabilities such as chain-of-thought reasoning and tool use (e.g., executing Python code). While they do not natively support multimodal inputs such as images or voice, such functionalities can be integrated through external connections to OpenAI’s commercial services.

Notably, OpenAI conducted additional safety testing and optimizations prior to releasing the gpt-oss models. The company emphasized that these models are currently part of an experimental rollout, and their broader availability will depend on feedback gathered during initial testing phases. One of the first launch partners is AI Sweden, Sweden’s national center for applied artificial intelligence, which will assist in deployment and application testing.

This strategic release comes at a pivotal moment, coinciding with Meta CEO Mark Zuckerberg’s recent remarks indicating a shift away from open-sourcing future models—a move that places OpenAI’s renewed open-weight initiative in stark contrast.

Related Posts:

• OpenAI Postpones Open-Weight AI Model Release Over Security Fears; WindSurf Acquisition Collapses
• QNAP Addresses High Severity Vulnerabilities in License Center and Operating Systems
• OpenAI to Require ID Verification for Advanced AI Models