OpenAI Launches GPT OSS: New Open-Source AI Model Family for Developers
OpenAI has marked a significant strategic shift with the release of its new “GPT OSS” open-weight model family, featuring gpt-oss-120b and gpt-oss-20b. Announced on August 5, 2025, and available on Hugging Face, this move represents OpenAI’s first release of open-weight large language models (LLMs) since GPT-2 in 2019, signaling a notable pivot towards greater openness in the AI ecosystem.
The introduction of gpt-oss-120b and gpt-oss-20b is distinct from traditional open-source releases. These models are “open-weight,” meaning OpenAI has made available the numerical values that the models learned during their extensive training. This allows developers and researchers to download, utilize, fine-tune, and build upon these systems, offering substantial flexibility. However, it does not include the underlying code or the proprietary data used for their training, differentiating them from fully open-source models. The models are released under the permissive Apache 2.0 license, which grants users considerable freedom to modify and deploy them for various applications.
OpenAI asserts that the new GPT OSS models demonstrate strong performance, outperforming similarly sized open models in complex reasoning tasks and exhibiting robust tool-use capabilities. Specifically, gpt-oss-120b shows competitive performance in coding challenges, nearing the capabilities of OpenAI’s internal state-of-the-art reasoning model, o3. The smaller gpt-oss-20b also positions itself effectively, falling between o3-mini and o4-mini in benchmarks. Both models are designed for efficient deployment, even on consumer-grade hardware, making advanced AI more accessible.
This strategic shift aligns with OpenAI’s foundational mission to ensure that Artificial General Intelligence (AGI) benefits all of humanity. By making these powerful models more accessible, the company aims to foster new research avenues and facilitate the creation of innovative products. The release also reflects a move to align with U.S. policy priorities that advocate for building AI on “American rails” rooted in democratic values, and offers solutions for governments and institutions with stringent data residency or security requirements, allowing them to harness advanced AI locally.
The launch of GPT OSS models also intensifies competition within the rapidly evolving AI landscape. OpenAI, traditionally known for its closed-source, proprietary models like the GPT series, is now entering a field increasingly influenced by open-weight models from companies such as Meta (with its Llama series) and Mistral AI. While this release marks a significant step, OpenAI views it as an experiment, with the potential for future open-weight model releases contingent on the community’s engagement and utilization of these new offerings.
Technically, both gpt-oss-120b and gpt-oss-20b are built as Mixture-of-Experts (MoE) models, which can enhance efficiency. It’s important to note that these initial releases are primarily instruct or fine-tuned versions, and they do not currently support multi-modal inputs like images, video, or voice. The models are readily available for download on Hugging Face, with the platform’s blog providing detailed guidance on their usage and optimization.