OpenAI Unveils GPT-OSS: First Open-Weight Language Models for Agentic Workflows

OpenAI has unveiled its first fully open-source large language model family, named GPT-OSS—short for Generative Pre-trained Transformer (Open Source Software). The release marks a significant shift for the company, which has previously kept its models proprietary. The new models are available under the permissive Apache 2.0 license, allowing broad access for developers, researchers, and organizations to use, modify, and distribute them. The GPT-OSS family includes two models that share the same underlying architecture but are optimized for different hardware capabilities. Both are mixture-of-experts (MoE) models, a design that improves efficiency by activating only a subset of parameters for each input, reducing computational load. They also leverage a 4-bit quantization scheme called MXFP4, which compresses model weights without sacrificing performance, enabling faster inference and lower memory usage. The larger model is designed to run efficiently on a single H100 GPU, making it accessible for research labs and enterprises with high-end hardware. The smaller variant is optimized to operate within 16GB of memory, a threshold that allows it to run on consumer-grade hardware, including high-end laptops and desktops. OpenAI emphasizes that these models are built with agentic workflows in mind—meaning they are particularly well-suited for applications that require autonomous decision-making, planning, and tool use. They perform strongly across tasks such as code generation, logical reasoning, multi-step problem solving, and natural language understanding. The release is seen as a strategic move to foster innovation in the open-source AI community while maintaining OpenAI’s leadership in the broader AI landscape. By providing high-performing, efficient models with open weights, the company aims to lower the barrier to entry for developers and researchers building next-generation AI applications.