What is gpt-oss? OpenAI's new open license model family

OpenAI has re-entered the open AI model space with the release of gpt-oss-120b and gpt-oss-20b, its first open-source large language models since GPT-2 in 2019, under the Apache 2.0 license. These models, designed for text-based tasks, utilize a mixture-of-experts architecture and are noted for their efficient use of parameters, offering high performance without excessive computational demands. The models are trained primarily on English textual data, with sensitive content filtered out to enhance safety and are post-trained with techniques to avoid harmful outputs. They are available for download on platforms like Hugging Face and can be run on different hardware setups, with gpt-oss-20b being accessible on consumer-grade devices. OpenAI has collaborated with various inference providers to offer these models through APIs, facilitating integration into workflows via tools like Zapier, which allows users to automate and operationalize AI tasks across business applications. This release marks a significant shift for OpenAI, which had previously focused on proprietary models, and suggests a potential influence on global AI research and development practices given the models' novel architecture and open availability.