ChatGPT creator OpenAI is introducing two open-source models (gpt-oss-120B and gpt-oss-20B) that it claims surpass similar-sized models in handling complex tasks. The US-based frontier AI company stated that these gpt-oss models are its most capable open-source models so far. The release of these models has been eagerly awaited, alongside the anticipated launch of GPT-5. OpenAI’s founder, Sam Altman, had previously mentioned a delay in the release of the open-source model. Altman expressed excitement about offering this model, a product of extensive research, to the world to widely disseminate AI. He hopes this release will enable new research and products.
Since its inception in 2015, OpenAI’s mission has been to ensure AGI benefits all humanity. OpenAI looks forward to global development on an open AI stack made in the U.S., based on democratic values, and available for free for widespread benefit. This marks OpenAI’s first open-source models since 2019, positioning it against models like Meta’s Llama and Mistral’s AI models, amidst growing popularity of Japanese open-source models. Unlike closed models, open-source models are accessible to developers for enhancing their products and tools. OpenAI stated that the models were trained using reinforcement learning and advanced techniques from internal systems like o3. These models outperform similar open models in reasoning tasks, exhibit strong tool use capabilities, and are optimized for consumer hardware efficiency.
OpenAI emphasized that open-source models offer developers a range of tools to accelerate research, foster innovation, and support safer, transparent AI development. They also lower usage barriers in emerging markets. The models will be available for developers in Europe and other regions to freely download and run locally.