OpenAI has launched two open-weight AI reasoning models, gpt-oss-120b and gpt-oss-20b, available for free on Hugging Face. These models offer state-of-the-art performance on various benchmarks, with the larger model usable on a single Nvidia GPU and the smaller model compatible with consumer laptops. This marks OpenAI’s first open language model since GPT-2, reflecting a shift towards open-source development influenced by competition from Chinese AI labs. The models employ mixture-of-experts technology for efficiency and have been trained using reinforcement learning.
Despite outperforming leading open models on select tasks, both models exhibit higher hallucination rates than OpenAI’s closed models. Released under the permissive Apache 2.0 license, the models’ training data will not be disclosed due to pending legal issues. OpenAI aims to align with democratic values and meet developer needs, indicating a potential strategic pivot in the AI landscape as it navigates safety considerations and regulatory scrutiny.
Source link
