The two new models operate under an Apache 2.0 license and excel in reasoning and speed.
OpenAI introduced GPT-OSS on Tuesday, a new series of language models with open weights. GPT-OSS is the company’s first open release since GPT-2 in 2019. The models come in two sizes: 117 billion and 21 billion parameters. They are expected to deliver similar performance to o4-mini and o3-mini.
Faster Thanks to Mixture of Experts
The models use a Mixture-of-Experts architecture, where only a portion of the sub-models is active for each prompt. This makes them faster and more efficient than traditional models of the same size. The 120B version runs on a single H100 GPU with 80 GB, while the 20B variant requires only 16 GB VRAM.
The models operate under an Apache 2.0 license, meaning they are as open as possible. OpenAI trained the models on English-language datasets focusing on science, coding, and general knowledge. They support context windows up to 128,000 tokens.
Security against Misuse
OpenAI states that it uses strict filters to prevent misuse, such as around hazardous chemical applications. Testers were unable to deceive the model during development, and there is a $500,000 reward for anyone who discovers a vulnerability.
Both models are freely available on Hugging Face. Meanwhile, OpenAI’s CEO Sam Altman hinted on X at a major upgrade this week. Could we perhaps get more information about GPT-5?