On Tuesday, OpenAI made a significant move by releasing two new open-source artificial intelligence models, marking the company’s first contribution to the open community since the launch of GPT-2 in 2019. The models, named gpt-oss-120b and gpt-oss-20b, are reported to deliver performance levels comparable to existing models o3 and o3-mini. Developed using a mixture-of-experts (MoE) architecture, these AI models have undergone extensive safety training and evaluation. Users can download the open weights of these models via Hugging Face.
OpenAI’s Open-Source AI Models Feature Native Reasoning
OpenAI CEO Sam Altman announced the launch in a post on X (formerly known as Twitter), emphasizing that “gpt-oss-120b performs about as well as o3 on challenging health issues.” Both models are currently available on OpenAI’s Hugging Face listing, allowing interested users to download and run the open weights locally.
According to information on OpenAI’s website, the models are compatible with the company’s Responses application programming interface (API) and can function in agentic workflows. They also support tools such as web search and Python code execution. Equipped with native reasoning capabilities, the models demonstrate a transparent chain-of-thought (CoT) feature, which can be tailored for either high-quality responses or faster output.
These models utilize the MoE architecture to optimize processing efficiency by activating a limited number of parameters. The gpt-oss-120b model engages 5.1 billion parameters per token while gpt-oss-20b activates 3.6 billion parameters per token. The former boasts a total of 117 billion parameters, whereas the latter has 21 billion, both accommodating a content length of 128,000 tokens.
Trained primarily on English text, OpenAI concentrated on various topics, including Science, Technology, Engineering, and Mathematics (STEM), coding, and general knowledge. Following the initial training, the models underwent reinforcement learning (RL)-based fine-tuning.
Benchmark performance of the open-source OpenAI models
Photo Credit: OpenAI
Internal assessments indicate that the gpt-oss-120b model surpasses o3-mini in areas such as competitive coding (Codeforces), general problem solving (MMLU and Humanity’s Last Exam), and tool usage (TauBench). However, both gpt-oss models perform slightly below o3 and o3-mini in certain benchmarks like GPQA Diamond.
OpenAI emphasizes that these models have undergone comprehensive safety training. In the pre-training phase, the company filtered out harmful data related to chemical, biological, radiological, and nuclear (CBRN) threats. The AI company also implemented techniques to ensure that the models decline unsafe prompts and are safeguarded against prompt injections.
Even though the models are open-source, OpenAI asserts that the training methods employed ensure they cannot be easily manipulated by malicious actors to produce harmful outputs.
