OpenAI releases two open-source AI models that perform on par with o3, o3-mini. These models are called GPT-OSS-120B and GPT-OSS-20B.
This is OpenAI’s first open-source release since GPT-2. Both models offer smart reasoning and support tools. They are now free to use by developers and businesses worldwide, including in India.
also read: OpenAI Introduces New ChatGPT Agent to Automate Complex Tasks
OpenAI Models Match o3 and o3-Mini Performance
OpenAI launched GPT-OSS-120B and GPT-OSS-20B together. Both models offer similar results to o3 and o3-mini. These models are based on MoE (Mixture-of-Experts) design.
This design makes them fast and efficient while using fewer resources. The models are open for commercial and personal use under Apache 2.0 license.
GPT-OSS-120B uses 5.1 billion active parameters per token. GPT-OSS-20B uses 3.6 billion active parameters per token. The larger model has 117 billion total parameters. The smaller version holds 21 billion total parameters. These models can handle up to 1,28,000 tokens of content length. Both models were trained mainly using English text.
Training focused on science, coding, math, and general topics. After training, both models went through advanced fine-tuning. OpenAI used reinforcement learning to improve the model’s skills. The models are tested and perform well on many tasks.
Tool Support, Reasoning, and Local Use
OpenAI made these models suitable for offline use. They work without needing internet or server access. GPT-OSS-20B runs on just 16GB RAM. GPT-OSS-120B can run on a single Nvidia GPU. This makes them very easy to use at home or office.
These models work with OpenAI’s Responses API. They also fit into agent-based systems. They can run Python code and do web searches. Their reasoning ability is flexible depending on task type. They can give fast answers or detailed step-by-step solutions.
They support chain-of-thought reasoning for clear explanations. Users can set them to respond quickly or with depth. These models are available on Hugging Face, Azure, AWS, and Databricks. Anyone can download and start using them today.
Safety Training and Benchmark Results
OpenAI added strong safety features to these models. Harmful content was filtered during early training. This includes chemical, biological, and nuclear threat data. Models are designed to reject unsafe or harmful prompts.
They are protected against prompt injections and misuse. Even though they are open, they cannot be fine-tuned for harmful tasks. GPT-OSS-120B performs very well in coding and tool use. It outperforms o3-mini in Codeforces and MMLU tasks.
It does slightly lower in GPQA Diamond tests. But overall, it delivers strong and reliable results. OpenAI CEO Sam Altman said it performs well in health topics too. These models show great balance between safety and smart performance.
OpenAI releases two open-source AI models that perform on par with o3, o3-mini. These models offer smart reasoning, offline use, and tool support. They are safe, fast, and ready for developers.
With Apache 2.0 license, Indian users can now explore powerful AI. OpenAI has made advanced AI available for everyone to use freely.
