OpenAI stuns AI community with release of free powerful reasoning models

OpenAI, the leading developer of proprietary AI, has released two open reasoning models that users can download and run for free locally. The smaller 20 billion parameter gpt-oss-20b model can be handled by a decent home computer, while the larger gpt-oss-120b requires a powerful datacenter GPU with 80GB of VRAM to run efficiently.

The tech company, which is expected to unveil its next-generation GPT-5 model in the coming days or weeks, surprised the AI community by suddenly entering the open-source arena, which is thriving without OpenAI participation. Meta’s popular Llama family, China’s advanced DeepSeek and Qwen models, France’s Mistral, and numerous other challengers currently lead the way in open-source models.

OpenAI’s last open-weight model was GPT-2, released in November 2019, which was superseded by the proprietary GPT-3 and GPT-4 series.

“We are excited for the world to be building on an open AI stack created in the United States, based on democratic values, available for free to all and for wide benefit,” Sam Altman, OpenAI CEO, said on X.

The company says that its “state-of-the-art” models “outperform similarly sized open models on reasoning tasks, demonstrate strong tool use capabilities, and are optimized for efficient deployment on consumer hardware.”

Both models are released compressed with MXFP4 quantization, where the weights are quantized to 4.25 bits per parameter. The reduced precision significantly lowers memory requirements and speeds up the inference, but it might sacrifice some accuracy compared to full-precision models.

gpt-oss is a big deal; it is a state-of-the-art open-weights reasoning model, with strong real-world performance comparable to o4-mini, that you can run locally on your own computer (or phone with the smaller size). We believe this is the best and most usable open model in the…

undefined Sam Altman (@sama) August 5, 2025

The size of gpt-oss-20b is 14GB, which means that it could fit in a laptop's RAM. More memory will be required for the context window. OpenAI claims that this model delivers similar results to OpenAI’s o3‑mini on common benchmarks.

The gpt-oss-120b model “achieves near-parity with OpenAI o4-mini on core reasoning benchmarks, while running efficiently on a single 80GB GPU.” Its size is 65GB.

“We believe this is the best and most usable open model in the world,” Altman boasted on X.

“We're excited to make this model, the result of billions of dollars of research, available to the world to get AI into the hands of the most people possible.”

OpenAI claims that both models excel on tool use, few-shot function calling, chain-of-thought reasoning, and medical benchmarks.

Both models leverage mixture-of-experts technology to reduce the number of active parameters required to process each token. The larger gpt-oss-120b model activates 5.1 billion parameters per token, while gpt-oss-20b activates 3.6 billion parameters.

“We trained the models on a mostly English, text-only dataset, with a focus on STEM, coding, and general knowledge,” OpenAI said.

Before training the models, the company filtered out certain harmful data related to Chemical, Biological, Radiological, and Nuclear (CBRN). The models are trained to refuse unsafe prompts and defend against prompt injections.

The models are currently available on Ollama and AWS, among other distribution channels.