👐 OpenAI Releases Its First Open-Weight Models Since GPT-2

The larger model, with 120 billion parameters, performs between o3 and o4-mini levels and is designed for launch in data centers.

The smaller one, with 20 billion parameters, is slightly less capable than o4-mini and optimized for laptops and consumer-grade machines.

Both models support advanced chain-of-thought reasoning and tool use, including web search and Python code execution. Users can adjust the level of reasoning effort: from fast and shallow (low) to deep and well-structured (high). This directly affects the quality of the model's responses.

Compared to Chinese open-weight models like DeepSeek-R1, Qwen 2.5, Yi-1.5, and GLM-4.1V-Thinking, OpenAI focuses on better reasoning and easier local deployment. Chinese models still lead in multimodality and the diversity of model sizes.

Both OpenAI models have undergone extensive safety evaluations, including external red-teaming. The company stresses that this is open weights, not full open source—datasets and training pipeline remain undisclosed.

🔜 Try the models here.

#OpenAI @hiaimediaen