OpenAI has released its first open-weight language models since GPT-2, marking a significant shift in strategy.
The new models, called GPT‑OSS, are fully downloadable and commercially usable under the Apache 2.0 license.
Two variants are available:
gpt‑oss‑120b: A 120-billion parameter model that matches or exceeds the performance of OpenAI’s internal o4‑mini model.
gpt‑oss‑20b: A smaller model designed to run on consumer-grade hardware (as low as 16 GB RAM) while matching or outperforming the o3‑mini model.
Both models support key capabilities such as tool use, chain-of-thought reasoning, and dynamic compute scaling based on latency/performance needs.
OpenAI conducted extensive safety evaluations before release, including adversarial fine-tuning and audits under its Preparedness Framework.
The full methodology, model card, and evaluation results have been published.
The release responds directly to increasing demand for high-performing, open-access models—driven by competition from Meta (Llama 3/4), Mistral, and DeepSeek.
With this move, OpenAI is re-entering the open ecosystem with a focus on transparency, extensibility, and local deployment.
Developers can now access the models via Hugging Face, AWS, Azure, Databricks, and other platforms.
Support for optimized inference libraries (like vLLM and NVIDIA TensorRT-LLM) is already available.
This release gives teams full control over language model deployment, fine-tuning, and private use without API dependencies or closed licensing.
