On 6 August 2025, OpenAI officially launched two open-weight language models—gpt-oss-120b and gpt-oss-20b—and made their full parameters available for download on Hugging Face, Databricks, Azure, and AWS under the Apache 2.0 license. This marks the first time since GPT-2 in 2019 that OpenAI has released the weights of a large model. In press releases and media briefings, the company stressed that the new series “outperforms any open model in its class,” while CEO Sam Altman said in an email statement that it is “the crystallization of billions of dollars of research, aimed at giving developers everywhere access to the most powerful open model.”

After Six Years of Silence, OpenAI Opens Up Again

This release includes two sizes at once: gpt-oss-120b with 117 billion parameters and gpt-oss-20b with 21 billion. The larger model can run inference on a single 80 GB GPU, while the smaller one runs on a laptop with 16 GB of RAM. Both allow commercial redistribution, breaking the technical barrier of the previous closed-source API approach.

Since halting open-sourcing in 2019 over “security concerns,” OpenAI has faced criticism from the community. The new move answers Altman’s earlier admission that he had “taken the wrong side” and is widely seen as a strategic response to rivals such as DeepSeek, Qwen, and Llama, who have been releasing open models in quick succession.

Key Difference from GPT-4o: Offline Deployment vs. Cloud-Only

Unlike GPT-4o, the new gpt-oss models are fully “deployable” offline: all weights and inference code are released under Apache 2.0, allowing enterprises to download the models to on-premise GPU clusters or private clouds, run them offline, fine-tune them, and integrate them into their own toolchains. GPT-4o, by contrast, remains a “server-hosted” product; developers can only access it via OpenAI’s remote API and cannot see its underlying parameters.

Put simply, the gpt-oss series lets you take the model home, whereas GPT-4o still requires OpenAI’s backend to unlock its full power.

Industry Impact: Accelerating—and Contesting—the Open Ecosystem

The launch solidifies a dual strategy of “closed API plus open weights.” Enterprises can keep sensitive data on-premise by running or fine-tuning the model locally, then call higher-level OpenAI APIs for more complex tasks when needed. Smaller developers gain a zero-cost path to experiment with AI applications, compressing the cycle from idea to product.

At the same time, the Apache 2.0 license lowers commercial barriers and may trigger waves of secondary fine-tuning, vertical-industry variants, and hardware-optimized versions. Yet openness brings tougher challenges in misuse and regulation: creating a unified global safety baseline and preventing malicious fine-tunes will test the coordination skills of OpenAI and policymakers alike.