On August 5, 2025, OpenAI introduced its first open-weight language models in more than five years: gpt-oss-120b and gpt-oss-20b. While not fully open-source in the traditional sense, these models are released under the Apache 2.0 license, allowing users to download the trained model weights, run them locally, fine-tune, or repurpose them freely.
OpenAI CEO Sam Altman announced the release on X (formerly Twitter), stating:
“gpt-oss is out! we made an open model that performs at the level of o4-mini and runs on a high-end laptop (WTF!!)(and a smaller one that runs on a phone). super proud of the team; big triumph of technology.”
Key features of the models
Two model variants are available under this release: gpt-oss-120b and gpt-oss-20b.
The 120b model is comparable in performance to OpenAI’s o4-mini, while the 20b model performs similarly to o3-mini. The larger model runs on a single high-end NVIDIA GPU, while the smaller one can operate on many consumer laptops with 16 GB RAM or even powerful smartphones.
Both models are distributed under the Apache 2.0 license, which permits commercial use and redistribution. They are available via platforms like Hugging Face or can be deployed on cloud services including AWS Bedrock, SageMaker, Azure, and Databricks.
Safety and testing process
Prior to the release, OpenAI delayed deployment twice in order to conduct rigorous safety evaluations. The company conducted adversarial simulations, exploring how a malicious actor might fine-tune the models to produce harmful content. These tests were reviewed by independent experts, who concluded that the models did not exhibit dangerous capabilities in sensitive areas such as bioweapons or cybersecurity.
Why this release matters
This launch signals a significant strategic shift for OpenAI, moving away from its traditionally closed API-based approach and toward broader accessibility for developers, researchers, and startups. It also places OpenAI in direct competition with other open-weight leaders, including Meta’s LLaMA, DeepSeek-R1, Alibaba’s Qwen, and Mistral.
By offering open weights under a permissive license, OpenAI empowers smaller teams to build advanced AI applications without the high costs associated with training models from scratch.
Technical details
The gpt-oss-120b model contains approximately 117–120 billion parameters, but due to its mixture-of-experts architecture and efficient quantization, only about 5 billion parameters are active per token. The gpt-oss-20b model requires just over 16 GB of RAM, making it suitable for modern laptops and some mobile devices.
In benchmark testing, gpt-oss-120b performs at the level of or better than o4-mini, while the 20b model performs near o3-mini. Both models demonstrate strong reasoning and tool-use capabilities, though they still show slightly more hallucination in factual tasks compared to the most advanced proprietary systems.
Comparison table
| Model | Parameter Count | Comparable Proprietary Model | Runs Locally | Ideal Use Case |
|---|---|---|---|---|
| gpt-oss-120b | ~120 billion | o4-mini | Yes, on high-end GPU | Advanced reasoning, complex tool integration |
| gpt-oss-20b | ~20 billion | o3-mini | Yes, on laptops/phones | Lightweight AI tasks, mobile applications |
Final thoughts
With this release, OpenAI moves toward greater openness and accessibility by providing trained model weights under the Apache 2.0 license. Although the training code and datasets remain closed, offering the model weights marks a significant departure from the company’s prior strategy of restricting access to its most powerful systems through proprietary APIs.
For independent developers and innovation-driven startups, this shift opens new opportunities to build state-of-the-art AI applications with significantly lower infrastructure demands.
