OpenAI has announced the release of its first open-source AI model since 2019, marking the end of years of delay and speculation. The company released gpt-oss-120b and gpt-oss-20b on Tuesday, allowing users to download the two customisable large language models that don’t require a license fee or API gate.
Both models are text-only, appearing to be shared under an Apache 2.0 license. This means that anyone can download the model weights from platforms like GitHub and Hugging Face. They also work on LM Studio and Ollama, and can run on everything from laptops to cloud servers. OpenAI said the models are optimized to function across a range of hardware, including consumer devices and chips from Nvidia, AMD, Cerebras, and Groq.
OpenAI says its model can block malicious fine-tuning
According to its statement, OpenAI said it delayed releasing the models because safety evaluations were still ongoing. During pre-training, the company said it filtered sensitive material, specifically chemical, biological, radiological, and nuclear data. It also tested scenarios where bad actors might try to fine-tune the models for malicious tasks. The company reported that none of those experiments led to models reaching its internal “high capability” threshold, a classification outlined in its Preparedness Framework, which it uses to assess harm potential.
Greg Brockman, OpenAI’s president, told reporters, “It’s been exciting to see an ecosystem develop, and we are excited to contribute to that and really push the frontier and then see what happens from there.” He added that the company also included three external expert groups to audit and comment on the fine-tuning of safety tests. These models aren’t open-source in the traditional sense.
OpenAI said it will only release the model weights and not the full training code or datasets. It said they are still open enough to allow users to test, fine-tune, and deploy them however they want to. The company framed this as a middle ground between full transparency and maintaining some control. The models will also be available through cloud platforms.
Meanwhile, Amazon, Microsoft, and Baseten are offering gpt-oss-120b and gpt-oss-20b directly to customers. This is the first time that Amazon Web Services will be hosting OpenAI models, making them accessible through both Bedrock and Sagemaker. Amazon said these models will help customers build AI agents capable of advanced reasoning and step-by-step thinking. Amazon CEO Andy Jassy has made AWS into a marketplace where companies can pick from multiple AI providers, not just Amazon’s in-house tools.
In addition to OpenAI, Amazon has also partnered with Anthropic, putting $8 billion into the AI startup. The deal will allow AWS clients to access Anthropic’s Claude models. On Tuesday, Anthropic said it would release a new Claude model, which it claims is better at coding, research, and data analysis than its previous versions.
The release puts OpenAI in the same lane as other companies offering open-weight models, like Meta, Mistral AI (which is backed by Microsoft), and DeepSeek, a Chinese startup that previously gained attention for building an AI model with human-like reasoning.
Reacting to the news, Jensen Huang, CEO of Nvidia, said, “OpenAI showed the world what could be built on Nvidia AI — and now they’re advancing innovation in open-source software.” Nvidia partnered with OpenAI to make sure the models perform well across its hardware. So did AMD, Cerebras, and Groq, giving users more flexibility when choosing where to run the models.
