OpenAI’s open-source models available on IBM watsonx.ai

OpenAI has released two artificial intelligence models to the public, allowing developers and businesses to download, run and customize them freely. One of the models is now hosted on IBM’s watsonx.ai developer studio.

It’s the first time OpenAI has made this kind of move since 2019. Over the last five years, the company has focused almost entirely on offering its most powerful models, such as GPT-4, GPT-4o and its o3 and o4-mini reasoning models, through tightly controlled interfaces. This latest release puts a pair of capable models directly into the hands of users, who can run them locally or on their own infrastructure without restrictions on commercial use.

One of the two models, the larger gpt-oss-120b, is being made available today on IBM’s watsonx.ai platform. IBM said the model can be deployed securely within enterprise infrastructure, customized using internal data and embedded into real-world workflows in industries such as healthcare, financial services and legal support. The second model, gpt-oss-20b, will also be added to the platform soon. Development teams will also soon be able to leverage these models from watsonx.ai when building Agents within watsonx Orchestrate.

IBM watsonx.ai empowers businesses to develop, deploy and scale AI applications in a unified development platform. It gives developers the flexibility and choice to customize a foundation model, access a wide range of pre-built, custom and open-source tools, take advantage of leading frameworks and utilize hybrid deployment options for cloud and on-prem, enabling rapid experimentation and accelerating time to value for the business .to customize a foundation model, access a wide range of pre-built, custom and open-source tools, take advantage of leading frameworks and utilize hybrid deployment options for cloud and on-prem, enabling rapid experimentation and accelerating time to value for the business. 

The availability of gpt-oss on watsonx.ai adds another option to a growing menu of open models hosted on the platform and will add more choice for teams building agents in watsonx Orchestrate. By offering users access to top-tier open models in watsonx including third-party models such as Meta and Mistral and IBM’s own Granite, the goal is to foster a culture of collaboration and flexibility for customers deploying generative AI.

The larger model, gpt-oss-120b, contains 116.8 billion parameters and uses a Mixture of Experts (MoE) architecture, meaning only a portion of the model is active at any time. This design allows it to run efficiently on a single NVIDIA H100 GPU, which is widely used in data centers. The smaller model, gpt-oss-20b, is optimized for consumer-grade devices such as laptops with 16GB of memory.

Both models support adjustable reasoning levels—low, medium or high—so users can balance output quality against cost and speed. The models also output full chains of reasoning, allowing users to see how a conclusion was reached, rather than receiving only the final answer. OpenAI says this improves transparency and makes the models easier to debug and trust.

OpenAI published benchmark scores showing that the larger model performs competitively in several reasoning and math tasks. On the MMLU benchmark, a standard test of general knowledge, it scored 90. On the GPQA science benchmark, it scored 80.1. On the AIME 2025 math test, it achieved a score of 97.9, suggesting strong capabilities in symbolic reasoning.

The models do not include image, audio or video understanding, and they are not equipped with content filters or moderation systems. They are text-only, and users are expected to implement their own safeguards depending on how the models are deployed. Users can leverage Granite Guardian alongside gpt-oss to serve as a guardrail model that helps detect risks in prompts and responses.

The release of gpt-oss comes at a moment when open models are gaining momentum, and further underscores IBM’s open, multi-model strategy and the company’s dedication to providing both developers and business users with access to the most advanced AI technologies available. Like IBM’s own Granite, Meta’s LLaMA, Mistral and Alibaba’s Qwen family have all found wide adoption among developers and researchers looking for tools that can be audited, improved and deployed independently.

Open-weight models offer an alternative to commercial black-box systems, allowing for greater scrutiny of model behavior.

For OpenAI, the release of gpt-oss signals a willingness to engage with these dynamics, even as it continues to offer its most advanced systems only through commercial interfaces. The Apache 2.0 license gives users broad freedom to build on the models, without restrictions on redistribution or use in proprietary applications.

“The future is hybrid,” said Bruno Aziza, IBM Vice President for Data, AI & Analytics Strategy, in a recent episode of the Mixture of Experts podcast. “There’s not going to be one model to rule them all ... not going to be only closed or open. The enterprise is going to get both.”

By making gpt-oss-120B available in watsonx.ai, IBM is empowering developers, data scientists and enterprises to:

• Deploy open models on their own infrastructure with full control over data and compute
• Customize and fine-tune the model for specialized use cases—from healthcare to finance to customer support
• Build AI applications that leverage the model’s native tool-use capabilities

Whether you’re building internal assistants, automating workflows or exploring new use cases for generative AI, gpt-oss-120B on watsonx.ai offers the flexibility and power to do it all—securely and transparently. 

Explore gpt-oss-120B in watsonx.ai and start building with one of the most capable open models available.