In a landmark move, OpenAI has launched its new open-weight models, gpt-oss-120b and gpt-oss-20b, on Microsoft’s Azure AI Foundry and Windows AI Foundry platforms. This release marks the first time since GPT-2 that OpenAI is offering fully open weights, allowing developers and enterprises to deploy, customize, and control advanced AI models end-to-end, whether running in the cloud or directly on local devices.

AI Moves from a Layer to the Entire Stack

Today’s breakthroughs signal a fundamental shift: AI is no longer just another technology layer—it’s rapidly becoming the stack itself. Microsoft’s AI Foundry, spanning from the Azure cloud to Windows at the edge, reflects this new reality. The platform gives developers, regardless of organization size, the flexibility to create, fine-tune, and securely deploy intelligent agents across diverse environments.

Azure AI Foundry unifies the process of building, training, and scaling intelligent agents, while Foundry Local extends AI capabilities to billions of edge devices.

Windows AI Foundry, built on Foundry Local, is now deeply integrated into Windows 11, supporting fast, secure, local inference and development—bringing advanced AI directly to PCs and workstations.

What Makes GPT-OSs Different?

Unlike previous releases, gpt-oss models empower users with complete transparency and adaptability:

• Open weights: No black boxes—inspect, fine-tune, retrain, and optimize these models using proprietary or domain-specific data.
• Flexible deployment: Whether you need high-performance cloud endpoints or offline, secure local inference, gpt-oss is built to adapt.
• Hybrid AI: Mix and match open and proprietary models, optimize for cost, compliance, and performance, and run AI wherever your data lives—cloud or edge, online or offline.

Meet the Models: gpt-oss-120b and gpt-oss-20b

Both models perform robustly on core reasoning and agentic tasks, with gpt-oss-120b achieving near GPT-4 levels and gpt-oss-20b making advanced AI accessible to resource-constrained environments. Both will soon be API compatible with OpenAI's mainstream infrastructure, paving the way for seamless integration into existing applications.

Hybrid, Secure, and Democratized AI for Everyone

Azure AI Foundry is more than a model marketplace—it is a fully featured AI app factory, already hosting over 11,000 models. Developers can spin up secure endpoints or deploy refined, domain-specific variants in just a few commands. For those prioritizing privacy, Foundry Local ensures sensitive data never leaves the device, dramatically reducing compliance risks while eliminating cloud dependency, making real-time, responsive AI possible even when disconnected.

• Full-stack customization: Quantize, distill, and fit models to exact device requirements.
• Low-latency and privacy: Local inference on CPUs, GPUs, or NPUs enables instant response and robust data governance.

Why This Matters to Builders and Enterprises

For developers, these models unlock new avenues of innovation—inspect, adapt, and control every part of the pipeline. Business leaders benefit from a new level of sovereignty and flexibility: no lock-in, no trade-offs between performance and privacy, and far-reaching deployment options.

Advancing an Open, Responsible Future

Microsoft and OpenAI’s combined efforts represent a deep commitment to openness in AI. In addition to gpt-oss, Microsoft recently open-sourced the GitHub Copilot Chat extension, moving flagship developer tools toward community-driven innovation. Integration of leading open and proprietary models within Azure AI Foundry signals a future where anyone—from individual enthusiasts to global enterprises—can harness the world’s best AI, responsibly and securely.

Models pricing