OpenAI Unveils First Open-Weight AI Models Since GPT-2

OpenAI, a company often associated with its powerful, proprietary AI models, has marked a significant strategic shift by unveiling its first open-weight artificial intelligence models since the release of GPT-2 in 2019. This move introduces two new models, gpt-oss-120b and gpt-oss-20b, designed to democratize advanced AI capabilities by enabling local deployment and sophisticated reasoning tasks. Amazon has been announced as an early customer, integrating these models into its cloud offerings.

The newly released gpt-oss models are made available under the permissive Apache 2.0 license, signalling OpenAI’s intention to foster broader innovation and adoption within the AI community. The gpt-oss-120b, the larger of the two, demonstrates performance nearing that of OpenAI’s proprietary o4-mini model on core reasoning benchmarks, yet it is optimized to run efficiently on a single 80 GB GPU. Its smaller counterpart, gpt-oss-20b, delivers capabilities comparable to OpenAI’s o3-mini and can operate on edge devices with as little as 16 GB of memory, making it ideal for deployment on laptops and other consumer hardware. Both models are built upon a Mixture-of-Experts (MoE) Transformer architecture, with gpt-oss-120b activating 5.1 billion parameters per token out of 117 billion total, and gpt-oss-20b activating 3.6 billion parameters per token from its 21 billion total. They support an expansive context window of up to 128,000 tokens and excel across a range of applications including coding, competition-level mathematics, health-related queries, and agentic tool use, featuring full Chain-of-Thought (CoT) reasoning and tool execution capabilities.

This release signifies a notable change in OpenAI’s approach, which for years had prioritized closed, API-driven models. The decision to offer open-weight models is driven by a desire to lower barriers for developers, researchers, and organizations, particularly those in emerging markets or with resource constraints. By providing models that can be run locally on a user’s own infrastructure, OpenAI aims to grant greater control over latency, cost, and data privacy, eliminating the reliance on remote cloud APIs for sensitive operations. While these are “open-weight” models, meaning the trained parameters are publicly accessible for fine-tuning and local execution, they are not “open-source” in the strictest sense, as the underlying training data and complete source code are not released, a distinction OpenAI attributes to legal and safety considerations.

The gpt-oss models are readily available for download on platforms like Hugging Face and GitHub, supporting deployment on various operating systems including MacOS, Linux, and Windows Subsystem for Linux (WSL) 2.0. Beyond direct downloads, these models are also integrated into major cloud platforms. Amazon Web Services (AWS) has made gpt-oss-120b and gpt-oss-20b available through Amazon Bedrock and Amazon SageMaker JumpStart, marking the first time OpenAI models are hosted on Bedrock. Similarly, Microsoft is offering these models via Azure AI Foundry and Windows AI Foundry, further broadening their accessibility across diverse computing environments. This multi-platform availability underscores a growing industry trend towards a more open and flexible AI ecosystem.

OpenAI has emphasized that the release of these open-weight models followed extensive safety evaluations, including rigorous internal testing and independent expert reviews. The company implemented advanced post-training techniques, mirroring those used for its proprietary frontier models, to align the gpt-oss models with its safety standards and to enhance their reasoning and tool-use capabilities. To further address potential risks, OpenAI has launched a Red Teaming Challenge, offering a $500,000 prize to identify and mitigate safety issues within these new open-weight models.

This strategic pivot by OpenAI, making its powerful reasoning models accessible for local deployment, promises to accelerate AI research and development across a wider spectrum of applications, giving developers unprecedented control and flexibility.