OpenAI Releases Two Open-Source AI Models: gpt-oss-120b & 20b

OpenAI, the artificial intelligence titan behind ChatGPT, has announced a significant strategic shift by releasing two open-weight AI models, gpt-oss-120b and gpt-oss-20b. This marks the company’s first public release of freely available AI model weights since GPT-2 in 2019, breaking a six-year period of focusing on proprietary, closed-source models.

The new models are available for download on platforms like Hugging Face and are licensed under the permissive Apache 2.0 license, making them accessible for both commercial and experimental use. This move allows developers and enterprises an unprecedented ability to run, adapt, and deploy OpenAI models entirely on their own terms, eliminating the reliance on remote cloud APIs or exposing sensitive in-house data to external services.

Details of the New Models

The gpt-oss series comprises two distinct models, both built on a Mixture-of-Experts (MoE) architecture with a Transformer backbone, which enhances efficiency by activating fewer parameters per token.

• gpt-oss-120b: This larger model has 117 billion total parameters, activating approximately 5.1 billion parameters per token. It is designed for production, general-purpose, and high-reasoning use cases, delivering performance near-parity with OpenAI’s o4-mini on core reasoning benchmarks. Despite its power, it is optimized to run efficiently on a single 80 GB GPU, making it suitable for data centers and high-end desktops.

• gpt-oss-20b: The smaller, more efficient model has 21 billion total parameters, with about 3.6 billion active parameters per token. It is optimized for lower latency and local or specialized use cases, delivering results similar to OpenAI o3-mini on common benchmarks. This model can run on edge devices with just 16 GB of memory, making it ideal for on-device applications, consumer hardware, and rapid iteration without costly infrastructure.

Both models support a context length of up to 128,000 tokens, feature Chain-of-Thought (CoT) reasoning with adjustable effort, and are capable of strong instruction following and tool use, including web search and Python code execution. They are also natively quantized in MXFP4 for efficient inference.

A Return to Open Roots

OpenAI’s decision to open-source these models marks a significant departure from its recent strategy. After GPT-2, the company largely pivoted to a closed-source approach for models like GPT-3 and GPT-4, prioritizing proprietary releases and API access. This shift was driven by a combination of factors, including competitive advantage, safety concerns, and maximizing profits.

However, the landscape of AI development has evolved, with open-source models from companies like Meta (Llama) and Mistral gaining significant traction. OpenAI CEO Sam Altman has previously acknowledged that the company might have been “on the wrong side of history” regarding open-sourcing its software. This latest release suggests a response to mounting competitive pressure and a recognition of the benefits that an open ecosystem can bring.

Implications for the AI Landscape

This move by OpenAI is expected to have far-reaching implications:

• Democratization of AI: By making powerful models freely downloadable and runnable locally, OpenAI is lowering barriers for developers, researchers, emerging markets, and smaller organizations that may lack the resources for extensive cloud infrastructure.

• Enhanced Control and Privacy: Running models locally offers full control over latency, cost, and privacy, as sensitive data can be processed in-house without being sent to external servers.

• Fostering Innovation: Access to open-weight models under a permissive license encourages experimentation, customization, and fine-tuning on domain-specific data, potentially accelerating research and development across various use cases.

• Cost Efficiency: Local deployment eliminates ongoing API costs and subscription fees associated with cloud-based AI services, offering a more cost-effective solution for scalable AI usage.

• Increased Competition: OpenAI’s re-entry into the open-weight space intensifies competition, pushing the entire industry towards more transparent and accessible AI development.

OpenAI has emphasized that safety remains foundational to their approach, and these models have undergone comprehensive safety training and evaluations, including adversarial testing. While the models are designed to follow OpenAI’s safety policies by default, the company notes that developers and enterprises will need to implement extra safeguards to replicate the system-level protections built into their proprietary API models.

This release signifies a potential future where AI development balances proprietary advancements with a commitment to open tools and standards, ultimately aiming to accelerate innovation and democratize access to advanced AI capabilities.