OpenAI Releases New Open-Weight gpt-oss Language Models
OpenAI has announced the release of gpt-oss-120b and gpt-oss-20b, two new state-of-the-art open-weight language models designed to deliver robust real-world performance at a low cost. Available under the permissive Apache 2.0 license, this move marks OpenAI’s first release of open-weight large language models since GPT-2 in 2019, signaling a significant step towards broader accessibility in the AI ecosystem.
The gpt-oss models are engineered to excel in reasoning tasks and demonstrate strong tool-use capabilities, outperforming similarly sized open models. The larger gpt-oss-120b model, boasting 116.8 billion total parameters with 5.1 billion active parameters per token, achieves near-parity with OpenAI’s proprietary o4-mini model on core reasoning benchmarks. Remarkably, this powerful model is optimized to run efficiently on a single 80 GB GPU.
For more accessible deployments, the gpt-oss-20b model, with 20.9 billion total parameters and 3.6 billion active parameters, delivers performance comparable to OpenAI’s o3-mini. Crucially, this smaller model can operate on consumer hardware with as little as 16 GB of memory, making it ideal for on-device use cases, local inference, and rapid development without requiring costly infrastructure. Both models leverage a 4-bit quantization scheme (MXFP4) for their Mixture-of-Experts (MoE) weights, which significantly reduces memory footprint and enables efficient inference.
The gpt-oss models are text-only autoregressive Mixture-of-Experts (MoE) transformers, built upon the foundational GPT-2 and GPT-3 architectures. They are designed for seamless integration into agentic workflows, featuring exceptional instruction following, advanced tool use such as web search and Python code execution, and customizable reasoning capabilities, including the ability to adjust reasoning effort for tasks requiring different levels of complexity or latency. Developers can also benefit from full Chain-of-Thought (CoT) and Structured Outputs, offering greater control and transparency over the models’ processes. Beyond general reasoning, these models show particular strength in areas like competition math, coding, and health-related queries, even surpassing some proprietary models on benchmarks like HealthBench.
This release is a notable development in the evolving AI landscape, where open-weight models are increasingly democratizing access to powerful AI technologies. By making these models available under the Apache 2.0 license, OpenAI enables developers and organizations to download, fine-tune, and deploy AI on their own infrastructure, reducing reliance on vendor-specific APIs and fostering greater control and customization. This shift aligns with a broader industry trend where open-weight systems are closing the performance gap with closed models, lowering barriers to entry, and accelerating innovation through community collaboration. The gpt-oss models are available for download on Hugging Face and can also be accessed via Amazon Bedrock and Amazon SageMaker AI on AWS, as well as through Ollama.
OpenAI emphasizes that safety remains a foundational aspect of its approach to releasing models, especially for open models where the potential for misuse exists once they are publicly available. The gpt-oss models underwent comprehensive safety training and evaluations, including testing adversarially fine-tuned versions. While designed to adhere to OpenAI’s safety policies by default, the company notes that developers and enterprises utilizing these models will need to implement additional safeguards to replicate the system-level protections typically built into OpenAI’s API-served models. This reflects a shared responsibility for ethical deployment as AI capabilities become more widely distributed.