OpenAI has unveiled GPT-OSS, a highly advanced open-weight AI reasoning model available for local deployment on personal computers and mobile devices. This release represents a milestone as OpenAI’s first major open-source large language models since GPT-2, designed to offer high-quality reasoning and performance comparable to proprietary counterparts like their o4-mini model.

Cutting-Edge Open-Source AI Models

The GPT-OSS family consists of two versions: GPT-OSS-20B and GPT-OSS-120B. The 20B model contains approximately 21 billion parameters and fits comfortably on devices with 16GB of VRAM, making it accessible for consumer laptops and edge devices. The larger 120B model, with about 117 billion parameters, requires more powerful hardware such as NVIDIA H100 GPUs but delivers significantly stronger reasoning capabilities, excelling in complex tasks such as math, coding, and domain-specific question answering.

Built with a Mixture-of-Experts (MoE) architecture and 4-bit quantization (mxfp4), these models offer efficient inference with reduced computational requirements, enabling fast, low-latency AI applications even on constrained hardware. They support a range of advanced features including instruction-following, chain-of-thought reasoning, tool use, and structured conversational formats.

Empowering Users with Local and Private AI

A core philosophy behind GPT-OSS is empowering individual users with full control over their AI. Unlike API-only closed models such as GPT-3.5 or GPT-4, GPT-OSS can be hosted and run locally, eliminating latency, cost concerns, and data privacy issues inherent in cloud-based services. This democratization allows users to customize, modify, and securely deploy AI for a wide variety of use cases without dependence on external providers.

OpenAI emphasizes that their models have undergone extensive safety and biosecurity evaluations. GPT-OSS’s performance on internal safety benchmarks is comparable to the company’s frontier models. This careful approach seeks to balance openness with responsibility, aligning with OpenAI’s mission to develop Artificial General Intelligence (AGI) that benefits all humanity.

Applications and Industry Adoption

The release of GPT-OSS is expected to accelerate innovation across multiple sectors. Enterprises can integrate these models into their workflows to build AI copilots that retrieve documents, invoke APIs, and automate data analysis with enterprise-grade governance and compliance features. The models are available across cloud platforms, including AWS, Azure, and Google Cloud, and on-premises environments, with full support for fine-tuning on domain-specific data.

Notably, Microsoft Azure AI Foundry supports deployment of GPT-OSS, allowing enterprises to spin up inference endpoints, customize models, and deploy solutions even in offline or hybrid cloud settings. The smaller 20B model’s efficiency makes it suitable for edge deployment on modern Windows PCs and smartphones, catering to scenarios requiring local, private AI inference.

Open Source, Open License, Open Future

GPT-OSS is licensed under the permissive Apache 2.0 license, encouraging broad community use, adaptation, and contribution. This openness marks a departure from OpenAI’s recent closed policies and represents a renewed commitment to transparency and accessibility in AI technology.

By offering these powerful open-weight models freely, OpenAI hopes to inspire a vibrant ecosystem of developers and researchers, boosting the AI field's innovation pace and expanding access to cutting-edge tools based on democratic and ethical values rooted in the United States.

Official Announcement