OpenAI has shaken up the artificial intelligence sector with the release of GPT-OSS-120B and GPT-OSS-20B, two open source models that can run on both high-performance servers and personal computers and mobile devices High-end. This move represents a shift from their traditional approach of keeping their AI systems under proprietary control and responds to the growing demand for flexible, auditable, and truly accessible tools.
The announcement has generated excitement in the technology community, as This is the first time since 2019 (with GPT-2) OpenAI is releasing models of these characteristics. Now, researchers, startups, companies, and independent developers have the opportunity to download and adapt these models free of charge and with commercial use permitted, an option that, until now, was only offered by a few competitors, such as Meta's Llama series or DeepSeek from China.
Two models, different approaches
The models presented cover different needs:
- GPT-OSS-120B has about 120.000 billion parameters and can run on a single 80GB GPU, making it ideal for research centers, universities, or technology companies with more advanced infrastructures.
- GPT-OSS-20B It is much lighter and works on computers with only 16 GB of memory, allowing its use to advanced users and small businesses that do not have large resources, even facilitating experimentation on laptops or decentralized environments.
Both models are deployed under Apache 2.0 license, which allows its modification, adaptation and commercial exploitation without the need for additional permits or payment of royalties.
The “mixture-of-experts” (MoE) architecture provides these systems with operational efficiency, since only a portion of the parameters are activated in each query, which reduces resource consumption and improves speed compared to other models with similar characteristics.
Capabilities and applications
Among the advantages of these models are:
- Chain-of-Thought Reasoning, without the need for direct supervised training.
- Support for structured functions y external tool calls such as web browsing or interpreting Python code.
- They allow you to adjust the level of reasoning depending on the context or task.
- Amplitude of context of up to 131.072 tokens, suitable for extensive and complex jobs.
- Support for automated workflows and agents, as well as text, code, and natural language response generation.
Also, these models can be fine-tuned by users for specific tasks, thanks to the technical guides and documentation provided. They are also ready to be integrated into open-source platforms such as Hugging Face, llama.cpp, or Ollama, and into cloud services such as AWS or Azure.
A return to open spirit and global competition
With this commitment, OpenAI seeks reconnect with the open source ecosystem, from which it had distanced itself in recent years. The initiative comes amid international competition, where Chinese AI labs such as DeepSeek and Alibaba, as well as Meta with its Llama series, have made a strong push to open up models, putting pressure on Western tech giants.
OpenAI executives emphasize that Privacy and control over AI are now in the hands of the end user. Sam Altman, CEO of the company, emphasizes that the goal is to maximize access and collective benefit, while Greg Brockman, co-founder, defends people's right to customize and run their own artificial intelligence within their own systems and firewalls.
This drive is reinforced by the collaboration with Amazon Web Services (AWS), which facilitates deployment at scale and increases the proposal's competitiveness against alternatives such as Google's Gemini or DeepSeek-R1. Internal tests indicate greater efficiency and a significant reduction in operating costs for business users.
Safety and risk assessment
The release of open weight models brings new security challenges. OpenAI has designed and released evaluation methodologies that simulate extreme malicious use scenarios, such as its use in biotechnology or offensive cybersecurity. To this end, both internal and external testing has been conducted, and the system has been fine-tuned to limit risks in these areas.
In comparative terms, the rates of incorrect responses or "hallucinations" are somewhat higher than in closed models, but the risk level has been considered low by the company itself and external entities. This allows for publication under secure conditions, with constant monitoring and evaluation.
A notable feature of these models is the transparency in the chain of reasoning, which facilitates auditing of behavior and helps prevent deviations or errors, increasing confidence in sensitive applications.
Where to download and how to get started?
The weights and documentation are freely available on Hugging Face and can be easily deployed on cloud platforms such as AWS, Azure, or Databricks. There's also a test site for experimenting with its capabilities before installation.
OpenAI offers instructions and resources to perform fine-tuning and facilitate integration into various AI tools, allowing both beginners and experts to take full advantage of GPT-OSS-120B and GPT-OSS-20B.
This launch represents a new stage in the democratization of artificial intelligence, allowing the community to analyze, adapt, and contribute to an open, flexible, and high-performance infrastructure. OpenAI reaffirms its leadership, promoting innovation across all sectors and returning control to those who want to work with AI on their own terms.