OpenAI has released two open-source artificial intelligence (AI) models—marking its first such contribution to the open-source community since the release of GPT-2 in 2019. The models, named gpt-oss-120b and gpt-oss-20b, are now available for public access and download on Hugging Face, a leading platform for hosting machine learning models.
This release is significant not only for the open-source AI ecosystem but also for developers, researchers, and organizations seeking powerful AI tools without relying on proprietary infrastructure. According to OpenAI, these models deliver performance on par with its o3 and o3-mini models, making them some of the most capable open models available today.
MoE Architecture and Model Specs
Both gpt-oss models are built on Mixture-of-Experts (MoE) architecture, a design known for its computational efficiency. MoE allows the model to selectively activate different subsets of parameters for each input, reducing the cost of computation without compromising accuracy.
-
gpt-oss-120b features a total of 117 billion parameters, of which 5.1 billion are activated per token during inference.
-
gpt-oss-20b includes 21 billion parameters, with 3.6 billion active per token.
This architecture enables faster inference times and reduced energy consumption, making it highly suitable for real-time applications. Both models support a context length of 128,000 tokens, allowing them to process large inputs, such as long documents, codebases, or detailed conversations.
Native Reasoning and API Compatibility
One of the standout features of the gpt-oss models is native reasoning support, which enhances the model’s ability to explain its answers and follow logical, transparent chains of thought (CoT). This makes the models particularly suitable for high-stakes domains like healthcare, education, and law, where accountability and explainability are critical.
The models also support agentic workflows, enabling them to execute tasks using external tools such as Python scripts or real-time web searches. This opens up possibilities for AI agents that can reason, research, and act dynamically based on real-world inputs.
Importantly, the models are compatible with OpenAI’s Responses API, which simplifies integration into existing applications and services. This gives developers the flexibility to leverage powerful AI capabilities without needing proprietary closed models.
Benchmarking and Performance
According to OpenAI’s internal evaluations, the gpt-oss-120b model performs competitively with o3 and outperforms o3-mini in several key areas:
-
Codeforces (competitive coding)
-
MMLU and Humanity’s Last Exam (general problem-solving)
-
TauBench (tool calling)
However, the models showed slightly lower performance compared to o3 and o3-mini in benchmarks like GPQA Diamond, which tests graduate-level physics knowledge. Despite this, the overall performance gap is small, and the models still rank among the most capable open-source large language models (LLMs) currently available.
Safety-First Approach
In line with OpenAI’s emphasis on responsible AI development, these models underwent rigorous safety training. During pre-training, datasets were filtered to exclude harmful content, particularly relating to chemical, biological, radiological, and nuclear (CBRN) threats.
In the post-training phase, OpenAI applied reinforcement learning (RL)-based fine-tuning and prompt injection protection. As a result, the models are better equipped to:
-
Refuse unsafe or harmful prompts
-
Minimize the risk of being manipulated by adversarial inputs
-
Generate more responsible, ethical outputs
This level of safety refinement is uncommon in open models and represents a significant step toward safer deployment of AI tools at scale.
A New Era of Open-Source AI
The announcement was made by OpenAI CEO Sam Altman on X (formerly Twitter), where he emphasized that the gpt-oss-120b performs on par with o3, especially in challenging domains like healthcare. The release has already generated considerable excitement among AI developers and researchers, as it brings state-of-the-art capabilities into the hands of the global open-source community.
By open-sourcing these models, OpenAI is responding to calls for greater transparency and accessibility in AI research, especially at a time when closed, proprietary models dominate the industry. The move also helps democratize access to powerful language models, fostering innovation across academia, startups, and enterprises.
Download and Availability
Both gpt-oss-120b and gpt-oss-20b are now available on Hugging Face under OpenAI’s official listing. The open weights can be downloaded and run locally, offering full control and customization opportunities for developers and organizations.
This is particularly beneficial for those who need full data privacy, want to run models offline, or wish to fine-tune the base models for specific use cases.
OpenAI’s re-entry into the open-source space with two cutting-edge models is a milestone for the AI community. With MoE architecture, native reasoning, long context support, and built-in safety features, these models strike an impressive balance between performance, transparency, and responsible AI deployment.
As developers begin to explore and experiment with these models, we can expect a wave of innovation in education, coding, content generation, virtual assistants, and much more.
