OpenAI just debuted a pair of open-weight foundation models: a larger, more capable gpt-oss-120b and a light-weight gpt-oss-20b. Built to run efficiently on local hardware and minimal infrastructure, both models perform quite well on tasks like coding, medical question answering and agentic tool use. In many cases they match (or even surpass) OpenAI’s o-series models, only they’re free to download and openly available.
What Are the gpt-oss models?
The gpt-oss models — known as gpt-oss-120b and gpt-oss-20b — are open-weight language models developed by AI company OpenAI. Available under the Apache 2.0 license, the models are designed for tasks like coding, reasoning and math, and optimized to run efficiently on local hardware.
Despite what its name suggests, OpenAI has not released an “open” AI model of any kind since 2019, when it came out with GPT-2. In its early days, the company championed transparency and open research, but its approach shifted after the success of ChatGPT and the release of its underlying GPT-3.5 model in 2022. Since then, OpenAI has prioritized closed development and commercialization, where it sells access to its proprietary AI models via an API to enterprises and developers — a strategy that enabled it to become one of the most influential (and valuable) tech companies on the planet.
However, OpenAI’s dominance came under threat in early 2025, when Chinese AI startup DeepSeek released DeepSeek-R1, a comparatively small, open-weight model that outperformed some of the world’s most advanced players, but at a fraction of the cost. R1 upended the entire generative AI industry and reignited enthusiasm for open-source AI, particularly in the United States. Even President Donald Trump has called on American developers to open up more of their models. Adding to the pressure is rival tech giant Meta, a major player in open-weight AI that’s reportedly been poaching some of OpenAI’s top talent with multi-million dollar contracts.
Amid all of this, OpenAI seems eager to return to its original mission of building open AI — and eventually artificial general intelligence (AGI).
“Going back to when we started in 2015, OpenAI’s mission is to ensure AGI that benefits all of humanity,” CEO Sam Altman said in a statement shared with reporters. “To that end, we are excited for the world to be building on an open AI stack created in the United States, based on democratic values, available for free to all and for wide benefit.”
Both gpt-oss-120b and gpt-oss-20b are now available for download on Hugging Face under the Apache 2.0 license.
What Are gpt-oss-120 and gpt-oss-20b?
Gpt-oss-120b and gpt-oss-20b are two of OpenAI’s newest open-weight language models. Their names reflect the approximate number of parameters they have, with parameters being the internal values that determine how the model processes and generates text.
The defining feature of a so-called “open-weight” model is that its “weights” — the numerical parameters that shape how it processes information — are publicly available, allowing anyone to peek under the hood and inspect the model’s inner workings. However, key details surrounding its architecture and training methods remain hidden. In the case of gpt-oss-120b and 20b specifically, OpenAI views them as “complementary” to the company’s proprietary models, co-founder Greg Brockman told Wired, noting that open-weight models have a “very different set of strengths.”
OpenAI says it trained gpt-oss-120b and 20b on a mostly English, text-only data set, with a focus on STEM, coding and “general knowledge,” incorporating similar training and alignment techniques to the company’s proprietary o3 and o4-mini models. Both gpt-oss models are designed for a wide range of use cases, including on-device applications and agentic workflows.
How Do gpt-oss-120b and gpt-oss-20b Work?
Both gpt-oss models are built on a mixture-of-experts (MoE) architecture, a design that routes each input through a small number of specialized sub-models, or “experts,” rather than activating the entire network. This allows the model to perform at the same level as a standard transformer-based large language model without incurring the full computational cost. Other top open-source models like DeepSeek-R1 and Mistral take the same approach.
Specifically, gpt-oss-120b contains 128 individual experts but activates just 4 experts per token, engaging 5.1 billion active parameters out of 117 billion total. Meanwhile, gpt-oss-20b activates only 4 of its 32 total experts, activating just 3.6 billion of its 21 billion parameters. This setup means gpt-oss-120b can run on a single 80GB GPU, and the 20b version works on devices with as little as 16GB of memory — a major advantage for researchers and developers who want high performance without the massive infrastructure. These models also open the door for edge AI use cases, where models can run on smaller, local devices like an appliance or a smartphone.
In addition to MoE, both models support chain-of-thought (CoT) reasoning, a technique where a model goes through multiple steps to produce an answer to a prompt instead of just providing an answer straight away. OpenAI first used CoT in its proprietary o1 reasoning model back in 2024. Its goal is to enable better performance on complex tasks that benefit from intermediate thinking, such as multi-step math problems or logic puzzles.
Other technical features of gpt-oss-120b and gpt-oss-20b include:
- Alternating sparse and dense attention patterns, which help with long-context performance.
- Grouped multi-query attention for faster, more memory-efficient inference.
- Rotary Positional Embeddings (RoPE), a method that helps language models better understand the order of words in a sentence — something most neural networks don’t naturally grasp. It’s especially helpful in models with long context windows, which both gpt-oss models have (128,000 tokens).
What Can gpt-oss-120b and gpt-oss-20b Do?
Both gpt-oss models are text-only, so they cannot process or generate things like images or audio. But they can browse the web, call on other cloud-based models to help with certain tasks, execute code and navigate external software as an AI agent.
Some of the models’ more noteworthy capabilities include:
- Tool use and agentic workflows: The models can perform what is known as function calling, meaning they can interact with external tools, APIs or custom functions based on a user’s input. As a result, they can perform an internet search, execute code and look up data, among other things.
- Math and reasoning-based tasks: The models performed particularly well on benchmarks like MMLU (multi-subject reasoning), AIME (competition math) and GPQA (PhD-level science), so they’re especially well-suited for tasks that involve these strengths, such as tutoring, technical writing, code generation and research assistance.
- Health-related tasks: The models performed well on medical question answering tasks in the HealthBench benchmark. But OpenAI emphasized they are not intended or approved for diagnostic use.
- Customizable reasoning effort: Similar to OpenAI’s o-series models, the gpt-oss models allow users to toggle between three different reasoning efforts (low, medium and high) to control latency and depth of reasoning — and all through a single system prompt.
Are gpt-oss-120b and gpt-ooss-20b Safe?
Open-weight models like gpt-oss-120b and 20b come with a distinct set of safety concerns compared to proprietary systems. Because anyone can access and modify an open model, there’s no way to manage who uses it or for what purpose. A bad actor could fine-tune the model to bypass safety constraints, or use the model in harmful ways — and the developer (in this case OpenAI) would have no way to intervene or revoke access. In other words: Once an open model is out, developer control goes out the window.
Recognizing this, OpenAI delayed the release of its gpt-oss models several times, and took extra steps to evaluate their safety. In addition to the battery of tests the company typically runs on its proprietary models, it also customized the open-weight options to simulate how they could be misused or modified after being downloaded, which helped to flag potential risks.
In the end, the gpt-oss models did not reach a high level of risk as measured by OpenAI’s preparedness framework, which assesses things like a model’s potential for autonomous replication, ability to assist in cyberattacks or misuse for biological or chemical threats. That said, the company says developers and enterprises will likely need to implement “extra safeguards” when deploying them on their own, since OpenAI cannot control the models after they’ve been downloaded.
How Do gpt-oss-120b and gpt-oss-20b Compare to Other OpenAI Models?
In addition to being open-weight, gpt-oss-120b and gpt-oss-20b stand out against OpenAI’s other models in several ways.
Performance
OpenAI evaluated gpt-oss-120b and 20b across a range of academic benchmarks to see how they stack up against its proprietary models — including o3, o3-mini and o4-mini — in areas like coding, competition math, tool use and health knowledge. Both open-weight models held their own.
Gpt-oss-120b surpassed o3-mini and matched or exceeded o4-mini in several categories, including general problem-solving (MMLU, HLE), competition coding (Codeforces) and agentic tool use (TauBench). It also outperformed o4-mini in health question and answering (HealthBench) and competition math (AIME 2024 & 2025). Meanwhile, gpt-oss-20b (though much smaller) met or exceeded o3-mini’s performance on many of the same evaluations — and even outperformed it in areas like health and competition math.
The company also emphasized that both gpt-oss models offer significantly lower latency and compute costs than its other models. More details on how these models compare to other OpenAI models are available in the official model card.
Modality
Unlike OpenAI’s most advanced proprietary models — namely GPT-4o — the gpt-oss models are text-only. They aren’t equipped for multimodal tasks like image generation or voice interaction. However, developers can connect the open models to one of the company’s more capable closed models and make them multimodal that way.
Price
Because the gpt-oss models are open-weight and self-hostable, users can run them for free, without incurring API usage costs, making them a much more affordable option than OpenAI’s proprietary models. It also means developers have more flexibility in how they use the models, as they don’t have to be constrained by limited budgets or specific privacy requirements. Of course, the tradeoff is that OpenAI does not provide hosting or continual safety monitoring for these open models once they’re downloaded, so users must assume the full responsibility for how they are used and maintained.
Availability
Unlike OpenAI’s paid models, gpt-oss-120b and 20b are not available through the OpenAI API. Rather, users can download them for free from Hugging Face, and run them locally — no internet or cloud connection required.
How Do I Access gpt-oss-120b and gpt-oss-20b?
Gpt-oss-120b and 20b are freely available on Hugging Face under the Apache 2.0 license. They can run locally or through various third-party platforms, including Azure, Ollama, AWS, Together AI, Vercel, Cloudflare and Databricks. Individual developers and organizations can download and fine-tune these models, then integrate them into their custom applications or experiment with them for research purposes.
Microsoft, which is a major strategic partner of OpenAI, is also distributing optimized versions of gpt-oss-20b to Windows devices via ONNX Runtime, which are available through Foundry Local and Visual Studio Code’s AI Toolkit.
Although the gpt-oss models are not currently supported in the OpenAI API, the company says it may consider API support in the future.
Frequently Asked Questions
Are the gpt-oss models open source?
Technically, gpt-oss-120b and gpt-oss-20b are open-weight models, not open-source. Their weights — the numerical parameters that guide how they make decisions — are publicly available, and anyone can download, fine-tune and study them. But OpenAI has not released key details about the models’ architectures and training data.
Are the gpr-oss models free?
Yes, both gpt-oss-120b and gpt-oss-20b are available for free on Hugging Face.
Are the gpt-oss models safe?
OpenAI put its gpt-oss models through rigorous safety training and evaluations, beyond even what its proprietary models go though, and the company found that it did not reach a high level of risk. That said, all open-weight models come with some safety risks, as anyone can modify or fine-tune them for any purpose without any oversight.
