What is gpt-oss? OpenAI’s new open license model family

Big news: OpenAI is once again in the business of open AI models. The maker of ChatGPT recently released two large language models (LLMs) under open licenses for the first time since GPT-2 back in 2019. The naming and punctuation is deeply upsetting, but I can forgive it this time since gpt-oss-120b and gpt-oss-20b are incredibly exciting models. 

So, let’s dig in.

Table of contents:

What is gpt-oss?

gpt-oss is a family of open AI models, released under the permissive Apache 2.0 license. They’re state-of-the-art reasoning models that anyone can download, fine-tune, and use for almost any purpose—though OpenAI has taken steps to limit the ways they can be used for malicious purposes or to generate harmful information. 

This is a big deal because since 2019, all the GPT and o-series models have been proprietary. With gpt-oss, OpenAI pulled back the curtain.

It’s also worth noting that gpt-oss-120b and gpt-oss-20b are the highest performing open models from North American and European AI labs. For people with concerns over how Chinese models are trained and the censorship inherent in their training data, this makes OpenAI’s latest models even more important. 

gpt-oss-120b and gpt-oss-20b

gpt-oss-120b and gpt-oss-20b are the first models in the family, and aside from the fact that they’re not proprietary like OpenAI’s other models, they look pretty similar to the rest:

• Both models use a mixture-of-experts architecture. The larger model, gpt-oss-120b, has 117 billion total parameters across 128 experts, though it only activates 5.1 billion parameters at any one time. The smaller model, gpt-oss-20b, has 21 billion parameters across 32 experts, though it only activates 3.6 billion parameters at any one time. 

• Both models are reasoning models, so they’re capable of using chain-of-thought (CoT) reasoning to work through complex problems. They have three available levels of reasoning: Low, Medium, and High. 

• Both models are LLMs, not large multimodal models (LMMs), so they only support text—not audio, images, or any other modality.

• Both models support a context length of 128k tokens.

• Both models are capable of tool use, so they can be used to browse the web, write code, and work in agentic systems. 

The two models were mostly trained on English text that focused on STEM, coding, and general knowledge content. Data related to chemical, biological, radiological, and nuclear (CBRN) threats was filtered out to make them as safe as possible.

In addition to filtering the training data, OpenAI post-trained the models using deliberative alignment and the instruction hierarchy to make them refuse to answer unsafe prompts and limit the risk of prompt injection. These are the same techniques that OpenAI uses to make its proprietary models safe.

How good is gpt-oss?

Open language models are having a major moment. In particular, Chinese research labs like DeepSeek, Qwen, Moonshot, and Z.ai are releasing incredibly competitive open models. While the best proprietary models still outperform the best open models, the gap has closed substantially. So where do gpt-oss-120b and gpt-oss-20b fit into the mix?

Well, OpenAI claims that gpt-oss-120b offers similar performance to o4-mini and gpt-oss-20b offers similar performance to o3-mini on key benchmarks, and independent analysis largely backs this up. In other words, they’re really good modern reasoning models.

But this analysis misses a major detail. 

Right now, gpt-oss-120b is the most intelligent model that can be run on a single NVIDIA H100 graphics card, and gpt-oss-20b is the most intelligent model that can be run on a consumer GPU (or even a laptop with just 16GB of RAM). These models aren’t optimized for raw performance, but are instead designed to be incredibly intelligent for their number of parameters and experts. 

For example, DeepSeek R1 outperforms gpt-oss-120b, but it has 671 billion total parameters and 37 billion active parameters (compared to 117 billion total parameters and 5.1 billion active parameters), making it more than ten times more memory-intensive to run. You have to really want every extra ounce of performance for the extra cost to make sense. 

While the open language model space is rapidly evolving, I think it’s safe to say that both gpt-oss-120b and gpt-oss-20b are high performing, state-of-the-art models with exceptional efficiency. If OpenAI continues to support them or releases more models in the gpt-oss family, they’re likely to remain relevant for the foreseeable future.

How to use gpt-oss-120b and gpt-oss-20b

Like with most open models, you can download gpt-oss-120b and gpt-oss-20b from Hugging Face right now. While gpt-oss-120b requires a server-class GPU to run, you can get gpt-oss-20b running on many modern MacBooks.

Or, if you just want to try out the new models, they’re also available in OpenAI’s open model playground. 

OpenAI has also partnered with inference providers like Azure, vLLM, Ollama, LM Studio, AWS, Fireworks, Databricks, Vercel, and OpenRouter to make gpt-oss-120b and gpt-oss-20b available to developers. They offer the two models as APIs at various price points with a range of features suited to different application needs. 

Using a provider like OpenRouter in combination with Zapier lets you connect gpt-oss directly to all the other apps you use at work. With Zapier’s OpenRouter integration, for example, you can pull gpt-oss into all your workflows and build fully automated, AI-powered systems. Generate responses in OpenRouter based on updated spreadsheet rows, form responses, chatbot conversations, or any other source, and then send the response wherever you need it. Here are a few templates to show you how it works.

As open weight models, gpt-oss-120b and gpt-oss-20b can be fine-tuned for specific purposes. That’s possible both by downloading the models yourself or using a third-party inference provider. 

OpenAI does open AI again

Despite the name, OpenAI has largely been a closed shop for the past five years. Even though they produced some of the most powerful models available, they’ve kept very quiet about how they were doing it. gpt-oss-120b and gpt-oss-20b change that. 

If you’re interested in AI, there’s a lot that can be gleaned from the information available in the gpt-oss-120b and gpt-oss-20b release. For example, it’s fascinating how few active parameters the two models use. The difference in performance between the two models is down to the number of experts and overall number of parameters, not the number of active parameters. Will this affect how other AI research labs develop their models? Absolutely.

Automate OpenAI models with Zapier

With Zapier, enterprises can move beyond experimenting with AI models and start operationalizing it. Zapier acts as the orchestration layer—triggering the models from business apps, routing outputs to the right teams or systems, and chaining them with other AI tools.

This makes it easy to build automated, AI-powered workflows that scale. Zapier’s built-in AI is powered by GPT, and you can also connect directly to ChatGPT. You might have OpenAI summarize research and push insights into your CRM, or use OpenAI models to draft customer communications that are reviewed and sent automatically. With AI and automation working together, you get the flexibility to innovate quickly without sacrificing enterprise governance.

Learn more about how to automate Zapier with OpenAI, or get started with one of these pre-made templates.

Related reading: