GPT-OSS 120B: Capabilities, Hardware, and How to Use It Free

GPT-OSS 120B is OpenAI's larger open-weight reasoning model, released on August 5, 2025. OpenAI's model card describes a mixture-of-experts model with 116.8 billion parameters, 5.1 billion of them active per token, and a 131,072-token context window. It takes text and returns text, under the Apache 2.0 license.

OpenAI publishes the weights but doesn't run the model for you: GPT-OSS is not in ChatGPT or OpenAI's API. That leaves three routes. You can run it on a GPU with at least 60 GB of memory, use a service that hosts it, or use it in Maple, where it is on the Free plan with the full 131,072-token window, through our protected inference path.

This guide separates what belongs to the model from what belongs to each route, so you can pick one.

What is GPT-OSS 120B?

It is a sparse model. Each of its 36 layers holds 128 experts, and a router picks four of them for every token, which is why only 5.1 billion of its 116.8 billion parameters are active for each token.

It does not shrink the memory you need. Every expert has to be available, because the next token may route to any of them. OpenAI's answer is MXFP4: it post-trained the model with the expert weights in this 4-bit format (4.25 bits per parameter), which brings the whole checkpoint to 60.8 GiB, a download of about 65 GB.

The model works through a problem before answering, and you can set how hard it thinks: low, medium or high reasoning. It was trained to use tools such as web browsing, a Python environment and functions a developer defines, but only when the app running it provides them. Its knowledge cutoff is June 2024, so it doesn't know about later events unless a tool supplies them.

OpenAI also notes that the model's raw reasoning isn't meant for end users. The card says it can contain hallucinated content and language that doesn't reflect OpenAI's safety policies, so developers should filter or summarize it before showing it.

Is GPT-OSS 120B open source?

The weights are. OpenAI publishes them on Hugging Face under the Apache 2.0 license, a standard permissive license, with reference code for running them. The accompanying usage policy is two sentences:

We aim for our tools to be used safely, responsibly, and democratically, while maximizing your control over how you use them. By using OpenAI gpt-oss-120b, you agree to comply with all applicable law.

OpenAI itself calls the models open-weight: the trained weights are public, while the model card describes the training data only as "a text-only dataset with trillions of tokens." If your definition of open source includes the data, GPT-OSS 120B doesn't meet it.

Can you use GPT-OSS 120B in ChatGPT or OpenAI's API?

No. OpenAI's Help Center is direct about it, for both GPT-OSS models and the Safeguard variants:²

These models are not served through the OpenAI API and are not available in ChatGPT.

OpenAI's developer docs also have a model page for GPT-OSS 120B listing its specifications. Use it for the specs. For access, you run the weights yourself or use a service that hosts them.

How good is GPT-OSS 120B?

In OpenAI's own evaluations, close to its small proprietary reasoning models. The model card's summary is that it "surpasses OpenAI o3-mini and approaches OpenAI o4-mini accuracy." These are OpenAI's results: it chose the benchmarks and ran them at the high reasoning setting, and they are not Maple measurements.

The pattern is consistent: slightly behind o4-mini on most tests, ahead of o3-mini on all five, and ahead of both on HealthBench. OpenAI adds that the models do not replace a medical professional and are not intended for diagnosing or treating disease.

The card is also candid about weaknesses. On SimpleQA, a test of short factual questions, it answers correctly less often than o4-mini and hallucinates more. Check its factual claims, especially about anything after June 2024.

Should you use GPT-OSS 120B or 20B?

OpenAI released two sizes together, and the choice comes down to your hardware. OpenAI says the 20B can run on systems with as little as 16 GB of memory; the 120B is the stronger model.

If your machine has less than 60 GB of GPU or unified memory, 20B is the realistic choice. If you can reach that much memory or use a hosted route, 120B gives you the stronger model.

What hardware does GPT-OSS 120B need to self-host?

At least 60 GB of GPU memory, or of unified memory on a workstation, in OpenAI's own guides. Running the weights yourself gives you the most control over the model, the serving stack and where your data goes. The cost is hardware and setup.

We checked the download size ourselves. Hugging Face's file listing for the repository shows 15 weight files totaling 65.25 GB, which is 60.8 GiB in the binary units the model card uses.³ The repository also holds two more versions of the weights for OpenAI's reference implementations, in original/ and metal/, for 196 GB in all; most runtimes need only the first set.

OpenAI's Transformers recipe says the 120B model "requires ≥60GB VRAM or multi-GPU setup" and calls it ideal for H100-class hardware.⁴ In Transformers, its MXFP4 weights need NVIDIA's Hopper generation or later, which includes the H100 and the RTX 50 series. OpenAI's Ollama guide, written for PCs and Macs, says the 120B is "best with ≥60GB VRAM or unified memory."⁷

The vLLM recipe, an open-source serving project's guide, says the model fits on a single A100 80GB.⁸ It also lists the hardware it has verified, from NVIDIA H100 and B200 to AMD MI300X and a set of four Intel Arc Pro B70 cards.

NVIDIA's GeForce RTX 5090 carries 32 GB and the RTX 4090 24 GB.⁹ Neither holds GPT-OSS 120B on its own. OpenAI's Ollama guide says you can offload part of the model to the CPU when VRAM runs short, but expect it to run slower.

You may see community versions in other formats. Check their file size before assuming they help: OpenAI's expert weights are already 4.25 bits each, and its Ollama guide notes that the models ship in MXFP4 with no other quantization from OpenAI.

Self-hosting is the right call when:

• your information must never leave hardware you control
• you need to fine-tune, modify or deeply instrument the model
• you already run GPU servers and have capacity to spare

If none of those apply, ask whether the operational work buys you anything a hosted route can't. Our DeepSeek V4.1 Flash guide works through the same question for a far larger open-weight model.

How do you use GPT-OSS 120B in Maple?

Sign up and choose GPT-OSS 120B from the model picker in Maple Research. On the Free plan it is also the model behind Quick, the picker's option for fast, everyday answers. We configure it with the full 131,072-token window, the same as the upstream model, so a long document or a long working session fits in one conversation.

The composer's thinking control sets its reasoning level to low, medium or high; medium is the default. Use high for multi-step math, code or analysis, and low when you want a fast answer.

Plan details are on our pricing page. GPT-OSS 120B appears in our model library beside the other models we offer. On paid plans, Quick moves to GLM-5.3 Flash, which our GLM-5.3 Flash guide covers, and GPT-OSS 120B stays in the picker.

For developer tools and existing applications, Maple Proxy is our API route, using the model ID gpt-oss-120b. API access comes with Pro, Max and Team. Check the Proxy model list for its current context limit and features before you build on it.

This route gives you two things. You get a model that needs 60 GB of GPU memory without buying or renting that GPU. And your requests run through our protected inference path rather than a conventional hosted API.

The route has limits to plan around:

• Text only. GPT-OSS 120B doesn't read images. On paid plans, Research has another model describe an uploaded image first, and GPT-OSS 120B answers from that description. The Free plan doesn't include image upload.
• Knowledge stops in June 2024. For anything newer, turn on web search in the composer. On Free, it can run up to five searches per answer.
• The API needs a paid plan. The model is free in the app; API access starts at Pro.

What information reaches each route?

That depends on who operates the model: you, a hosting service, or Maple. Three questions get blurred together here. Who can read your prompts while they are processed, what happens to them afterward, and is any of it used for training?

Self-hosted weights. Your prompts and outputs stay on infrastructure you operate. OpenAI's Help Center notes that OpenAI doesn't receive the data you send to a copy you host yourself, unless you choose to share it. The trust boundary is your own: your access controls, logging and retention, which makes this the strongest option when information must stay on your hardware.

A hosting service. OpenAI partnered with deployment platforms ahead of launch, and many services now run GPT-OSS 120B.¹⁰ Each sets its own terms. Read them for the service and plan you would use: whether prompts are retained, who can access them, and whether they are used for training.

Maple. We run OpenAI's open weights inside an attested enclave. Your requests are encrypted before they leave your device, and our proof page explains the attestation you can check yourself.

The model runs in its own attested enclave, and Maple's enclave verifies it before forwarding a prompt. It works like certificate chains on the web: you check Maple's enclave, and Maple's enclave checks the model's. Our privacy notice covers everything outside your chats.

That design has a scope. It protects the contents of your chats: your prompts and the responses. It doesn't make information private if you've already shared it elsewhere.

It is also separate from the quality of the answer. A private route doesn't make the model's output more accurate, so review its code, facts and suggestions as you would any untrusted input.

How is GPT-OSS Safeguard different?

GPT-OSS Safeguard 120B is a separate model with a narrower job. OpenAI released it on October 29, 2025, as a research preview fine-tuned from GPT-OSS, under the same Apache 2.0 license.¹¹ You give it a written policy and the content to check, and it returns a verdict with its reasoning.

It is built for moderation and trust-and-safety pipelines, not conversation; OpenAI recommends against making it the core of what end users interact with. We offer it as gpt-oss-safeguard-120b through Maple Proxy only, not in the Research picker.

Which route fits your task?

It depends on where your information may go, how much hardware you can run and whether you need the API. The table compares the three routes on the same terms.

A rough decision rule:

• Choose self-hosting if information must stay on your hardware, you need to fine-tune the weights, or you already have 60 GB of memory to spare.
• Choose a hosting service if your prompts aren't sensitive and its terms fit your data.
• Choose Maple if you want GPT-OSS 120B without a 60 GB GPU and without a host that can read your chats.

Sources

1. OpenAI, gpt-oss-120b & gpt-oss-20b Model Card, August 5, 2025. Source for the architecture, parameter and expert counts, MXFP4, checkpoint size, context length, knowledge cutoff, training data description, reasoning levels, tools, the chain-of-thought guidance, the benchmark and SimpleQA results at high reasoning, and the medical caveat.
2. OpenAI Help Center, “OpenAI open-weight models (gpt-oss)”. Source for the models not being served in ChatGPT or the OpenAI API, the open-weight terminology, free use under Apache 2.0 and the usage policy, and self-hosted data not reaching OpenAI.
3. Hugging Face, repository file listing for openai/gpt-oss-120b. Our sum of the 15 safetensors files (65.25 GB, 60.8 GiB) and of the repository's original/ and metal/ weight files (196 GB in all).
4. OpenAI, “How to run gpt-oss with Transformers”, OpenAI Cookbook, August 5, 2025. Source for the 60 GB memory requirement, H100-class hardware and MXFP4 support on Hopper or later.
5. Maple, public model configuration and Research model picker. Source for the Free plan tier, Quick on Free, the configured context window, text-only input and the reasoning levels in Maple.
6. OpenAI, gpt-oss-20b model card on Hugging Face and repository file listing. Source for the 20B's parameter counts and our sum of its weight files (13.8 GB).
7. OpenAI, “How to run gpt-oss locally with Ollama”, OpenAI Cookbook, August 5, 2025. Source for 60 GB of VRAM or unified memory, CPU offload and MXFP4 as the only quantization OpenAI ships.
8. vLLM project, GPT-OSS serving recipe (source file), updated September 22, 2026. Source for the single A100 80GB fit and the verified hardware, including four Intel Arc Pro B70 cards.
9. NVIDIA, GeForce RTX 5090, GeForce RTX 4090 and A100 product pages. Source for 32 GB, 24 GB and 80 GB of GPU memory.
10. OpenAI, “Introducing gpt-oss”, August 5, 2025. Source for the release date and the deployment partners at launch.
11. OpenAI, “Introducing gpt-oss-safeguard”, October 29, 2025, and the gpt-oss-safeguard technical report. Source for the release, license, inputs and outputs, and OpenAI's recommended use.