GPT-OSS-20B Locally: Requirements, Files and Ollama Setup

👁 121 reads

GPT-OSS-20B is OpenAI’s open-weight model that you can run locally, including through Ollama. It’s suitable for experiments with text, reasoning, and tools, but the model file alone doesn’t give you a ready-made equivalent of all ChatGPT features.

We checked the documentation and file sizes in the repository. We did not perform speed or quality tests on a specific computer for this guide. Below, we separate the manufacturer’s requirements from our own measurement of metadata and first-run instructions.

What the Name GPT-OSS-20B Means

OpenAI’s model card describes approximately 21 billion parameters total and 3.6 billion active parameters in a mixture-of-experts architecture. This doesn’t mean you only need to store the active portion on disk. The model is released under the Apache 2.0 license, and available reasoning levels are low, medium, and high.

The weights let you run the model independently, but you still need to choose an environment, handle conversation formatting, and verify results. We don’t treat the term “open” as a guarantee that every element of training data and the learning process has been published. We explain the difference between a model and an application in the AI models section.

How Much Space the Files Take and Why Numbers Vary

On September 9, 2026, we read the metadata from a specific revision of OpenAI’s repository. Three main safetensors parts totaled 13,761,316,904 bytes. Separately, the repository contained a file in the original directory and a Metal version. These are data variants for different execution methods, not instructions to download everything at once.

We calculated that downloading all these groups would mean 41,273,504,288 bytes of weights alone, approximately 41.27 GB decimal. That’s roughly three times one variant. The tokenizer and configuration are additional. This measurement concerns files in the pinned revision, not Ollama’s transfer or memory occupied during response generation.

In Ollama’s library, the 20b variant is described as approximately 14 GB. Different unit notation, package format, and file overhead can produce different numbers in the interface. Download size, free disk space, and RAM are three different pieces of information. We don’t choose a graphics card based solely on file size.

Is a Computer with 16 GB Memory Enough

OpenAI describes running the MXFP4 variant in 16 GB of memory. This is a requirement for a specific execution method, not a promise of smooth operation on every laptop with that amount of RAM. The system, other programs, context buffer, and environment used also need memory.

Before buying hardware, we would check GPU compatibility, available free memory, and a small test in the target environment. First a short conversation, only then a long document. If your goal is simply occasional writing assistance, an easier reference point might be free ChatGPT. We choose a local model when independent execution has concrete value.

How to Run GPT-OSS-20B Through Ollama

This is an instruction from the documentation, not a record of an installation we performed here. Install current Ollama from its official site and make sure you have space for the model. Then in the terminal, use the command provided in the library:

ollama run gpt-oss:20b

On first run, the model will need to be downloaded. For a test, type a short text and ask for a summary without adding facts. Record the time and check whether the response preserved numbers and names. Don’t start with confidential documents or automatic access to your entire disk.

Ollama’s documentation describes a command to check loaded models:

ollama ps

The PROCESSOR column shows whether CPU, GPU, or both are being used. This lets you distinguish a configuration problem from an assumption that the model “is always slow.” Also save the model name, environment version, and configured context. Without these, someone else’s tokens-per-second number tells you little about your computer.

Harmony Format and Tools Require Proper Handling

OpenAI’s repository describes the Harmony format and example environments. With independent integration, it’s not enough to glue arbitrary text fragments together and assume conversation roles will be recognized correctly. Use format handling prepared for this model.

A tool call by the model and actual operation execution are two separate stages. The application must provide the function, receive the request, and return the result. Only adding a search engine allows fetching current pages, and only adding a file tool gives access to files. The model doesn’t get these capabilities from the name “agentic” alone.

Where Local Models Fall Short and What to Check

We’re not providing quality results that we didn’t measure here. We would prepare our own test set: name declension, data preservation in summaries, and a question that has no answer in the provided material. In the last one, we check whether the model admits lack of information instead of adding a plausible-sounding answer.

Being local doesn’t automatically solve all privacy issues. Separately check the interface, logs, add-ons, and external tool calls. A chat connected to a remote server still sends content to it, even if it uses an open-weight model. For a first test, a made-up document is sufficient.

How to Decide Whether to Stick with GPT-OSS-20B

Choose one repeatable process, for example summarizing your own notes. Prepare a fixed instruction and examples of acceptable and incorrect responses. You’ll find an instruction template in the guide what is a prompt. Compare the number of corrections needed with the tool you already use.

If the result is correct and handling fits within your time and resources, only then increase material length or add tools. When the model loses facts in a short note, larger context isn’t proof the problem is solved. First determine whether the instruction, data, configuration, or the response itself is failing.

See also: Local AI: text models, images and video on your own computer · test of 7 local models

// read moreRun Llama Locally: Ollama Setup, VRAM Requirements & BenchmarksGuidesOllama: How to Run AI Models Locally on Your ComputerGuidesbasal-omni: Run the Decision Model Locally on an RTX 4090AI Tools

Kinetyra