gpt-oss is OpenAI's family of open-weight language models: you download the model files and run them on your own hardware. OpenAI released two versions, gpt-oss-20b and gpt-oss-120b, in August 2025 under the Apache 2.0 licence, which Ollama describes as ideal for experimentation, customisation and commercial deployment.
The smaller model is realistic on a well-specified gaming PC. This guide explains what each version needs and how to run it with Ollama or LM Studio. Forge PC hasn't tested gpt-oss performance on specific hardware; the requirements below come from OpenAI, Ollama and LM Studio.
gpt-oss-20b vs gpt-oss-120b
Both models use a mixture-of-experts (MoE) design, which means only part of the model is active for each token. We explain the idea in MoE vs dense LLMs.
| gpt-oss-20b | gpt-oss-120b |
|---|
Total parameters (OpenAI) | 21 billion | 117 billion |
Active parameters per token | 3.6 billion | 5.1 billion |
Context length | 128k | 128k |
Memory target (OpenAI) | 16GB | A single 80GB GPU |
Ollama download size | 14GB | 65GB |
Licence | Apache 2.0 | Apache 2.0 |
OpenAI says gpt-oss-20b delivers results similar to its o3-mini on common benchmarks and "can run on edge devices with just 16 GB of memory", while gpt-oss-120b achieves "near-parity" with o4-mini on core reasoning benchmarks. Those are OpenAI's own benchmark claims.
The reason the models are relatively compact is quantisation. Ollama explains that OpenAI post-trained the models with the MoE weights quantised to the MXFP4 format, about 4.25 bits per parameter, and that Ollama supports MXFP4 natively. For background, see LLM quantisation and GGUF explained.
What hardware do you need?
gpt-oss-20b: OpenAI targets 16GB of memory, and LM Studio lists 12GB as the minimum system memory. The Ollama download is 14GB, so a graphics card with 16GB of VRAM is the comfortable way to keep it all on the GPU. With less VRAM, Ollama will split the model between GPU and system memory, which is slower.
gpt-oss-120b: a 65GB download designed for a single 80GB data centre GPU. On a home PC that realistically means a machine with a very large pool of memory; see our coverage of Ryzen AI Max PRO 495 mini PCs with 192GB for one example.
Our best GPU for local LLMs and how much VRAM you need guides go into more detail.
Option 1: Run gpt-oss with Ollama
Install or update Ollama from ollama.com. Ollama's library page recommends the latest version for gpt-oss.
Open a terminal (Windows Terminal or PowerShell) and run:
ollama run gpt-oss:20b
Wait for the 14GB download to finish, then type a prompt.
For the larger model, use ollama run gpt-oss:120b. Ollama's gpt-oss:latest tag lists the same 14GB size as gpt-oss:20b.
To check whether the model fitted in your graphics card's memory, run ollama ps in a second terminal. Ollama's FAQ explains that the Processor column shows "100% GPU" when the model is entirely in VRAM, "100% CPU" when it's in system memory, and a split such as "48%/52% CPU/GPU" when it's shared.
Ollama's default context window is 4,096 tokens, according to its FAQ, far below gpt-oss's 128k maximum. You can raise it with the OLLAMA_CONTEXT_LENGTH environment variable or /set parameter num_ctx inside a chat, but more context uses more memory; see LLM context length and VRAM.
Option 2: Run gpt-oss with LM Studio
LM Studio is the easier choice if you'd rather not use a terminal.
Install LM Studio and open the model search.
Search for gpt-oss and choose openai/gpt-oss-20b.
Download it and load it in the chat window.
LM Studio's model page says it is based on GGUF and MLX builds of the model, supports a context length of 131k, and exposes a Reasoning Effort setting with low as the default. For more on choosing downloads, see how to choose models in LM Studio.
OpenAI designed gpt-oss with configurable reasoning effort: low, medium or high. Higher effort makes the model think for longer before answering, which can improve tricky answers but takes more time. Ollama and LM Studio both list this as a feature. The models also show their full chain of thought, which OpenAI says helps with debugging, though LM Studio notes this output "is not intended for end users".
Ollama's page also lists native function calling, structured outputs and optional web search, and you can use gpt-oss with tools such as Claude Code or OpenCode through Ollama's ollama launch commands. We cover similar set-ups in how to use Claude Code with Ollama.
Tips if it's slow
Our verdict
gpt-oss-20b is the version to run on a gaming PC: OpenAI targets 16GB of memory, the Ollama download is 14GB, and a 16GB graphics card is the comfortable way to keep it entirely in VRAM. Run it with ollama run gpt-oss:20b or download openai/gpt-oss-20b in LM Studio, and check ollama ps to see whether it fits on your GPU. gpt-oss-120b needs about 65GB and targets an 80GB GPU, so it's for high-memory systems only. Forge PC hasn't tested gpt-oss speeds itself.