OpenAI finally gave the open-source community a model worth bragging about, and if you own an RTX 5090, you can run GPT-OSS locally right now, fully offline, with zero API bills and zero rate limits. In this 2026 guide, we walk through exactly how to run GPT-OSS locally on Windows using Ollama and LM Studio, what hardware you actually need, and how fast OpenAI’s open-weight models perform on NVIDIA’s flagship 32GB card.

What Is GPT-OSS? OpenAI’s Open-Weight Model Family

In August 2025, OpenAI released GPT-OSS, its first open-weight language models since GPT-2. The family ships in two sizes: gpt-oss-120b, a 117-billion-parameter mixture-of-experts (MoE) model with 5.1B active parameters that fits on a single 80GB GPU, and gpt-oss-20b, a smaller 21B-parameter model with 3.6B active parameters designed to run within 16GB of memory. Both are released under the permissive Apache 2.0 license, so you can use, modify, and even commercialize them freely. Both also support a massive 131,072-token context window, and both use native MXFP4 quantization on the MoE weights, making them the first open models to ship with day-one NVIDIA RTX GPU support for this format.

Test Config Specifications:

Monitor: MSI MAG342CQR E2 Curved Gaming Monitor
Motherboard: MSI PRO Z890-S WIFI Motherboard
CPU: Intel Core Ultra 7 265K
RAM: 32GB (2X16) Corsair Vengeance Performance DDR5 Memory
Primary SSD: Western Digital SN850 500GB PCIe Gen 4 SSD
Secondary/Game SSD: 480GB Crucial SATA SSD, WD Green 960GB SATA SSD
Power Supply: Antec HCG-1000-EXTREME PSU
GPU: INNO3D GeForce RTX 5090 X3 OC
CPU Cooler: DEEPCOOL GAMMAXX L360 ARGB
Cabinet: MSI MAG PANO 130R PZ
OS: Microsoft Windows 11 Pro

GPT-OSS-20B vs GPT-OSS-120B: Specs at a Glance

Model	Total Params	Active Params	Min. Memory	Context
gpt-oss-20b	21B	3.6B	16GB	131,072
gpt-oss-120b	117B	5.1B	80GB	131,072

Why the RTX 5090 Is Built for GPT-OSS

NVIDIA’s GeForce RTX 5090 packs 32GB of GDDR7 VRAM, more than enough headroom to run GPT-OSS locally at full precision without offloading layers to system RAM. Because GPT-OSS-20b needs only 16GB and GPT-OSS-120b’s MXFP4 MoE weights are natively supported on Blackwell-generation Tensor Cores, the RTX 5090 can load either model with room to spare for a long context window. NVIDIA’s own RTX AI Garage benchmarks clocked the card at up to 256 tokens per second on gpt-oss-20b, among the fastest consumer-GPU numbers recorded for an open-weight model of this class.

NVIDIA RTX AI Garage benchmark chart showing throughput when you run GPT-OSS locally on RTX AI PCs

How to Run GPT-OSS Locally With Ollama (Step-by-Step)

Ollama is the fastest way to run GPT-OSS locally on Windows, macOS, or Linux. It handles model downloads, GPU offloading, and quantization automatically.

Download and install Ollama from ollama.com for your operating system.
Open a terminal (or PowerShell on Windows) and pull the model.
Run gpt-oss:20b for the lighter model, or gpt-oss:120b if you have the VRAM headroom.
Once the download finishes, Ollama drops you straight into a chat prompt, fully offline and running on your RTX 5090.

Running GPT-OSS in LM Studio (Chat GUI Alternative)

Prefer a graphical interface over the command line? LM Studio lets you search, download, and chat with GPT-OSS through a polished desktop app. Open LM Studio, go to the model browser, search ‘gpt-oss’, and pick the 20B or 120B GGUF build. The listing shows minimum system memory (13GB for the 20B variant), tags, and a configurable reasoning-effort slider so you can trade speed for depth on the fly. LM Studio also supports Apache 2.0-licensed GPT-OSS out of the box, and on an RTX 5090 it offloads every layer to the GPU automatically.

LM Studio model browser page for gpt-oss-20b showing minimum system memory and Apache 2.0 license details

Ollama vs LM Studio: Which Should You Use?

Ollama’s model page for GPT-OSS lists a simple one-line command, ollama run GPT-OSS, plus ready-made integrations for Claude Code, OpenCode, and other agent frameworks. It is the better pick if you’re comfortable with a terminal and want the fastest possible setup on your RTX 5090. LM Studio trades a bit of that simplicity for a full GUI with chat history, model comparison, and a visual reasoning-effort control, which makes it the friendlier option if you’d rather point and click than type commands.

Ollama official model library page for gpt-oss with the ollama run gpt-oss command and model size table

Licensing, Costs, and Why Local Matters

Because GPT-OSS is Apache 2.0 licensed, running it locally costs nothing beyond your electricity bill, no API keys, no per-token pricing, and no data leaving your machine. That makes it a natural companion to other open models we’ve covered, like Google’s Gemma 4 12B, if you’re building a fully local AI workflow on your RTX 5090. For the full benchmark methodology and architecture notes, see NVIDIA’s official RTX AI Garage write-up.

Final Thoughts

OpenAI’s GPT-OSS models, paired with an RTX 5090, give you a genuinely capable, fully local ChatGPT-class assistant with no subscription and no data privacy concerns. Whether you go with Ollama’s one-line simplicity or LM Studio’s polished GUI, you’ll be up and running in minutes. Which one are you trying first? Let us know in the comments, and keep checking back on TechnoSports for more hands-on local AI guides, RTX benchmarks, and the latest from OpenAI, NVIDIA, and the open-weight model ecosystem.

If you’re planning your own local AI rig, the RTX 5090’s 32GB of GDDR7 VRAM headroom means you’re not limited to gpt-oss-20b either — with quantization and offloading tricks, ambitious users can experiment with the 120B variant too, though 20B remains the sweet spot for consumer hardware. Both Ollama and LM Studio receive frequent updates as the open-weight ecosystem matures, so expect performance and compatibility to keep improving through 2026.

Buy the INNO3D GeForce RTX 5090 X3 OC from here