GPT-OSS · 9B to 32B · Mixture of experts
GPT OSS 20B
20.91 billion parameters in total, but only 4.18 billion are read for each token. That split is the whole point of the design: it costs the memory of a large model and the speed of a small one. 12 of its 24 layers use a sliding window of 128 tokens, so only 12 keep growing with the conversation.
About GPT OSS 20B
OpenAI's open-weight model from August 2025, under Apache 2.0. It is a mixture-of-experts model with about 4B parameters active and ships in the MXFP4 format, so its file is 12.1 GB whatever you do to it. It offers low, medium and high reasoning effort.
GPT OSS 20B is a mixture-of-experts model with 20.91B parameters, 4.18B of them active for each token, built from 24 layers with a hidden size of 2880. Its repository was published on 2025-08-04. The license listed on Hugging Face is apache-2.0. It is released in a native MXFP4 format, so the download is 11.28 GiB however it is labeled. 12 of its 24 layers use a sliding window of 128 tokens, so only 12 keep growing with the conversation.
Memory needed, by quantization
At an 8k context window. Quality is our rough ranking of how much the compression costs you: anything at or above Q5 is hard to tell apart from the original in normal use.
| Quantization | File size | KV cache | Total VRAM | Quality | What it costs you |
|---|---|---|---|---|---|
| MXFP4 | 11.28 GB | 0.19 GB | 11.55 GB | 98% | The 4-bit format some models, gpt-oss among them, are released in. It is their native precision, so there is no quality step down. |
What a longer conversation costs
Same model at MXFP4, only the context window changes. 12 of its 24 layers use sliding-window attention, so their cache stops growing past 128 tokens and only the other 12 keep growing.
| Context | KV cache | Total VRAM | Fits in 8 GB | Fits in 12 GB | Fits in 24 GB |
|---|---|---|---|---|---|
| 4k | 0.1 GB | 11.46 GB | no | no | yes |
| 8k | 0.19 GB | 11.55 GB | no | no | yes |
| 16k | 0.38 GB | 11.74 GB | no | no | yes |
| 32k | 0.75 GB | 12.11 GB | no | no | yes |
| 64k | 1.5 GB | 12.86 GB | no | no | yes |
| 128k | 3 GB | 14.36 GB | no | no | yes |
The "fits" columns allow for the roughly 0.8 GB Windows keeps for the desktop.
Which hardware runs GPT OSS 20B
75 of 118 consumer devices run it at a quantization worth using, at an 8k context window.
| Device | Memory | Verdict | Quantization | Used | Speed |
|---|---|---|---|---|---|
| GeForce RTX 5090 NVIDIA | 32 GB | Runs great | MXFP4 | 11.55 GB | 320.2 t/s |
| RTX PRO 6000 Blackwell NVIDIA | 96 GB | Runs great | MXFP4 | 11.55 GB | 320.2 t/s |
| GeForce RTX 4090 NVIDIA | 24 GB | Runs great | MXFP4 | 11.55 GB | 180.1 t/s |
| GeForce RTX 3090 Ti NVIDIA | 24 GB | Runs great | MXFP4 | 11.55 GB | 180.1 t/s |
| GeForce RTX 5080 NVIDIA | 16 GB | Runs great | MXFP4 | 11.55 GB | 171.5 t/s |
| RTX 6000 Ada Generation NVIDIA | 48 GB | Runs great | MXFP4 | 11.55 GB | 171.5 t/s |
| GeForce RTX 3090 NVIDIA | 24 GB | Runs great | MXFP4 | 11.55 GB | 167.3 t/s |
| GeForce RTX 5070 Ti NVIDIA | 16 GB | Runs great | MXFP4 | 11.55 GB | 160.1 t/s |
| GeForce RTX 5090 Laptop NVIDIA | 24 GB | Runs great | MXFP4 | 11.55 GB | 160.1 t/s |
| Radeon RX 7900 XTX AMD | 24 GB | Runs great | MXFP4 | 11.55 GB | 145.5 t/s |
| GeForce RTX 5080 Laptop NVIDIA | 16 GB | Runs great | MXFP4 | 11.55 GB | 137.2 t/s |
| RTX A6000 NVIDIA | 48 GB | Runs great | MXFP4 | 11.55 GB | 137.2 t/s |
| RTX A5000 NVIDIA | 24 GB | Runs great | MXFP4 | 11.55 GB | 137.2 t/s |
| Apple M3 Ultra 96GB Apple | 96 GB | Runs great | MXFP4 | 11.81 GB | 135.2 t/s |
| Apple M3 Ultra 256GB Apple | 256 GB | Runs great | MXFP4 | 11.81 GB | 135.2 t/s |
| Apple M3 Ultra 512GB Apple | 512 GB | Runs great | MXFP4 | 11.81 GB | 135.2 t/s |
| Apple M1 Ultra 64GB Apple | 64 GB | Runs great | MXFP4 | 11.81 GB | 132.1 t/s |
| Apple M1 Ultra 128GB Apple | 128 GB | Runs great | MXFP4 | 11.81 GB | 132.1 t/s |
| Apple M2 Ultra 64GB Apple | 64 GB | Runs great | MXFP4 | 11.81 GB | 132.1 t/s |
| Apple M2 Ultra 128GB Apple | 128 GB | Runs great | MXFP4 | 11.81 GB | 132.1 t/s |
| Apple M2 Ultra 192GB Apple | 192 GB | Runs great | MXFP4 | 11.81 GB | 132.1 t/s |
| GeForce RTX 4080 SUPER NVIDIA | 16 GB | Runs great | MXFP4 | 11.55 GB | 131.5 t/s |
| Radeon PRO W7900 AMD | 48 GB | Runs great | MXFP4 | 11.55 GB | 130.9 t/s |
| GeForce RTX 4080 NVIDIA | 16 GB | Runs great | MXFP4 | 11.55 GB | 128.1 t/s |
| Radeon RX 7900 XT AMD | 20 GB | Runs great | MXFP4 | 11.55 GB | 121.2 t/s |
Show 35 more devices
| GeForce RTX 4070 Ti SUPER NVIDIA | 16 GB | Runs great | MXFP4 | 11.55 GB | 120.1 t/s |
| GeForce RTX 4090 Laptop NVIDIA | 16 GB | Runs great | MXFP4 | 11.55 GB | 102.9 t/s |
| Radeon RX 9070 XT AMD | 16 GB | Runs great | MXFP4 | 11.55 GB | 97.7 t/s |
| Radeon RX 9070 AMD | 16 GB | Runs great | MXFP4 | 11.55 GB | 97.7 t/s |
| Radeon RX 7800 XT AMD | 16 GB | Runs great | MXFP4 | 11.55 GB | 94.6 t/s |
| Apple M4 Max 36GB Apple | 36 GB | Runs great | MXFP4 | 11.81 GB | 90.2 t/s |
| Apple M4 Max 48GB Apple | 48 GB | Runs great | MXFP4 | 11.81 GB | 90.2 t/s |
| Apple M4 Max 64GB Apple | 64 GB | Runs great | MXFP4 | 11.81 GB | 90.2 t/s |
| Apple M4 Max 128GB Apple | 128 GB | Runs great | MXFP4 | 11.81 GB | 90.2 t/s |
| Radeon RX 7900 GRE AMD | 16 GB | Runs great | MXFP4 | 11.55 GB | 87.3 t/s |
| GeForce RTX 5060 Ti 16GB NVIDIA | 16 GB | Runs great | MXFP4 | 11.55 GB | 80.1 t/s |
| RTX A4000 NVIDIA | 16 GB | Runs great | MXFP4 | 11.55 GB | 80.1 t/s |
| Radeon RX 6900 XT AMD | 16 GB | Runs great | MXFP4 | 11.55 GB | 77.6 t/s |
| Radeon RX 6800 AMD | 16 GB | Runs great | MXFP4 | 11.55 GB | 77.6 t/s |
| Arc A770 16GB Intel | 16 GB | Runs great | MXFP4 | 11.55 GB | 77.3 t/s |
| Apple M1 Max 32GB Apple | 32 GB | Runs great | MXFP4 | 11.81 GB | 66 t/s |
| Apple M1 Max 64GB Apple | 64 GB | Runs great | MXFP4 | 11.81 GB | 66 t/s |
| Apple M2 Max 32GB Apple | 32 GB | Runs great | MXFP4 | 11.81 GB | 66 t/s |
| Apple M2 Max 64GB Apple | 64 GB | Runs great | MXFP4 | 11.81 GB | 66 t/s |
| Apple M2 Max 96GB Apple | 96 GB | Runs great | MXFP4 | 11.81 GB | 66 t/s |
| Apple M3 Max 36GB Apple | 36 GB | Runs great | MXFP4 | 11.81 GB | 66 t/s |
| Apple M3 Max 48GB Apple | 48 GB | Runs great | MXFP4 | 11.81 GB | 66 t/s |
| Apple M3 Max 64GB Apple | 64 GB | Runs great | MXFP4 | 11.81 GB | 66 t/s |
| Apple M3 Max 96GB Apple | 96 GB | Runs great | MXFP4 | 11.81 GB | 66 t/s |
| Apple M3 Max 128GB Apple | 128 GB | Runs great | MXFP4 | 11.81 GB | 66 t/s |
| GeForce RTX 4060 Ti 16GB NVIDIA | 16 GB | Runs great | MXFP4 | 11.55 GB | 51.5 t/s |
| NVIDIA DGX Spark 128GB NVIDIA | 128 GB | Runs great | MXFP4 | 11.84 GB | 48.8 t/s |
| Apple M4 Pro 24GB Apple | 24 GB | Runs great | MXFP4 | 11.81 GB | 45.1 t/s |
| Apple M4 Pro 48GB Apple | 48 GB | Runs great | MXFP4 | 11.81 GB | 45.1 t/s |
| Apple M4 Pro 64GB Apple | 64 GB | Runs great | MXFP4 | 11.81 GB | 45.1 t/s |
| Radeon RX 7600 XT AMD | 16 GB | Runs great | MXFP4 | 11.55 GB | 43.6 t/s |
| Jetson AGX Orin 32GB NVIDIA | 32 GB | Runs great | MXFP4 | 11.84 GB | 36.6 t/s |
| Jetson AGX Orin 64GB NVIDIA | 64 GB | Runs great | MXFP4 | 11.84 GB | 36.6 t/s |
| Ryzen AI Max+ 395 32GB AMD | 32 GB | Runs great | MXFP4 | 11.84 GB | 35.3 t/s |
| Ryzen AI Max+ 395 64GB AMD | 64 GB | Runs great | MXFP4 | 11.84 GB | 35.3 t/s |
Can my GPU run GPT OSS 20B?
A direct answer for popular devices, each with the settings we would use.
How it compares
Other models in the same family or of a similar size, on a 12 GB card and a 24 GB card, each at the best quantization that fits.
| Model | Weights at 4 bits | GeForce RTX 3060 12GB | GeForce RTX 4090 |
|---|---|---|---|
| GPT OSS 20B | 11.3 GB | 65.7 t/s at MXFP4 | 180.1 t/s at MXFP4 |
| GPT OSS 120B | 59.0 GB | 16.9 t/s at MXFP4 | 23.9 t/s at MXFP4 |
| Gemma 4 31B Instruct | 17.6 GB | does not fit well | 35.5 t/s at Q4_K_M |
| Gemma 4 26B A4B Instruct | 14.5 GB | 37.6 t/s at Q4_K_M | 117 t/s at Q6_K |
| Qwen2.5 14B Instruct | 8.3 GB | 28.3 t/s at Q4_K_M | 48.3 t/s at Q8_0 |
| Qwen3.5 27B | 15.6 GB | does not fit well | 35.5 t/s at Q6_K |
What to buy to run GPT OSS 20B
The cheapest hardware that runs it at a quantization worth using and a speed you would not resent, at an 8k context window.
Cheapest that works
GeForce GTX 1650
$100 used
MXFP4 · uses 11.55 GB of its 4 GB · about 18.1 tokens/s
Best value
Arc A770 16GB
$280 used
MXFP4 · about 77.3 tokens/s · 560 GB/s
| Device | Price | Memory | Runs it at | Speed | |
|---|---|---|---|---|---|
| GeForce GTX 1650 NVIDIA | $100 used | 4 GB | MXFP4 | 18.1 t/s | Check price |
| GeForce GTX 1660 SUPER NVIDIA | $110 used | 6 GB | MXFP4 | 27.7 t/s | Check price |
| GeForce GTX 1080 Ti NVIDIA | $150 used | 11 GB | MXFP4 | 65.5 t/s | Check price |
| GeForce RTX 2060 12GB NVIDIA | $160 used | 12 GB | MXFP4 | 61.7 t/s | Check price |
| GeForce RTX 3060 12GB NVIDIA | $200 used | 12 GB | MXFP4 | 65.7 t/s | Check price |
| GeForce RTX 2080 Ti NVIDIA | $250 used | 11 GB | MXFP4 | 75.7 t/s | Check price |
| Arc A770 16GB Intel | $280 used | 16 GB | MXFP4 | 77.3 t/s | Check price |
| Radeon RX 6800 AMD | $330 used | 16 GB | MXFP4 | 77.6 t/s | Check price |
Indicative prices reviewed 2026-09-01; used prices are marketplace typical. Options that cost more than a cheaper one with no more memory and no more speed are hidden. Price links are Amazon affiliate links. Change the standard, the context window or the budget in the buying tool.
These numbers also cover
Fine-tunes share their base model's architecture, so they need exactly the same memory. If you are looking for one of these, the figures above apply unchanged.
unsloth/gpt-oss-20b-BF16
Where it sits in the catalog
Source: openai/gpt-oss-20b on Hugging Face. Downloaded 6.5M times in the last month. Published 2025-08-04. Architecture figures are read from the repository's own configuration file, so they move when the model does.
Questions people ask about GPT OSS 20B
How much VRAM does GPT OSS 20B need?
At MXFP4 with an 8k context window, GPT OSS 20B needs about 11.55 GB: 10.99 GB of weights, 0.19 GB of KV cache and 0.37 GB of overhead. At the highest-quality file it needs about 11.55 GB.
What is the cheapest GPU that runs GPT OSS 20B?
The smallest consumer device in our catalog that fits it is the GeForce RTX 5080 with 16 GB, running it at MXFP4 at roughly 171.5 tokens per second.
How much does a longer context window add for GPT OSS 20B?
On the memory side, a 32k window adds 0.8 GB of KV cache on top of the weights, and the cache grows more slowly than a conventional model because of its attention design. The longest window the model supports natively is 128k tokens.