gpt-oss-20b
Pilot $4,900 fixed
- Sizes
- ~21B MoE
- Active params
- 3.6B
- Context
- 128K
- Licence
- Apache 2.0
- Fits on
- 1× T4 16GB at 4-bit, 1× RTX 5090 32GB at FP8
- Endpoint
- llm.yourcompany.internal/v1
- Egress
- 0
- Owner
- YOU
gpt-osson your own GPUs.
OpenAI's open-weight gpt-oss models on hardware you own. The 120b model fits on one 80GB GPU in its native 4-bit format. The 20b model runs on a 24GB card.
This card is the installation. Pick a size on its edge and it is re-issued.
The 72-hour pilot with gpt-oss.
$4,900 fixed. Both are mixture-of-experts models with few active parameters, so they are cheap to serve for their size.
Live on your data
72 h$4,900 fixed
gpt-oss live on your data in 72 hours. The clock starts after access and prerequisites.
Your endpoint
/v1OpenAI-compatible
OpenAI-compatible endpoint. Your applications switch by changing the base URL.
Your hardware
- gpt-oss-20b1× T4 16GB at 4-bit, 1× RTX 5090 32GB at FP8
- gpt-oss-120b1× H100 80GB at 4-bit, 1× H200 141GB at FP8
The readout
Evalon your data
Reference evaluation on your data. Readout and production scoping.
We size the hardware and source it at cost, with no markup. Ballpark: one H100 80GB ≈ $25–30k; a 4×H100 node ≈ $180–240k; smaller models run on far less.
Same SDK. Change the base URL.
The model is one parameter. Everything else in your application stays as it is.
client = OpenAI(
base_url="https://llm.yourcompany.internal/v1", # inside your network
api_key="your-key",
)
reply = client.chat.completions.create(
model="gpt-oss-20b",
messages=[{"role": "user", "content": "Summarise this contract."}],
)Licence, plainly.
- gpt-oss-20bApache 2.0Commercial use: permitted
- gpt-oss-120bApache 2.0Commercial use: permitted
gpt-oss is released under Apache 2.0: commercial use, modification and private deployment are permitted.
OpenAI also publishes a usage policy alongside the weights. We go through it with you during scoping.
Published by OpenAI. We are not affiliated with the vendor. This is a summary, not legal advice.
Related models.
- Qwen3AlibabaCPU only at 4-bitApache 2.0
- Gemma 4Google2 models, from 1× T4 16GB at 4-bitApache 2.0
- MistralMistral AI4 models, from 1× T4 16GB at 4-bitApache 2.0 / Mistral Research License (MRL) (research only)
Talk to an engineer.
Tell us the model, the data and the GPUs you have. An engineer from Limassol or Auckland scopes the 72-hour pilot with you.