Runs on a private AI box you own
One compact GB10-class device — 128 GB unified memory — runs it on-premise, sold and provisioned by us. No cloud API keys, no data leaving the building.
Run gpt-oss-120b on-premises
○ Memory-fit arithmetic — checkable math, not a deployment claim
gpt-oss-120b is OpenAI's open-weight flagship: about 117 billion total parameters with ~5.1 billion active per token (mixture-of-experts), released under the permissive Apache-2.0 license and shipped in a native MXFP4 4-bit format. It is designed to run on a single high-memory accelerator, which makes it the most comfortable big-model fit for the hardware we sell.
Here is the arithmetic on a GB10-class unit — the 128 GB unified-memory device we sell and operate.
Memory fit on a 128 GB unit
| Parameters | 117B total, ~5.1B active per token (MoE) |
|---|---|
| Native format | MXFP4 (~4-bit) — the format it ships in |
| Weights on device | ~63 GB |
| Device memory | 128 GB unified (one GB10-class unit) |
| Left for context cache + system | roughly 45–55 GB |
| Units required | 1 |
Because gpt-oss-120b ships in MXFP4, the ~63 GB weight figure is the format's own footprint, not an aggressive down-quantization — one unit runs it with generous context headroom.
What it serves well
- General reasoning and tool-use / agentic workflows on your own hardware.
- A strong default 'house model' behind chat UIs and coding agents on the box.
- Workloads where an Apache-2.0 license removes commercial-use friction entirely.
Honest limits
- This page is arithmetic, not our own benchmark: we have not published a production deployment of gpt-oss-120b. The memory envelope, however, is smaller than models we do run measured on this hardware (Qwen3.5-122B at 78 GB), so the fit is conservative.
- MoE models trade some raw depth for serving efficiency; whether that suits your task is something we test in the assessment rather than assert here.
Frequently asked questions
- Does gpt-oss-120b fit one GB10-class unit?
- Yes, comfortably: in its native MXFP4 format the weights are ~63 GB against 128 GB of unified memory, leaving tens of gigabytes for context and system.
- What license is it under?
- Apache-2.0 — a permissive license with no commercial-use restriction, which is part of why it is a strong default for a business box.
- Is it your production model?
- Not yet — we label this arithmetic, not measured. We run a larger-envelope model (Qwen3.5-122B) measured on the same hardware, which validates the fit.