gpt-oss-120b: GPU Memory and System Selection
The model's parameter count does not directly translate into a single universal memory requirement. Use the intended artifact and execution path, then validate the complete request profile on candidate hardware.
Publisher reference
The official model card describes 117 billion total parameters with 5.1 billion active parameters. OpenAI's MXFP4 path is designed to run on a single 80 GB GPU; its examples name H100 and MI300X. The latter example names a device, not a claim that every example device has exactly 80 GB. The card specifies the Harmony conversation format.
Active parameters do not define the whole weight store
An MoE model activates part of its parameter set during a token's computation, but the complete artifact and expert storage still matter to the deployment. Do not size the system as if it only stored the active parameters.
Account for runtime allocations, sequence-dependent state and simultaneous requests. Memory offload or a different representation can make another installation possible while changing the performance and integration costs.
Evaluate candidate hardware without promising compatibility
H100 is a natural candidate to investigate because it appears in the publisher's example. A100 systems may be attractive commercially, but equivalent capacity does not prove support for the same low-precision kernels or the same memory footprint. Check the runtime's documented path and benchmark it before committing.
Likewise, a compact unified-memory computer needs its own backend and artifact validation. A card capacity, a shared-memory capacity and an application memory requirement are different things.
Define the service you intend to run
- Exact artifact, revision and runtime version.
- Maximum prompt length and output length.
- Number of concurrent active requests and acceptable latency.
- Required tool use, application interfaces and output quality.
- Host RAM, storage, networking and monitoring requirements.
Use a reproducible acceptance result
Record the complete system and model setup beside time to first token, generation rate and peak memory. Include long requests and the desired parallel load. Compare the smaller gpt-oss-20b on the same business task if it could meet the quality requirement at lower cost.
A sourcing enquiry can then request a system with an explicit acceptance target. The publisher's memory example is documented guidance; this page does not claim an EZ Mode-measured result.
Sources and documentation
- gpt-oss-120b model card ↗OpenAI · checked 17 Nov 2026
Model overview and MXFP4; publisher's 80 GB example
- gpt-oss reference repository ↗OpenAI · checked 17 Nov 2026
Reference implementation, backend and format information