gpt-oss-20b: Local Hardware and Deployment Requirements
Start with the official artifact and a runtime that supports its format. The publisher's memory example is a useful starting point, but the production requirement also depends on context, execution settings and concurrent requests.
Model reference
OpenAI describes gpt-oss-20b as an open-weight reasoning model with 21 billion total parameters and 3.6 billion active parameters. The official card lists Apache 2.0 and an MXFP4 execution path; it says the model can run within 16 GB of memory under its supported setup. It also requires the Harmony conversation format.
Treat the memory example as a starting point
Do not turn the publisher's example into a guarantee for every 16 GB device. Hardware and runtime support, weight representation, cache policy and request lengths affect the actual allocation. A different artifact or a fallback implementation can change both memory and speed.
A small local deployment should define the maximum prompt and answer length and test them together. If several people will use it, add concurrent requests to the acceptance test before selecting the final hardware.
Choose the runtime before ordering
Use the model's current instructions and record the model revision, runtime, framework, driver and artifact. Confirm how the runtime handles the official format on the intended accelerator. On another platform, inspect the conversion and backend support rather than assuming the same file runs identically.
Keep formatting and tool interfaces correct. Successful weight loading does not establish correct conversation behavior or working tool calls. Run a small application-level test as well as a hardware load test.
Hardware paths to evaluate
| Route | Validation question |
|---|---|
| Discrete-GPU workstation | Is the supported artifact runnable with the required request profile and headroom? |
| DGX Spark | Does the selected runtime support this model and its required kernels on this platform? |
| Apple silicon system | Is a suitable maintained artifact and execution path available for the exact requirements? |
| Shared GPU server | Does simultaneous usage meet latency and memory targets? |
A useful acceptance workload
Test representative requests, a near-maximum prompt and the expected parallel load. Record output quality on your task, time to first token, generation rate, peak memory and failures. Include the surrounding application because document retrieval and tool calls can dominate the perceived response time.
Share your intended application, request lengths and number of users to compare hardware. No measured EZ Mode compatibility result is asserted on this page.
Sources and documentation
- gpt-oss-20b model card ↗OpenAI · checked 17 Nov 2026
Model overview, MXFP4, memory statement, Harmony and license
- gpt-oss reference repository ↗OpenAI · checked 17 Nov 2026
Reference implementation, backend and format information