Short answer: for OpenAI’s default gpt-oss-20b MXFP4 path, start with a supported 16GB RTX 50-series GPU only if you will validate context and peak memory; choose 24GB or 32GB when you want more working room; and treat 48GB as the documented bfloat16 loading class, not guaranteed spare capacity. OpenAI lists about 16GB for MXFP4 and about 48GB for bfloat16. Capacity alone does not prove that the required kernels, runtime or complete workload will fit.
The most important compatibility check: OpenAI’s Transformers guide says MXFP4 is supported on Hopper-or-later architectures, including H100/GB200 and RTX 50-series cards. Do not assume an older 16GB GPU can use that exact path merely because its capacity label matches.
gpt-oss-20b hardware paths at a glance
| Path | What it buys | Main risk to verify | Good next page |
|---|---|---|---|
| RTX 5060 Ti / 5070 Ti / 5080 16GB | Official MXFP4 target capacity on the cited RTX 50-series generation | Little nominal room above the ~16GB guidance; display use, cache and runtime allocations can still cause OOM | Compare 16GB RTX paths |
| 24GB supported or converted-format path | More room for cache, buffers and the rest of the workload | The exact runtime and format may differ from OpenAI’s cited Transformers path | Plan the complete workload |
| RTX 5090 32GB | RTX 50-series compatibility with substantially more capacity than the ~16GB MXFP4 figure | 32GB does not establish speed, maximum context or correctness | Check the RTX 5090 platform |
| 48GB workstation/server GPU | Large MXFP4 headroom or the approximate documented bfloat16 loading class | ~48GB bfloat16 is a load estimate, not room for every cache and concurrency target | Compare 48GB paths |
| CPU offload or shared memory | An alternative when the exact GPU path does not fit or is unsupported | Different memory behavior and potentially much lower measured speed | Understand offload |
Is 16GB enough for gpt-oss-20b?
It is the official approximate requirement for the cited MXFP4 Transformers path, but it is also a tight boundary. A 16GB label does not mean 16GB is completely available to the model: a display workload, runtime workspaces, KV cache and other processes consume memory too. The practical test is whether the longest representative request completes without an out-of-memory error while peak device memory remains below the usable limit.
Among 16GB consumer cards, the first filter is not a gaming chart. It is whether the precise card and software stack support the required path. The RTX 5060 Ti 16GB, RTX 5070 Ti 16GB and RTX 5080 16GB are all Blackwell-generation candidates; their equal capacity does not imply equal measured throughput, power, price or physical requirements.
Why 24GB or 32GB can be the safer one-GPU target
Moving above the approximate 16GB model guidance gives the cache and runtime more room, but only after compatibility is established. A 24GB card with an unsupported MXFP4 path is not automatically a better fit than a supported 16GB card. If another runtime uses a converted format, record that exact artifact and do not present its memory behavior as OpenAI’s documented MXFP4 result.
An RTX 5090 combines 32GB with the RTX 50-series generation named by OpenAI’s guide. That makes it a straightforward headroom path for this workload, not a universal value winner. Measure tokens per second, time to first token, output quality and peak memory separately.
What 48GB changes
OpenAI lists approximately 48GB for gpt-oss-20b in bfloat16. A 48GB accelerator therefore serves two different decisions: generous capacity for the default MXFP4-sized workload, or a tight documented loading class for bfloat16. Those are not interchangeable claims. If bfloat16 is the goal, cache, buffers and concurrency may require deliberate placement or more than the nominal 48GB label.
Professional and server 48GB cards also differ in active versus passive cooling, display outputs, ECC, power delivery, software generation and host requirements. Buy the complete deployable system, not the memory number alone.
Do active parameters make this a 3.6B memory problem?
No. OpenAI describes gpt-oss-20b as 21B total parameters with 3.6B active parameters. The active number describes MoE routing per token; it does not mean the other experts vanish from stored-weight memory. Use the published runtime requirement or the exact artifact as the starting floor, then add the complete workload.
Compatibility checklist before buying
- Record the exact GPU model, architecture and memory capacity—not only the product family.
- Choose the runtime and exact gpt-oss-20b format: default MXFP4, bfloat16 or a clearly labeled conversion.
- Confirm driver, framework, kernel and Harmony-format support from the runtime’s current documentation.
- Set prompt length, maximum output, cache precision, batch size and concurrency to the intended workload.
- Measure peak device memory and complete one representative long request.
- Measure speed and output behavior separately; memory fit is not a benchmark.
- Check card dimensions, cooling, power supply, connectors, host RAM and storage.
A defensible buying sequence
- Use the gpt-oss memory guide to confirm the official format boundary.
- Compare supported 16GB consumer paths or 48GB deployment paths.
- Use the GPU comparison directory for exact hardware pairs.
- Estimate the complete workload with the VRAM calculator.
- Inspect the benchmark database without treating unlike tests as rankings.
- Submit a controlled result if you can document the format, context, runtime and hardware.
- Build a private hardware watchlist; it stays in the browser unless the visitor explicitly saves it.
Sources and limits
- OpenAI gpt-oss-20b model page
- OpenAI: run gpt-oss with Transformers
- OpenAI: verifying gpt-oss implementations
- OpenAI Harmony response format
- HighVRAM methodology
Reviewed October 25, 2026. The ~16GB MXFP4 and ~48GB bfloat16 figures are OpenAI’s published guidance for the cited Transformers path. HighVRAM has not independently benchmarked every listed configuration. Availability, runtime support and measured performance can change.