HighVRAM’s new gpt-oss-20b GPU guide turns OpenAI’s published memory guidance into four practical buying paths. The key distinction is format and compatibility: OpenAI lists about 16GB for the default MXFP4 Transformers path and about 48GB for bfloat16, while stating that the cited MXFP4 path requires Hopper-or-later hardware including RTX 50-series cards.
The four decisions
| Path | Why it belongs | What still needs proof |
|---|---|---|
| 16GB RTX 50-series | Matches the approximate official MXFP4 capacity and cited consumer generation | Usable headroom after display memory, cache and runtime allocations |
| 24GB | More capacity for the complete workload | The exact runtime and format; capacity alone does not establish MXFP4 support |
| RTX 5090 32GB | RTX 50-series path with substantially more room than ~16GB | Measured speed, context, output behavior, power and value |
| 48GB | Large MXFP4 headroom or the approximate documented bfloat16 class | Whether bfloat16 plus cache, buffers and concurrency fits safely |
Three mistakes the guide avoids
- Every 16GB GPU is not equivalent: the official MXFP4 statement includes an architecture and software condition, not only a capacity number.
- 3.6B active is not the stored-weight requirement: gpt-oss-20b has 21B total parameters; active MoE routing does not make the remaining experts disappear from memory.
- Fit is not speed: memory capacity, published bandwidth, time to first token, generation throughput, output quality, power and complete-system cost are separate measurements.
Follow the decision path
- Read the complete gpt-oss-20b GPU guide.
- Check the official memory and model-family boundaries.
- Compare 16GB RTX paths, the RTX 5090 32GB platform, or 48GB deployment paths.
- Use the VRAM calculator as a planning range and validate in the intended runtime.
- Inspect the benchmark database or submit a controlled gpt-oss result.
- Build a private hardware watchlist before checking current offers.
Published October 25, 2026. The ~16GB MXFP4, ~48GB bfloat16 and architecture statements are OpenAI guidance for the cited Transformers path. HighVRAM has not independently benchmarked every configuration. Inclusion is not endorsement or ranking.