Model → hardware
gpt-oss-20b
Plan for 16 GB of dedicated GPU memory at MXFP4 profile and a 4,000-token context. 24 GB gives more headroom under the same settings.
Reasoning and tool-oriented tasks. Memory fit is an estimate, not a speed or quality guarantee.
Hardware starting points
Planning estimatesMXFP4 profile · 32 GB system RAM · one active request · full GPU residency
| Dedicated GPU memory | At 4,000 tokens | At 16,000 tokens | Example from these guides |
|---|---|---|---|
| 16 GB | Smallest listed tier that fits14.8 GB + 0.48 GB reserve | Exceeds this budget15.8 GB + reserve | RTX 4060 Ti 16 GB |
| 24 GB | More headroom14.8 GB + 0.72 GB reserve | Fits the estimate15.8 GB + reserve | RTX 4090 24 GB |
The smallest tier fits the calculation; it is not a verified minimum. Cache precision and runtime buffers can change the fit. “More headroom” means at least 2 GB remains after the estimate and reserve.
The publisher describes operation within 16 GB with native MXFP4. Use a compatible runtime and its Harmony chat template. This is a native MXFP4 planning profile, not an ordinary Q4_K_M or Q5 file. CPU offloading needs adequate system RAM and changes performance; it is not included in these full-GPU estimates.
Compare quantisation and context
| File profile | Weight input¹ | GPU estimate at 4K | GPU estimate at 16K |
|---|---|---|---|
| MXFP4 profile | 13.50 GB | 14.8 GB | 15.8 GB |
¹ Catalogue planning weights. These can differ from an exact downloaded file; runtime memory also includes cache and buffers.
Keep the exact file, runtime version, cache type and context limit consistent when comparing. Generic Q5/Q8 extrapolations are omitted for this native MXFP4 model.
Setup and common questions
Which format and runtime do I need?
Use a runtime with explicit gpt-oss, MXFP4 and Harmony support. MXFP4 is not interchangeable with an ordinary GGUF Q4_K_M conversion. Harmony must handle reasoning and final-answer channels; a generic chat template can give misleading output.
How should I choose context and reasoning effort?
The publisher lists up to 131,072 context tokens. Runtime support and memory are separate limits. Start with the context you need, and keep low, medium or high reasoning effort fixed when comparing time to an answer.
Does the active parameter count reduce the memory needed?
The model has about 20.9B total parameters and 3.6B active per token. The active count describes computation, not the amount of weights that must be stored. It does not establish inference speed.
How are these recommendations chosen?
We compare dedicated GPU memory tiers using the main Finder’s memory calculation at the stated quantisation and context. These tiers describe memory fit, not speed or output quality.
GPU memory includes weights, context cache and a 0.9 GB runtime allowance, plus a 3% reserve (0.15–1.5 GB). Cache format and runtime buffers can change the fit. See the calculation assumptions.
Sources and related guides
Specification sources checked 2 October 2026.
These comparisons are planning estimates. No reviewed or independently reproduced measurements are available yet. About test evidence.