Models › gpt-oss 20B (MoE)

Can I run gpt-oss 20B (MoE) in Australia?

Yes. The cheapest hardware we track that holds gpt-oss 20B (MoE) at MXFP4 with an 8k context is the AMD Radeon RX 9060 XT 16GB at A$699 (Mwave, checked 20 Oct 2026, 7:38 am AEST). Expect roughly 134 tokens/sec (estimate).

OpenAI · released 4 Aug 2025 older · superseded but still widely used · 5.9M HF downloads (last 30 days)

GGUF builds: unsloth · bartowski · ggml-org · lmstudio-community

Memory needed by quantisation (8k context)

All hardware that fits at MXFP4

The tok/s figures are bandwidth-based estimates for this model. Where a device has a measured benchmark, the estimate is calibrated to that run. They are not measurements. Methodology.

Graphics card + system RAM (MoE experts in RAM)

Mixture-of-experts models only read a few experts per token, so llama.cpp can keep the expert weights in system RAM and everything else on the GPU (e.g. --n-cpu-moe / -ot exps=CPU). That lets a 12–32GB card run a much bigger model. All figures below are estimates (MXFP4, 8k context, dual-channel DDR5-5600). How we estimate.

DDR5 kit prices today: 32GB: A$610 (Mwave, checked 20 Oct 2026, 7:44 am AEST) · 64GB: A$1,499 (Mwave, checked 20 Oct 2026, 7:44 am AEST) · 128GB: A$2,511 (Mwave, not in stock online, checked 20 Oct 2026, 7:44 am AEST). Card + kit only; the rest of the PC isn't included. RAM-offload speeds are estimates and can be optimistic, so check community results below.

Community measured

No reader results yet. Be the first to submit one →

Try other quantisations and context lengths in the Can I run it? tool → · Submit your own gpt-oss 20B (MoE) benchmark →