GPT-OSS — Open MoE
Open-weight 20B MoE. Both the base and safeguard variants are pre-compiled for NPU with dedicated expert-dispatch xclbins.
Models
| Model | Params | 1BP Size | Backend(s) | Perf |
|---|---|---|---|---|
| GPT-OSS-20B | 20B | — | NPU / CPU | — |
| GPT-OSS-Safeguard-20B | 20B | — | NPU / CPU | — |
Notes
- NPU: 6 xclbins each, including
expert.xclbinfor MoE dispatch. MoE expert batched GEMM generator ingenerators/n1_core_moe_expert.py. - GPU HIP / Vulkan: not yet. CPU: universal GGUF backend.
See also: full model support detail · benchmarks SSOT · all families