Ollama has no --n-cpu-moe switch. Here's what it does with MoE experts instead

The short answer

No, Ollama has no setting like llama.cpp's --n-cpu-moe. The request, issue 11772, has been open since 7 August 2025. But since Ollama 0.30 (13 May 2026), every GGUF model runs inside llama.cpp's own…

Leave num_gpu alone. Any value you set counts as a manual choice, and the expert-aware fit step steps aside.

There is an unofficial route. Ollama passes its own environment to llama-server, so LLAMA_ARG_N_CPU_MOE reaches the engine. Ollama doesn't document it, and it applies to every model.

Want a per-model number? Run llama.cpp directly, or use LM Studio's expert-weights toggle.

We read code, we didn't benchmark. No speed numbers here are ours.

Read this before you set it

The variable is server-wide. It applies the same N to every model that server loads, including dense models (where it does nothing) and small MoE models that would have fit entirely on the GPU…