Server: debian-ai (192.168.1.50, Tailscale 100.126.244.3), Debian 13.
- RTX 2080 Ti 11GB, llama-swap on :9292
- Two llama.cpp builds: sm75 (b9341, MTP-Turbo) for qwen35/qwopus/qwen36 with --n-cpu-moe 128; upstream (ac4cdde) for Gemma 4 QAT with --spec-type draft-mtp
- Hermes gateway :18789, dashboard :9119 (no-auth/--insecure preferred)
- 7 models: qwen36-35b-mtp 7.2GB/42t/s, nex-n2-mini 2.8GB/27t/s, gemma-26b 10.8GB/45t/s (tight), qwopus-9b ~8GB, qwen35-9b-mtp 7.5GB/126t/s, gemma-12b 7.7GB/123t/s, bonsai-8b ~2GB
- --n-cpu-moe fixes MoE OOM (NOT lowering -ngl)
- auto-model-switch.py at ~/.hermes/scripts/
- fallback_providers must match custom provider ID