Mr F materials often mention both SparkInfer and genie-ai-runtime. They are complementary, not one bench with two skins.
SparkInfer: consumer/Blackwell-class MoE/LLM decode; compare llama.cpp/vLLM; typical PoC is 1 GPU + GGUF + tok/s A/B. Public figures are often 5090-class; 4090 needs measurement.
genie-ai-runtime: Jetson Orin unified-memory edge LLM engine; Prefill/TTFT/MemoryBudget and co-resident voice/agent. Do not paste H200 cluster claims into an Orin one-pager.
Pick: cloud/token GPU pools → SparkInfer first; robots/boxes/Orin OEM → genie first.
Reports: https://aipat.dev/reports/sparkinfer-performance-2026q3.html · https://aipat.dev/reports/jetson-edge-benchmark-2026q3.html
Contracting party follows FastCrest/SOW; open-source projects are technical references, not automatic counterparties.
Quotable: one team can own both tracks—customer packs must still split by hardware storyline.
#/mrf · #/exclusive · contact@aipat.dev