🛰️ AI Fleet — Benchmark Ledger

Ground-truth tok/s straight off each Ollama API (not Mux/OpenWebUI). Last run: 2026-06-19. Click any column header to sort.
EVO always-on
qwen2.5-coder:7b · 9.4 t/s · ~5GB
EVO heavy (on-demand)
qwen2.5-coder:32b · 2.7 t/s
RTX 4090 daily driver
qwen3.6:35b · 41.6 t/s
Vision quick / quality
moondream (EVO) / llama3.2-vision (4090)

📊 Coding / General models

ModelSizeUsageDeviceQuantDateMUX WorkspaceInterfacesGen t/sPrefill t/sLoad sFootprintNotes
qwen2.5-coder:7b7BCoding / agenticEVO ArcQ4_K_M2026-06-19evo-qwen25coder-7bCLI9.4223 @35582.3~5GBFastest EVO coder. Always-on pick — leaves ~55GB free.
qwen2.5-coder:14b14BCoding / agenticEVO ArcQ4_K_M2026-06-19evo-qwen25coder-14bCLIMux5.5112 @355810.1~9GBBalanced quality/speed. num_ctx 16K.
qwen2.5-coder:32b32BCoding / heavyEVO ArcQ4_K_M2026-06-19evo-qwen25coder-32bCLI2.753 @355823.7~19GBBest EVO quality, slow. On-demand only.
qwen3:14b14BCoding (thinking)EVO ArcQ4_K_M2026-06-19evo-qwen3-14bCLIMux5.4111 @35458.8~9GBnothink template; repeat_penalty 1.05.
gemma3:12b12BGeneral + VisionEVO ArcQ4_K_M2026-06-19evo-gemma3-12bCLI5.1119 @354010.6~8GBMultimodal — also a vision model (free bonus).
qwen3.6:35b35B A3BCoding / agenticRTX 4090Q42026-06-19rtx-qwen36-35bCLI41.61620 @353953.9~20GB VRAMMux rtx4090 default. ~3B active/token.
qwen3-coder:30b30B A3BCoding / agenticRTX 4090Q42026-06-19rtx-qwen3coder-30bCLI182 ⚡7436 @353738.5~18GB VRAMA3B MoE blazes (3B active). 182 is short-run burst — sustained ~27.
llama3.2:3b3BSmall / edgeJetsonQ4_K_M2026-06-19jet-llama32-3bCLI3.8280 @355414.7~2GBTier-0 edge.
qwen2.5-coder:7b7BCoding / edgeJetsonQ4_K_M2026-06-19jet-qwen25coder-7bCLI❌ failed~5GBRemoteDisconnected — likely OOM/Frigate contention on the 8GB box.

🖼️ Vision / Image models

ModelSizeUsageDeviceQuantDateMUX WorkspaceInterfacesGen t/sFootprintNotes
moondream1.8BVision (fast)EVO ArcQ42026-06-19evo-moondream-2bpulling…pending~1.7GBTiny, quick image Q&A.
llava:7b7BVision (general)EVO ArcQ42026-06-19evo-llava-7bpulling…pending~4.7GBStandard vision-language.
gemma3:12b12BVision + generalEVO ArcQ4_K_M2026-06-19evo-gemma3-12bCLI5.1~8GBAlready pulled — multimodal.
llama3.2-vision:11b11BVision (quality)RTX 4090Q42026-06-19rtx-llama32vision-11bpulling…pending~7.8GBStrong image reasoning.
llava:13b13BVision (general)RTX 4090Q42026-06-19rtx-llava-13bpulling…pending~8GBLarger LLaVA.