Qwen VL on Apple M4: llama-server 3.1 s vs Ollama 7 s per new image
Benchmark on an Apple M4 with 16 GB: 25 robot-camera photos at 640x480, the same 'detect every object' prompt, one uncached call per image. llama-server (llama.cpp b12110) running Qwen2.5-VL-3B Q4_K_M, sending about 400 image tokens: median 3.1 s. Ollama 0.40.13 with qwen2.5vl:3b: median 7.0 s with the same accuracy, because Ollama upscales small images to about 1000 visual tokens. Ollama with qwen3-vl:2b-instruct: 4.2 s, and shrinking the input to 448x336 did not make it faster. Two gotchas. First, qwen3-vl often copies the prompt's description into the label (for example 'separate object lying on the floor or table surface'), so a label filter that drops words like 'floor' or 'table' silently removes correct detections. Second, a benchmark that repeats one image hits Ollama's prompt cache and reports about 0.75 s, so change a pixel per call. Also, homebrew llama-server cannot load Ollama's qwen2.5vl blob ('unknown model architecture qwen25vl'); use the ggml-org GGUF instead.