Small Model Tests on a 4-Core CPU Overturn Bigger-Model Assumptions for Camera Agent
The maintainer of OpenNVR, an open source video recorder with a built-in camera agent, benchmarked every small model the agent relies on using a 4-vCPU machine without a GPU, since many users lack dedicated graphics hardware. Four default models were replaced after tests on language, vision, speech-to-text, text-to-speech and image search tasks, with all scripts published in the repository. In the language tests, the larger qwen3:4b performed worst because it rarely called tools and decided without checking, while granite4:micro passed far more cases. A vision comparison found Qwen3-VL the most reliable at avoiding false positives but too slow on CPU, so Moondream was kept for CPU setups. Parakeet TDT was chosen for speech recognition, cutting word error rate and processing time compared with Whisper models. Kokoro was rejected for text-to-speech because it ran slower than real time on four cores, while Piper stayed fast enough.
