3º. Entity extraction with a 2B model: benchmarks from a personal knowledge graph

The article benchmarks qwen3-vl:2b-instruct-q4_K_M, a 2B-parameter quantized multimodal model running locally via Ollama, for entity extraction in a personal knowledge graph. Across 15 text cases and 10 vision cases, it achieved an overall text F1 of 0.645 with zero parse errors, excelling at person-name extraction while struggling more with projects, locations, and topics. The model runs in under 2GB RAM on CPU with 2–4 second latency, offering a practical alternative to large cloud models for on-device use.