Open Source 4B VLM Beats Qwen 122B in GeoGuesser Benchmarks

Reinforcement learning was used to train a small scale vision model to identify locations on digital maps. The 4 billion parameter VLM beat Qwen 3.5 122B and G…