GeoVista open-source agentic geolocation (the-decoder.com)

Researchers from China have unveiled GeoVista, an innovative open-source AI model designed for geolocating images through a combination of visual analysis and live web searches. Developed by Tencent alongside various Chinese universities, GeoVista is engineered to rival commercial offerings like Gemini 2.5 Flash. Its standout features include a zoom function for examining specific image details, and a search tool that scours platforms such as Tripadvisor and Instagram for relevant signals, making real-time web search integration a significant advantage over existing models focused solely on image analysis. The researchers structured GeoVista's learning process into two phases: an initial supervised phase using 2,000 curated examples to instill basic reasoning and tool usage, followed by reinforcement learning with 12,000 examples aimed at refining geographic accuracy. The results are promising, with GeoVista achieving 92.64% accuracy at the country level and outperforming many existing open-source models. With the introduction of GeoBench—a new dataset providing a challenging evaluation framework—the study not only raises the bar for geolocation accuracy but also highlights the capabilities of open-source tech in a field often dominated by proprietary advancements. As AI models increasingly decode location from public images, the implications for privacy and data usage become paramount.