A 9B local model handled photo mapping and approval-gated deletes
In a test running Qwen 3.5 9B through Ollama, the model read EXIF sidecar files for three photos, organized their capture times, cameras, and GPS locations into a table, and placed all three locations on a map. When asked to clean up the files, it used real deletion tools, but every delete operation stopped for human approval because the tool was marked as destructive. The multi-step work and file access happened entirely on the local computer.
The smaller model chose the correct tools but struggled to assemble the final result, sometimes producing fake inline maps or ending without an answer. The 9B model completed the same requests cleanly. The test used Spring AI Playground, an for macOS, Windows, and Linux that with Ollama and can start it automatically.
It can also search an index for the tools it needs instead of loading all 115 tool descriptions at once, and it supports connecting to, inspecting, and testing MCP servers.
Key points
- Qwen 3.5 9B read EXIF data from three photos and produced a table and map.
- Every file deletion paused until a person approved it.
- selected the right tools but did not reliably finish the result.
- Spring AI Playground with Ollama on macOS, Windows, and Linux.
- The app searches for needed tools and can connect to and test MCP servers.