Models are not isolated from each other

[imkira]

LocalAI version: 3.12.1

Environment, CPU architecture, OS, and Version: Linux

Describe the bug

I've noticed that because Qwen3.5-4B-GGUF and Qwen3.5-9B-GGUF share the same file mmproj/mmproj-F32.gguf but the files are different files, they get overwritten by LocalAI when you install them and so only the last file remains.

To Reproduce

1. install unsloth/Qwen3.5-9B-GGUF
2. Make sure 9B has mmproj: llama-cpp/mmproj/mmproj-F32.gguf
3. Initiate a conversation with 9B
4. Conversation works
5. install unsloth/Qwen3.5-4B-GGUF
6. Initiate a conversation with 9B
7. Conversation fails

Expected behavior

Conversation works.

Logs

Additional context