Podman/docker works fine and doesn’t need the entire gpu like a vm. Hermes has a supported method for just that.

Yeah, it’s all the rage now. Qwen3.6 35B or Google Gemma 26B should work fine for the task. Llmfan on hugging face uses the heretic framework to “abliterate” them and remove any safe guards that might prevent working with “pirated” content.

You can run hermes in a container or vm if you’re worried about the ai hallucinating, though I haven’t seen that happen. Use as high of a Q quant value as you can and run llama.cpp for speed. Or just try a free cloud model with hermes and see if it works.

The agent installed a bunch of mp3 scanning tools, did an inventory of my library and generated a list of actions for me to approve before it ran. Feels like the future.

I actually made a copy of my library (for safety) and had hermes agent scan it and fix it so that it would register well with navidrome and be named coherently. It was very impressive what local AI could do. If you have a gpu with 12+ GB VRAM, you can do it local with an uncensored model. Cloud providers can probably do it too, but might throw safety warning.