Hermes

Hermes is a family of open-weight models from Nous Research, built by fine-tuning open bases like Llama. Open-weight means you can download it and run it on your own hardware, no permission needed, which is the pitch for privacy and no per-token bill. It's tuned to follow instructions closely and drive tools well, so people reach for it in agent setups.

The distinctive thing is what Nous calls neutral alignment. Instead of shipping with a fixed set of built-in refusals, Hermes follows your system prompt and takes on whatever role you give it. That makes it flexible to build on, and it puts more of the responsibility for behavior on you. The weights and the technical reports are public on Hugging Face.

An open-weight model like Hermes makes sense when you need to run things yourself, for privacy or for cost at scale. Its instruction-following and tool use make it a common pick for agents, where you want the model to do exactly what the system prompt says.

It's not the move if you just want the strongest general answer with zero setup, since the big closed models still tend to lead there and need no infrastructure. And neutral alignment cuts both ways: you get flexibility, but the guardrails are yours to add, so think about that before you point it at anything sensitive.