When choosing a model for Hermes Agent, don't just look at the list score. The most important thing is three things: whether the tool call is stable, whether the context is long enough, and whether your task is to write code, research, or automate daily tasks. Models are smart but don't stabilize tools, and they can be difficult to use in agent scenarios.
The master model prioritizes tool capabilities
The value of Hermes lies in its ability to call tools such as terminals, files, web pages, MCPs, messaging platforms, etc. The master model must be able to correctly output tool calls and continue to reason after the tool results come back. Many cases of "it just chats and doesn't work" end up being related to model compatibility.
Long tasks also depend on the context
If you often ask Hermes to read projects, handle long sessions, or do multi-step research, the model context should not be too small. The official configuration also reminds that the summary model used for compression should have a context window no lower than the main model, otherwise the compression call may fail or even cause the intermediate context to be lost.
Practical selection advice
- Code fix: Select a tool to call a stable, long-contexted model.
- Everyday Q&A and light tasks: Cheaper Mini or Flash models are available.
- Local model: Make sure OpenAI is compatible with the interface and tool call format, not just chat.
- Auxiliary tasks such as compression, title, and vision: Configure the auxiliary model separately, and don't default to the most expensive main model.
The most reliable way to test is to let Hermes do a small real task: read a file, change a line, run a command, and summarize the results. If you can complete this chain, consider setting it up as a daily master model.
