Building a Local LLM · 19/30 · Connect with an Agent
Full series and learning sequence · Practice code ZIP
Hermes can refer to a model family or an agent application. This part covers the Nous Research chat models and their conversation templates. The next part connects Hermes Agent, which is a separate program and does not require choosing a Hermes model.
Prepare an evaluation sheet for 18 pieces with Ollama. Compare the Hermes model’s summary and information extraction results with the same questions.
Base model and post-learning
Hermes 3 Llama 3.1 8B is a post-learning model based on Llama. It is a model that adjusts instruction implementation and conversation behavior to language patterns learned through prior learning. Even if the basis is the same, the answer tendency may vary depending on the post-learning materials and purpose.
SFT is a method of learning through instructions and examples of good responses. There are also methods from the preference optimization family that reflect preferences between answers. Don’t just assume what technique is used by looking at the name; check the model card and technical report for that version.
Run and compare Hermes models
ollama pull hermes3:8b
ollama show hermes3:8b
ollama run hermes3:8bIf memory is insufficient, stop the currently loaded model before loading another. Look for smaller variants in the official tags. Hermes 3 is used here for a comparison under the same conditions. Start with the summary and extraction questions from Part 18 to check whether source facts are preserved and missing information is handled correctly.
Just because it worked well in the English example does not mean that the same result will be obtained in the Korean data. Evaluate whether you like the tone of the speech and whether it gets the necessary facts right. In addition to short questions, we also check whether the initial constraints are maintained in long conversations.
Check conversation templates
A conversation model can be trained with specific token formats that represent the boundaries of roles. If the server uses a different template, problems may arise in distinguishing between system instructions and user input or in calling tools. The Ollama deployment model is first checked by the included templates.
ollama show hermes3:8b --modelfileCheck the TEMPLATE and SYSTEM of the output, but do not arbitrarily overwrite other series templates. When importing GGUF directly, check the template for the model card and experiment with short dialogues, multiple turns, and tool calls. Just because metadata is there doesn’t mean all servers will interpret it the same way.
Weights and program memory
The model does not automatically adjust the weights with each answer. The program can re-insert previous conversations or save memory files. Hermes Agent’s persistent memory and skill creation are also understood separately from these application operations.
Terms of use and redistribution check both the Hermes variant and the underlying Llama. It is not mandatory to link this model to Hermes Agent. If you think of agents and models separately, it is easier to select the necessary tools and models respectively.
reference material
Documentation and model tags checked: October 18, 2026. Primary lab environment: Windows 11 x64.
.jpg){kind=link}