Hermes is a family of open-weight instruction-tuned LLMs developed by Nous Research. Rather than training models from scratch, Nous fine-tunes strong base models (primarily Meta's Llama family) to create assistants that excel at instruction following, coding, tool use, and role-playing.
Why Hermes is popular
Hermes models are known for:
-
Excellent instruction following
-
Strong code generation
-
Reliable function/tool calling
-
Structured JSON output
-
Long multi-turn conversations
-
Good role-playing and creative writing
-
High steerability through system prompts
For developers building AI agents, Hermes has long been considered one of the better open models because it tends to follow system prompts faithfully and produces clean structured output.
Available sizes
Hermes 3 is available in several sizes:
| Model | Typical Use |
|---|
| 8B | Local PCs, coding assistants, chat |
| 70B | High-quality local/server inference |
| 405B | Enterprise/HPC deployments |
Running locally
Hermes is supported by nearly every major inference engine:
-
llama.cpp
-
Ollama
-
LM Studio
-
vLLM
-
Text Generation WebUI
-
KoboldCpp
Since you're already running llama.cpp, Whisper, Intel Arc SYCL, and hot-swappable GGUF models, Hermes 3 8B GGUF should integrate easily into your existing setup.
Performance
For coding:
-
★★★★☆ C++
-
★★★★☆ PHP
-
★★★★☆ JavaScript
-
★★★★★ JSON generation
-
★★★★★ Tool calling
-
★★★★☆ General reasoning
It isn't as strong as the newest frontier reasoning models, but it remains a very capable local assistant with lower hardware requirements.
Example (llama.cpp)
Should you use it?
Given your Gentoo workstation (Intel Core Ultra 9 + Intel Arc GPU) and your focus on:
-
C++ development
-
PHP/Smarty
-
Bootstrap applications
-
Local AI
-
Speech-to-text integration
Hermes 3 8B is a solid choice, especially if you want:
-
better instruction following than many base Llama models,
-
excellent JSON/function-calling support,
-
good coding performance,
-
efficient local inference.
If your primary goal is coding, however, there are newer models that generally outperform Hermes:
-
Qwen3 Coder (best overall for code)
-
DeepSeek Coder V2
-
Hermes 3 8B
-
Gemma 3
For your development workflow, I'd lean toward Qwen3 Coder as the primary coding model and keep Hermes 3 available as a versatile general assistant and agent model.