Hermes Agent Exposes Ollama Tooling Limits

This looks less like a “best model” question and more like a tool-calling stack problem.

• –Hermes' own docs recommend the official `hermes3` Ollama model for local use and call out 32k context as a practical floor for agent workflows: https://hermes-agent.ai/how-to/use-hermes-with-ollama
• –The docs also stress that Ollama, vLLM, and llama.cpp need the right context length and parser settings, or tool calls degrade into plain text: https://hermes-agent.nousresearch.com/docs/integrations/providers/
• –If GLM Flash is outperforming larger LLaMA/Qwen variants here, the win is probably in tool-call fidelity and prompt adherence, not raw parameter count.
• –For this use case, “best model” really means “best model plus the backend that preserves tools, context, and schema exactly.”