Author: Umi4Life / Zekamashi(zekamashi@umi4.life)

I started with a simple question: if Hermes is already running GPT-5.5 as the orchestrator, does it make sense to hand the typing to a separate worker drone?

Not a magical “AI saves all tokens” kind of question. A boring operational one.

If GPT-5.5 can plan and review while another agent writes code, maybe Hermes can stretch its useful work across more tasks. Or maybe the handoff overhead eats the whole benefit. The only way to find out was to measure it.

Published on 2026-06-12 at Umi4Life's Blog, last modified on 2026-06-12

Author: Umi4Life / Zekamashi(zekamashi@umi4.life)

Can a local multimodal model reduce GPT/Codex vision usage without making the assistant unreliable?

Summary

I tested whether a local Gemma multimodal model, served through LiteLLM, could act as a first-pass image interpretation helper for my Hermes/Sky Feather assistant.

The goal was not to replace GPT-5.5/Codex vision entirely. The goal was more practical:

Can Gemma handle routine image descriptions, while escalating uncertain, text-heavy, or high-risk images back to GPT?

---

Published on 2026-06-05 at Umi4Life's Blog, last modified on 2026-06-05

]]>automationaihomelab