Comparison

Which model behind the agent

Which model sits behind the agent, and when small or local models are enough.

Type
Comparison
Confidence
High
Sources
7 sources
First published
Updated

Posters mostly say to pay for a hosted model rather than run local ones, with one qualified exception. A commenter running Hermes on mini PCs says skip local models and pay about $20 a month for a hosted subscription. A developer running small local models says small local models cannot replace state-of-the-art models for agentic coding beyond small refactorings, but that ministral-3:8b or qwen3.5:9b is a boon for many smaller use cases, for example internal tools exposed over MCP. @coldemailchris’s team prefers Codex to Claude Code as cheaper and steadier across thousands of rows. Loïc Berthelot routes everything through OpenRouter so any provider can be plugged in. In client work, an SME automation consultant uses the client’s own ChatGPT account, and Marc Bara runs jobs on existing subscriptions while choosing carefully which model does which job. A founder who replaced their helpdesk says SOTA models are fairly tight on not saving user data. No benchmarks are given; all views are opinion or self-report.

Key points

  • Use a hosted model subscription; local models are more headache than they are worth — 2 posts. (a commenter running Hermes on mini PCs · My Hermes setup (Notion knowledge base)… ¶ L270) (a developer running small local models · MCP on internal business tools with small… ¶ L5)
  • Small local models still suit narrow, everyday tasks — 1 post. (a developer running small local models · MCP on internal business tools with small… ¶ L6)
  • Match the model to the job and route via subscriptions or a router — 3 posts. (Marc Bara · Marc Bara: 15 days with Hermes Agent, 50+… ¶ L94) (Loïc Berthelot · Hermes God-Mode Instruction Manual… ¶ L62) (a Reddit commenter · Anyone else deploying Hermes into SME… ¶ L374)
  • Codex preferred over Claude Code by one team for bulk work on cost — 1 post. (@coldemailchris · Our team built an AI agent that builds our… ¶ L11)

Where posters disagree

Are local models worth running? ⚠️

  • No: pay $20 a month for a hosted model and skip the headache (a commenter running Hermes on mini PCs · My Hermes setup (Notion knowledge base)… ¶ L270)
  • Yes for small, well-scoped uses, though not for agentic coding (a developer running small local models · MCP on internal business tools with small… ¶ L6)

Tactics

  • Route all models through OpenRouter so one setup can use OpenAI, Anthropic or others. — Plug Hermes into OpenRouter and connect whichever provider you want. (Loïc Berthelot · Hermes God-Mode Instruction Manual… ¶ L62)
  • Run jobs and sessions on models through existing subscriptions to limit API cost. — Choose carefully which model does which job. (Marc Bara · Marc Bara: 15 days with Hermes Agent, 50+… ¶ L94)
  • Expose internal tools through MCP so a small local model can handle everyday queries over chat. (a developer running small local models · MCP on internal business tools with small… ¶ L4)

Stories & examples

  • Italian local AI solution on Hermes: a Reddit commenter says they are in a similar situation in Italy, built a local AI solution based on Hermes and other things, and are talking to companies now. (a Reddit commenter · I made €2,700 last month installing Hermes… ¶ L550)

Caveats & warnings

  • The comment says local models cannot replace SOTA models for agentic coding beyond small refactorings, and the thread it replies to is not in the source, so what ‘them’ refers to is implied. (a developer running small local models · MCP on internal business tools with small… ¶ L2)
  • One has to be careful about which model does what, because of cost. (Marc Bara · Marc Bara: 15 days with Hermes Agent, 50+… ¶ L95)

How much to trust this

  • Evidence: 8 posts, 9 independent speakers (repeat posters counted once).

Sources

  1. (01)

    Anyone else deploying Hermes into SME businesses?

    Post

  2. (02)

    Hermes God-Mode Instruction Manual (Dropmagic operator)

    Post

  3. (03)

    MCP on internal business tools with small local models

    Post

  4. (04)

    Our team built an AI agent that builds our entire TAM (Slack)

    Post

  5. (05)

    My Hermes setup (Notion knowledge base), crons and use cases: attorney-operator on a Raspberry Pi

    Post

  6. (06)

    Marc Bara: 15 days with Hermes Agent, 50+ capabilities

    Post

  7. (07)

    I made €2,700 last month installing Hermes Agent for French companies (post removed; comments remain)

    Post

(By email)

Compounding Insights