Skip to content
archive

Hermes Routing Policy

AI · · 3 min read

Hermes Routing Policy

This is the default model-routing policy for Hermes jobs.

Principles

  1. Use the cheapest model that can do the job well.
  2. Escalate only when the task is ambiguous, high-stakes, long-horizon, or quality-sensitive.
  3. Never route reminders, trivial formatting, or tiny maintenance tasks through frontier models.
  4. Use strong premium models for synthesis, planning, and user-facing writing.
  5. Use explicit fallback chains so rate limits do not break workflows.

Global Defaults

  • Default internal worker: DeepSeek V4 Flash
  • Default reasoning worker: DeepSeek V4 Pro
  • Default premium synthesis: Claude Opus 4.7
  • Default ideation / creative strategy: GPT-5.4
  • Default long-horizon / architecture: GLM-5.1

Exact Model Assignments by Job Type

1) Daily reminders, alerts, status pings

Use when the job is mostly formatting, sending, or checking a single condition.

  • Primary: gpt-5.4-mini
  • Fallback 1: DeepSeek V4 Flash
  • Fallback 2: Qwen3.5 Plus

Examples:

  • daily reminders
  • cron delivery wrappers
  • threshold alerts
  • healthcheck notifications
  • simple summaries with no deep synthesis

2) Knowledge-base ingest, extraction, and cleanup

Use when the job is mostly parsing, deduping, or summarizing source material.

  • Primary: DeepSeek V4 Flash
  • Fallback 1: Qwen3.6 Plus
  • Fallback 2: DeepSeek V4 Pro

Examples:

  • URL ingest
  • article capture
  • transcript cleanup
  • note normalization
  • metadata extraction

3) User research / market research / article synthesis

Use when the job needs synthesis across messy inputs and judgment.

  • Primary: Claude Opus 4.7
  • Fallback 1: Kimi K2.6
  • Fallback 2: DeepSeek V4 Pro

Examples:

  • research briefings
  • competitor analysis
  • product discovery
  • source comparison
  • “what does this mean?” tasks

4) Brainstorming / ideation / naming / product strategy

Use when the job needs breadth, taste, and non-obvious alternatives.

  • Primary: GPT-5.4
  • Fallback 1: Claude Opus 4.7
  • Fallback 2: Qwen3.6 Plus

Examples:

  • feature ideation
  • naming
  • product positioning
  • content hooks
  • strategic options

5) Coding, implementation, and normal agentic workflows

Use when the job is a real build task but not yet a huge refactor.

  • Primary: DeepSeek V4 Pro
  • Fallback 1: Kimi K2.6
  • Fallback 2: DeepSeek V4 Flash

Examples:

  • feature implementation
  • bug fixes across a few files
  • terminal-heavy automation
  • standard agent workflows
  • code review and PR feedback

6) Large refactors, architecture, long autonomous runs

Use when the task spans many files, requires deep planning, or runs for hours.

  • Primary: Kimi K2.6
  • Fallback 1: GLM-5.1
  • Fallback 2: Claude Opus 4.7

Examples:

  • multi-file refactors
  • architecture decisions
  • long-horizon autonomous runs
  • complex orchestration
  • agent-swarm style coordination

7) Debugging and terminal-heavy automation

Use when the task is exploratory and changes direction a lot.

  • Primary: Qwen3.6 Plus
  • Fallback 1: DeepSeek V4 Pro
  • Fallback 2: DeepSeek V4 Flash

Examples:

  • runtime debugging
  • script repairs
  • log analysis
  • environment troubleshooting
  • cron job recovery

8) Multimodal / screenshot-to-code / visual reasoning

Use when the task involves images, UI screenshots, or mixed modalities.

  • Primary: MiMo-V2-Omni
  • Fallback 1: GPT-5.4
  • Fallback 2: Claude Opus 4.7

Examples:

  • screenshot-to-code
  • UI audits
  • visual debugging
  • mixed image + text reasoning

Hermes Job Routing Rules

Internal ops

  • Use DeepSeek V4 Flash by default.
  • Escalate to DeepSeek V4 Pro only when the job spans multiple steps or depends on uncertain interpretation.

User-facing synthesis

  • Use Claude Opus 4.7 or GPT-5.4.
  • Do not use flash-tier models for polished external writing unless it is truly trivial.

Research

  • Start with Claude Opus 4.7.
  • If the task is broad or long-horizon, escalate to Kimi K2.6.
  • If the task is price-sensitive but still analytical, use DeepSeek V4 Pro.

Code work

  • Start with DeepSeek V4 Pro.
  • Escalate to Kimi K2.6 for refactors or complex orchestration.
  • Use DeepSeek V4 Flash for tiny fixes and mechanical edits.

Reminders and simple notifications

  • Keep on gpt-5.4-mini or DeepSeek V4 Flash.
  • Never spend premium reasoning on routine delivery.

Suggested Hermes Defaults

If you want a simple universal setup:

  • Default Hermes model: gpt-5.4-mini
  • Default worker for automation: DeepSeek V4 Flash
  • Default research model: Claude Opus 4.7
  • Default coding model: DeepSeek V4 Pro
  • Default architecture model: Kimi K2.6

Fallback Discipline

Use this fallback order whenever a job fails because of rate limits or weak output:

  1. Primary model
  2. Fallback 1
  3. Fallback 2
  4. Degrade to a smaller model only if the task is mechanical

Do not keep retrying the same model after a 429, 503, or clearly bad-output loop.