Profile Model Fleet

Profile Model Fleet — Three-Provider Assignments

The fleet uses three providers:

• Xiaomi MiMo (provider: xiaomi, base_url: https://token-plan-sgp.xiaomimimo.com/v1) — primary heavy lifter, 1B token giveaway
• DeepSeek Direct (provider: deepseek, base_url: https://api.deepseek.com) — premium reasoning for research/market analysis
• Nous Portal (provider: nous, base_url: https://inference-api.nousresearch.com/v1) — free tier for lightweight workers (deepseek-v4-flash:free)
• OpenRouter (provider: openrouter) — used by secretary for owl-alpha

Profile → Model Map (as of 2026-05-27)

Senna (default chat session)

Oracle (market intelligence)

NOTE (2026-05-27): Researcher was briefly switched to xiaomi/mimo-v2.5-pro due to missing DEEPSEEK_API_KEY in its profile .env. Key was copied from root .env, researcher switched back to deepseek-v4-pro with reasoning_effort: high.

Tier 1 — Lightweight (Xiaomi + OpenRouter, reasoning: none)

Fast, cheap workers for routing, triage, light tasks.

Deleted (2026-05-27): council, explorer, librarian — generic boilerplate, no custom SOUL, no Discord bots.

Tier 2 — Paid Heavy Lifters (Xiaomi MiMo v2.5 Pro)

Tier 3 — DeepSeek Direct (deepseek-v4-pro)

Premium reasoning via DeepSeek API. Best quality for research, analysis, and code generation.

Model Selection Guide

When choosing a model for a new profile, match to the work type:

Rule of thumb: If the profile generates code (even HTML/CSS for UI), use deepseek-v4-pro. If it routes or triages, use mimo. If it writes prose, use owl-alpha.

Config templates

Xiaomi:

model:
  provider: xiaomi
  default: xiaomi/mimo-v2.5-pro
  base_url: https://token-plan-sgp.xiaomimimo.com/v1

DeepSeek Direct:

model:
  provider: deepseek
  default: deepseek-v4-pro
  base_url: https://api.deepseek.com

DeepSeek via OpenRouter (no DEEPSEEK_API_KEY needed):

model:
  provider: openrouter
  default: deepseek/deepseek-chat-v3-0324

Nous free tier:

model:
  provider: nous
  default: deepseek/deepseek-v4-flash:free
  base_url: https://inference-api.nousresearch.com/v1

Reasoning Effort Map

Notes

• Each profile needs its own API keys in its own .env — profiles do NOT inherit from root ~/.hermes/.env. Root .env is only read by processes that resolve to the default profile (senna). All other profiles must have the relevant key in ~/.hermes/profiles/<name>/.env.
• OpenRouter as aggregator fallback: If a profile's direct provider key is missing (e.g. DEEPSEEK_API_KEY), switch to OpenRouter with provider/model format instead of adding a new key. Designer uses this pattern. Verify current model IDs at openrouter.ai/models before committing.
• Xiaomi credentials: XIAOMI_API_KEY in each profile's .env. Check with grep "XIAOMI_API_KEY" ~/.hermes/profiles/<name>/.env.
• DeepSeek credentials: DEEPSEEK_API_KEY in each profile's .env.
• OpenRouter credentials: OPENROUTER_API_KEY in each profile's .env.
• Model name format: Always use provider/model format (e.g. xiaomi/mimo-v2.5-pro, deepseek/deepseek-v4-pro). Bare model names like mimo-v2.5-pro may work but the provider prefix is the canonical form.
• The :free suffix on deepseek v4 flash maps to Nous Portal's free-tier endpoint.
• User explicitly avoids deepseek-r1-0528.
• Senna stays on xiaomi/mimo — user has 1B MiMo tokens, reasonable for session profile.
• Oracle and Researcher use DeepSeek v4 Pro directly for best reasoning quality.
• Always verify with actual config: grep -A5 "^model:" ~/.hermes/profiles/<name>/config.yaml
• See references/profile-key-inheritance.md for the full explanation of why profiles don't inherit root .env keys.