⚡ OMP Model Quickswitch

Current: nemotron (Nemotron-3-Ultra + Omni-Vision, Owl-Alpha fallback on all roles)
Updated: 2026-06-23 | SSOT: ops/omp-model-quickswitch.md

🎯 Config at a Glance

# ~/.omp/agent/config.yml  — modelRoles (validated selectors from `omp models`)
modelRoles:
  default: nvidia/nemotron-3-ultra-550b-a55b:free
  smol:    nvidia/nemotron-3-ultra-550b-a55b:free
  plan:    nvidia/nemotron-3-ultra-550b-a55b:free
  advisor: nvidia/nemotron-3-ultra-550b-a55b:free
  task:    nvidia/nemotron-3-ultra-550b-a55b:free
  commit:  nvidia/nemotron-3-ultra-550b-a55b:free
  vision:  nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free
  designer: nvidia/nemotron-3-ultra-550b-a55b:free
  slow:    nvidia/nemotron-3-ultra-550b-a55b:free

# ~/.omp/agent/models.yml  — OpenRouter provider override (8 lines, zero redundancy)
providers:
  openrouter:
    compat:
      extraBody:
        models: [openrouter/owl-alpha]   # server-side fallback: rate-limit, downtime, moderation, context

🔄 Swap Profiles Instantly

swompswap              # show current + available
swompswap nemotron     # 🏆 nemotron-ultra (text) + omni (vision) + owl fallback
swompswap claude       # 📜 original claude-sonnet-4-6 / opus-4-8
swompswap free         # 🆓 owl-alpha everywhere (text-only, 1M ctx)

~/bin/swompswap — PyYAML, timestamps backup before write, zero deps beyond python3+pyyaml.

🧠 Model Specs (Free Tier)

Model	Ctx	Max Out	Modalities	Tools	Thinking
`nvidia/nemotron-3-ultra-550b-a55b:free`	1M	65K	text	✅	✅ (auto)
`nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free`	256K	65K	text+img+vid+aud	✅	✅ (auto)
`openrouter/owl-alpha`	1.05M	262K	text only	✅ (no tool_choice)	✅

⚙️ Fallback Mechanism Comparison

Mechanism	Layer	Triggers On	Config	Current
`compat.extraBody.models`	OpenRouter server	any primary failure	`models.yml` provider `compat`	✅ active
`retry.fallbackChains`	OMP client	429/5xx/network only (≠ context)	`config.yml` `retry:` block	❌ unused
`contextPromotionTarget`	model-level	`context_length_exceeded` ONLY (larger ctx sibling)	`models.yml` `modelOverrides`	❌ wrong tool

Pick one. extraBody wins: covers more errors, zero per-role verbosity, transparent.

✅ Pre-Flight Validated

Check	Status
`models.yml` schema	✅ clean — no registry errors in log
OpenRouter API key	✅ stored in `agent.db` (cred id=3, `api_key`)
Selectors match registry	✅ verified via `omp models`
`swompswap` round-trip	✅ nemotron↔claude↔free, detects correctly
Auth on `getAvailable()`	✅ OpenRouter models are available

🚀 Go Live

# Restart OMP session (config loaded at init, not hot-reloaded)
# First turn will exercise: auth + selector + extraBody end-to-end
omp --allow-home

📝 Notes

Vision fallback to owl-alpha silently drops image/video/audio — documented, accepted per user intent.
:free is an OpenRouter route suffix, NOT a thinking level — OMP handles it correctly.
Selectors are flat (e.g. nvidia/nemotron-3-ultra-550b-a55b:free, NOT openrouter/nvidia/...).
OMP flattens OpenRouter namespaces as provider prefix.
Zero redundancy, zero hand-defined models — all three IDs exist in OMP’s bundled OpenRouter catalog.

Generated from live config. Single source of truth for OMP model topology.

moincraft SSOT

Explorer

OMP Model Quickswitch — Nemotron Ultra + Omni + Owl Fallback