local-model-route

name: local-model-route description: > Routes any task to the correct local Ollama model based on task type, context size, and privacy requirements. Prevents GPU overload on 12GB VRAM systems. ug-ug: lite model: sonnet gstack_stage: Think

SKILL: local-model-route

Bot: any Role: Given a task description and constraints, select the correct local Ollama model and emit a ready-to-use routing decision with ollama run or API call pattern. Ug-ug mode: lite Status: beta Parallelizable: yes — no shared mutable state detected (auto-inferred; verify) Model: sonnet Tool compatibility: Claude Code · OpenCode · Aider · n8n HTTP node

When to invoke

Trigger phrases: "which model should I use?", "route this task", "what ollama model for X", "local model for [task]", "offload to ollama"

Invoke before any local Ollama session when the right model isn't obvious.

Routing table

Hard limits on 12GB VRAM:

• Never run 14B + any other gen model simultaneously
• Default context: 4k (Ollama default under 24GB VRAM). Raise to 16k only when OpenCode tool use or PRD work needs it: OLLAMA_CONTEXT_LENGTH=16384 ollama serve
• 30B+ models: not viable locally

Step-by-step execution

Step 1 — Classify the task

task_type:       autocomplete | code-edit | prd-summary | estimate-review | embeddings
context_needed:  small (<4k) | medium (4k–16k) | large (>16k)
privacy:         local-only | cloud-ok
concurrent_jobs: list any other models currently loaded

Step 2 — Apply routing logic

IF privacy == local-only → always route local regardless of task
IF context_needed == large AND task != embeddings → warn: may need cloud exception path
IF concurrent_jobs includes 14b model → do not add another gen model
ELSE → pick from routing table above

Step 3 — Emit decision

MODEL:    qwen3:8b
CONTEXT:  16384  (raise only if needed)
PATTERN:  ollama run qwen3:8b
          OR: curl http://localhost:11434/api/generate -d '{"model":"qwen3:8b","keep_alive":"30m"}'
REASON:   PRD draft — general reasoning, no long context needed

Cloud exception path

Only use when:

• Context clearly exceeds 16k–32k effective local capacity
• Multi-repo architectural reasoning required
• Deliverable has zero margin for error

Never send: raw secrets, broad repo snapshots, PII, client-confidential code. Send: Repomix-compressed slices, redacted excerpts, issue-specific context only.

Key rules

• Autocomplete model stays loaded all day — 3B never competes with main model
• nomic-embed-text stays loaded alongside any gen model (only 274MB)
• deepseek-r1:8b is for review passes only — not interactive sessions
• Never recommend cloud for proprietary client code (PrevenDebt, WebMeet, etc.)

Handoffs

Permissions