Model Selection Matrix (Local-First Practical Policy)

Status: Active Owner: Burt Prime Updated: 2026-02-26

Policy Goal

Use local models as much as practical while preserving execution quality and speed for high-impact tasks.

API Gate (Active)

Budget-conservation mode is active.

  • API usage requires explicit approval token: APPROVE API [task-id]
  • Non-critical tasks are local-only.
  • Emergency reserve policy: 8% emergency + 8% planned critical ops.
  • See: generated/API_GATE_POLICY_ACTIVE.md

Current Constraints

  • Kimi tokens are constrained this week → do not use by default.
  • Use API models only when local models are likely to fail task quality/speed requirements.

Routing Rules

Tier 1 — Local First (default)

Use local by default for:

  • Drafting, summarization, brainstorming
  • Routine analysis
  • Low-risk formatting/transforms
  • Background tasks

Preferred:

  1. Qwen3-14B (local/qwen-14b alias → qwen3:14b) — default local generalist
  2. Mistral (local/mistral-small) — local fallback / second-opinion style
  3. Qwen-Coder (local/qwen-coder-32b) — coding-heavy local, off-peak preferred
  4. Qwen2.5-14B (local/qwen2.5-14b) — legacy fallback only

Tier 2 — API Escalation (when needed)

Escalate from local when task is high-stakes or local quality is insufficient:

  • Critical code changes, debugging blockers
  • External-facing final copy with high consequences
  • Complex reasoning requiring stronger model consistency

Preferred:

  1. Codex (Codex) — technical execution + debugging
  2. GPT-4o (GPT-4o) — polished writing/comms
  3. Gemini (Gemini) — broad research/synthesis

Tier 3 — Constrained/Explicit-Only

  • Kimi (Kimi) — use only if explicitly requested until recharge

Safety Overrides

Always require explicit human confirmation for:

  • Financial commitments
  • Irreversible external operations
  • Public statements representing organization policy

Sub-Agent Model Defaults

  • artifact architect, coding, automation, debug, infra → Codex
  • research, scan, summarize → Qwen3-14B first, escalate to Gemini if weak
  • email, inbox, simple tasks → Qwen3-14B (local) by default
  • draft, announcement → Qwen3-14B first, escalate to GPT-4o for final polish
  • bulk/background → Qwen3-14B / Mistral

Escalation Trigger Checklist

Escalate from local if any true:

  • Output quality fails once after revision prompt
  • Task requires deep repo-wide bug root cause under time pressure
  • User requests final production-grade polish immediately
  • Context pressure exceeds practical local performance

Operator Note

When spawning sub-agents, pass model explicitly to enforce this matrix.

Empirical Local Profiles (2026-02-26)

Benchmarked local Ollama models for speed + quality taste checks:

  • qwen3:14b (via local/qwen-14b): new default; ~2x faster average vs qwen2.5 in current 3-case benchmark set
  • qwen2.5:14b (via local/qwen2.5-14b): retained as legacy fallback route
  • mistral-small3.2:24b-instruct-2506-q4_K_M: style-diverse fallback / second-opinion lane
  • qwen2.5-coder:32b-instruct-q3_K_L: coding specialist, slower but useful for hard code tasks

Artifacts:

  • /tmp/qwen_benchmark_results_2026-02-26.json
  • generated/subagent_assignment_matrix_2026-02-23.md (superseded guidance in progress)

Smart spawn helper

Use:

~/.openclaw/workspace/scripts/spawn_smart.sh "<task description>" [low|normal|high] [run|session]

Examples:

~/.openclaw/workspace/scripts/spawn_smart.sh "debug stripe api sync" high run
~/.openclaw/workspace/scripts/spawn_smart.sh "summarize inbox for today" normal session