OpenClaw Field Guide

Section 6: Selecting Models - Daily Use vs. Coding Tasks

No single model is best at everything. Match the model to the task and you'll get better speed, quality, and cost control.

Four Model Categories

Category Example Models Best For
Fast & efficient qwen2.5-7b, gpt-4.1-mini Daily chat, reminders, quick Q&A
Smart & capable claude-sonnet-4.5, gpt-4.1 Complex reasoning, writing
Coding specialists deepseek-coder-v2, codestral Code generation, debugging
Vision/image analysis gpt-4.1, llava Image descriptions, diagrams

Default vs. Task-Specific Overrides

OpenClaw uses one default model for most work, but you can override by task. For example:

  • Use claude-sonnet-4.5 for drafting a long email.
  • Switch to deepseek-coder-v2 for debugging a script.

Failover Chains

If your primary model fails (rate limit, outage, timeout), OpenClaw tries the next model in the chain. This keeps workflows moving without manual intervention.

Cost Awareness

Long prompts on premium models (for example, gpt-4o) can get expensive. Reserve them for high-value tasks, and use lower-cost models for routine work.

Free Downloadable Local Models

If you want to avoid recurring API costs, this is the best place in the guide to discuss free models you can download and run locally. These models usually work through tools like Ollama or other local inference runtimes.

Good beginner categories:

  • Small fast models such as Gemma, Qwen, or small Llama variants for daily chat and utility tasks
  • Coding-focused models such as DeepSeek Coder or Codestral-family local options for programming help
  • Vision-capable local models such as LLaVA-style models if you want basic image understanding on your own machine

Main tradeoff: downloadable models are free to obtain, but they shift the cost to your hardware. A lightweight laptop can run small models, while larger models often need a stronger desktop or GPU.

Rule of thumb: if you want the easiest start, begin with free cloud models. If you want privacy, offline use, or predictable long-term cost, add downloadable local models next. Check provider docs occasionally, because model lineups change quickly.
Per-agent routing: You can assign different models to different OpenClaw agents. Example: Agent A uses claude-sonnet-4.5 for emails, while Agent B uses deepseek-coder-v2 for code reviews.

Starter Config Strategy

  1. Pick one default model (for example, qwen2.5-7b for daily chat).
  2. Add one fallback model (for example, gpt-4.1-mini).
  3. Add task overrides for specialized work.

Example config snippet:

⚙️ Reference only — do not paste this into any file:

{
  "default_model": "qwen2.5-7b",
  "fallback_models": ["gpt-4.1-mini"],
  "task_overrides": {
    "coding": "deepseek-coder-v2",
    "writing": "claude-sonnet-4.5"
  }
}

::: action Run openclaw onboard to set up your first model, then tune config choices as you learn what works best. :::