Which model should I use with OpenClaw?

For most agents, start with a frontier Claude model as the default because it handles reasoning, judgment calls, and long conversations most reliably. Then route cheaper or background tasks to a lower-cost model, and reach for Gemini when the work is multimodal. OpenClaw is model-agnostic, so you are choosing per task, not committing to one provider forever.

Is running OpenClaw on a local model actually free?

The API cost is zero because inference runs on your own hardware through Ollama, so there are no per-token charges and your data never leaves the machine. It is not free in the fuller sense: you pay in hardware, electricity, and lower capability than a frontier cloud model. For private or high-volume background tasks that do not need frontier reasoning, that trade is often worth it.

What hardware do I need to run a local model?

8GB of RAM runs small 7B models slowly, 16GB is comfortable for 7B to 13B models, and 32GB or a dedicated GPU is needed for 30B-plus models at acceptable speed. Apple Silicon Macs punch above their weight thanks to unified memory, so an M-series Mac with 16GB is a strong entry point for local inference.

Can I mix cloud and local models in one OpenClaw agent?

Yes, and it is the recommended setup for serious deployments. OpenClaw supports primary and secondary model configurations, so you can run a capable cloud model for hard reasoning and fall back to a local or cheaper model for routine, high-volume steps. That two-tier approach is exactly what cuts monthly cost without hurting quality where it matters.

Does switching models lose my agent's memory or setup?

No. Model-agnostic architecture means the model is a swappable component. Your agent keeps its memory, personality, tools, and connected channels when you change providers, so you can move from GPT to Claude to a local model as the landscape shifts without rebuilding the agent.

Are cheaper models good enough for production?

For a lot of tasks, yes. Lower-cost models, including strong open and Chinese models, handle classification, extraction, summarization, and routine replies well. Reserve the frontier model for the steps that genuinely need judgment. Matching model strength to task difficulty, rather than defaulting everything to the most expensive option, is the core of cost optimization.

How to Choose the Best OpenClaw Model: Cloud vs Local (Complete Guide)

This is Day 3 of the OpenClaw Bootcamp. Yesterday you installed OpenClaw and connected Telegram. Today you will learn which AI model to use for which tasks, compare every major provider OpenClaw supports, and connect a fully local model so your agent can run for $0 on your own hardware.

The full video walkthrough covers everything below with live demos:

Why Model-Agnostic Architecture Matters

One of OpenClaw's most important design decisions is that it is model-agnostic. Your agent is not locked to any single AI provider. You can swap between Claude, GPT, Gemini, local models, or any compatible endpoint, and your agent keeps its memory, personality, tools, and channels intact.

This matters because the model landscape changes fast. A model that is best today might be outperformed tomorrow. Model-agnostic architecture means you are never stuck.

The Model Landscape: What OpenClaw Supports

Anthropic (Claude)

Claude models are the default recommendation for most OpenClaw deployments. They excel at reasoning, following complex instructions, and maintaining coherent long conversations. Claude is particularly strong for agents that need to make judgment calls, support agents, research assistants, and workflow orchestrators.

The tradeoff is price. Claude is premium-tier, which is why Day 4 covers cost optimization in detail.

OpenAI (GPT)

GPT models offer a strong balance of capability and ecosystem support. If you are already in the OpenAI ecosystem or your clients use OpenAI, this is a natural fit. GPT is strong at structured output, function calling, and code generation.

Google (Gemini)

Gemini models bring competitive performance at lower price points. They are particularly interesting for multimodal tasks, Gemini handles images, video, and audio natively. If your agent needs to process visual content, Gemini is worth serious consideration.

Lower-Cost Chinese Models

Models from providers like DeepSeek and others offer surprisingly strong performance at a fraction of the cost of Western providers. For budget-sensitive deployments or tasks that do not need frontier reasoning, these models can cut costs dramatically without sacrificing much quality.

Local Models (Ollama)

This is the zero-cost option. Run a model on your own hardware and your agent operates for free, no API calls, no usage billing, complete data privacy. The tradeoff is that local models are less capable than cloud models, and performance depends on your hardware.

How to Install Ollama and Connect It to OpenClaw

Ollama makes local model deployment dead simple. Install it, pull a model, and point OpenClaw at it:

Install Ollama from the official website. It supports macOS, Linux, and Windows.
Pull a model, start with something small like llama3 or mistral to test your setup.
Point OpenClaw at Ollama through the config file or the interactive setup. OpenClaw detects local Ollama instances automatically.
Test it, send a message through Telegram and confirm your agent responds using the local model.

Hardware Requirements for Local Models

Local model performance depends entirely on your hardware:

8GB RAM, enough for small 7B parameter models. Responses will be slow but functional.
16GB RAM, comfortable for 7B-13B models. Reasonable response times for most tasks.
32GB+ RAM or a dedicated GPU, required for larger models (30B+) with acceptable speed.

Apple Silicon Macs are particularly good for local inference because of their unified memory architecture. An M1 Pro with 16GB can run 7B models comfortably.

Cloud vs Local: The Real Tradeoffs

The cloud-versus-local decision is not about which is better in the abstract, it is about which constraint dominates your deployment. Cloud models win on raw capability and zero setup: you get frontier reasoning instantly with nothing to maintain, and you pay per token. Local models win on cost and privacy: once the hardware is bought, inference is free and no data ever leaves the machine, but you accept a real drop in capability and you own the maintenance. Most OpenClaw builds do not pick one side. They run cloud for the hard reasoning steps and local for the private or high-volume routine work, which is only possible because the framework is model-agnostic by design.

Cost Comparison: Cloud API vs Local Hardware

The cost picture flips depending on volume. At low volume, cloud APIs are almost always cheaper because you avoid buying hardware and you only pay for what you use. At high, steady volume, a local model on hardware you already own can undercut cloud APIs dramatically, because your marginal cost per message approaches zero. The two-tier strategy captures most of the savings without the capability hit: send the small share of genuinely hard requests to a frontier cloud model and route the large share of routine requests to a cheaper or local model.

Relative Monthly Cost by Model Strategy (illustrative)

Frontier cloud model for everything100%

Two-tier: frontier for hard, cheap for routine32%

Budget cloud model (DeepSeek / Gemini Flash)18%

Local model on owned hardware4%

The percentages are illustrative, but the ranking is stable: running a frontier model for every task is the most expensive habit in agentic AI, and the largest single lever you have is routing routine work off it. We break the mechanics down in the two-step OpenClaw cost fix.

Latency, Reliability, and Data Privacy

Beyond cost and capability, three operational factors decide the cloud-versus-local split. Latency: cloud frontier models are fast and consistent, while local latency depends entirely on your hardware and can be sluggish on smaller machines. Reliability: cloud providers occasionally have outages you cannot control, whereas a local model is only as reliable as the box it runs on, which you can control but must maintain. Data privacy: this is where local is uniquely strong, because sensitive customer data never leaves your infrastructure, which matters for regulated clients and for any agent handling personal information. If privacy is a hard requirement, local inference is not a cost optimization, it is a compliance decision.

When to Choose Local and When to Stay on the Cloud

Choose local when: data privacy is non-negotiable, volume is high and steady, the tasks are routine, and you have the hardware to run a capable model at acceptable speed.
Stay on the cloud when: you need frontier reasoning, volume is low or spiky, you want zero maintenance, or the agent makes high-stakes judgment calls where capability outweighs cost.
Do both when: you can identify a clear split between hard and routine work, which is the case for most production agents.

If you are still setting up your first agent, start with the install and Telegram guide, then come back here once you have a working agent to route.

How to Choose the Right Model for Each Task

Model Selection Framework

Complex reasoning & judgment callsClaude (Sonnet/Opus)

Structured output & code generationGPT-4o / Claude

Multimodal (images, video, audio)Gemini

Budget-sensitive background tasksDeepSeek / Gemini Flash

Full privacy, zero costOllama (local)

The key insight is that you do not need to pick one model for everything. OpenClaw supports primary and secondary model configurations, which is exactly what we optimize in Day 4.

What is Next

Now that you understand the model landscape and have a local option connected, Day 4 focuses on cost optimization, setting up a two-tier model strategy that cuts your monthly bill by 70% or more without sacrificing quality where it matters.

Need help choosing the right model configuration for your use case? OpenClaw Consult is an OpenClaw consulting team, we help with setup, troubleshooting, and custom agent builds.