Free models for OpenClaw

The search for a free AI model for OpenClaw has an honest answer in two parts. Yes: OpenClaw runs local models through Ollama, LM Studio, vLLM, SGLang and other local servers, and a local model costs nothing per token. And no: the docs are explicit that smaller models are more susceptible to prompt injection and tool misuse, so a free model is fine for a chat-only assistant and a bad idea for one that runs tools on an inbox you do not control. Here is the setup, the discovery rules, and where the line is.

Ollama

• OpenClaw talks to Ollama’s native API, not the OpenAI-compatible endpoint, which breaks tool calling; the base URL is the host without a version path.
• Three modes: cloud plus local through one host, cloud only, or local only. Local and LAN hosts need no real token; remote hosts and Ollama Cloud need a real credential.
• Onboarding discovers only models already loaded in memory, with tool support and at least sixteen thousand tokens of context confirmed; an installed but unloaded model is not a candidate, and the route still needs a real completion before it is saved.
• Vision models, node-local inference on a paired machine, embeddings for memory search and a web-search mode each have their own page.

LM Studio and the others

LM Studio serves GGUF or MLX models from a desktop app or a headless daemon on a local port; vLLM and SGLang serve models for people with the hardware; there are pages for a local DeepSeek and for a generic local server. All of them are providers in the same list as the hosted ones, switched in config, so a hybrid setup, a local model for cheap turns and a frontier model for the ones that matter, is ordinary.

Do not use older/weaker/smaller tiers for tool-enabled agents or untrusted inboxes.

The rule

The prompt-injection page draws the line: for a chat-only personal assistant with trusted input and no tools, smaller models are usually fine; for a bot that can run tools or touch files and networks, use the latest best tier, and if you must use a smaller model, reduce the blast radius with read-only tools, strong sandboxing and strict allowlists. Prompt injection and OpenClaw has the whole argument; what an OpenClaw agent actually does shows which of the things people run it for need tools, which is most of them.

Hardware, honestly

A local model that can call tools and hold a useful context needs memory the average laptop does not have to spare, and the community keeps a lite-mode skill for machines with two to four gigabytes precisely because OpenClaw itself already wants some. OpenClaw system requirements covers what the runtime needs before the model is even counted.

On Diali

Diali does not run local models; every plan includes a monthly balance of AI credits across managed models, and bring-your-own-key covers ten providers. That is the trade: no free tier, and no small model quietly running your inbox. Hosted OpenClaw on Diali is the assistant; the AI credit calculator turns a month of expected use into a number.

• Ollama’s native API, a loaded model, a real completion before it saves.
• Local for chat-only; frontier for anything with tools.
• Free per token is not free of memory or risk.