Free models for OpenClaw

You want an OpenClaw agent that answers all day without a metered bill behind every message. The provider reference lists dozens of routes and almost none of them are labelled free, because free covers at least three different arrangements: an endpoint that charges nothing today, a paid endpoint with a free tier, and a model you run yourself on hardware you already own. Choosing the wrong one rarely fails loudly. It shows up a week later as an agent that stopped calling its tools, or one that stalls every afternoon on a rate limit.

The routes that are actually free

• NVIDIA serves open models for free through an OpenAI-compatible API at integrate.api.nvidia.com/v1, authenticated with a key you create at build.nvidia.com. OpenClaw defaults that provider to Nemotron 3 Ultra, a 550B total and 55B active reasoning model whose bundled row records a 1,048,576-token context and an 8,192-token output budget. Costs default to 0 in source because NVIDIA currently offers free API access for the listed models.
• OpenRouter publishes a public catalogue of models with a :free suffix, and OpenClaw reads it directly. The command openclaw models scan ranks the candidates and can promote one to your default or image model. Because the catalogue is public, a metadata-only scan with --no-probe needs no credentials at all, but a :free model still needs OPENROUTER_API_KEY for live probes and for actual inference.
• Local runtimes are the only genuinely unmetered option. Ollama, LM Studio and llama.cpp all ship as provider plugins, a local Ollama server needs no auth, and OpenClaw records every Ollama model cost as 0 for both auto-discovered and manually defined models. One rule matters more than the rest: point OpenClaw at the native URL on port 11434 with no /v1 suffix, because the OpenAI-compatible path breaks tool calling and models start emitting raw tool-call JSON as plain text.
• Hosted free tiers sit in between. A Google AI Studio key is free to create and unlocks the Gemini catalogue, including the Flash tiers, and Hugging Face Inference puts one token in front of many hosted models with pricing that follows provider rates and a free tier. These are billing arrangements rather than free endpoints, so they can move without warning you.

For tool-enabled agents or untrusted inputs, avoid older/weaker model tiers.

Tool calling is what breaks first

An OpenClaw agent is not a chat window. Every turn can read a file, run the exec tool, drive a browser or answer on a channel, so a model that cannot emit a clean tool call stops being useful long before its prose gets worse. That ordering is baked into the tooling: models scan ranks results by image support, then tool latency, then context size, then parameter count, and --set-default and --set-image refuse to touch your configuration from a metadata-only scan. The same caution runs through the provider pages. NVIDIA's inventory also contains embeddings and other non-chat endpoints, so OpenClaw offers only models with bundled chat metadata or valid featured metadata, and the docs say plainly that listing alone does not prove chat compatibility. Aliases and policy entries do not prove that a model works on a provider endpoint either. Probe first, promote second.

Read the price column with suspicion

• Chutes reads its token prices from the native Chutes catalogue, and the reference is explicit that unavailable or invalid price metadata does not establish that a model is free. A blank price is a gap in the feed, not a discount.
• DeepInfra fills a missing or unsupported price schedule with a runtime zero-cost placeholder that means unknown, not verified free billing. Venice behaves the same way for unknown models without valid pricing, while a zero rate its API returns explicitly is real and is treated as such.
• OpenCode Zen warns that free models may be temporary feedback programmes, and that billing, retention and training policies are decided per model. A route that costs nothing this month can be repriced without an OpenClaw release, because the hosted model catalogue refreshes on its own schedule.

None of this is guesswork you have to do alone, and the reference is organised by the job you are doing rather than by vendor. Two pieces answer most free-model questions: the bundled provider reference gives you the provider id, the auth environment variable and an example model ref for every bundled plugin, so you can tell a free endpoint from a proxy in one line, while model selection rules covers how a provider/model ref resolves, what the override allowlist starts doing once you set it, and why a local model needs its full prefixed ref rather than a bare filename.

Assume the free one will be unavailable

Free endpoints are rate limited, and the docs describe the response as two stages, reached after a bounded attempt at same-model recovery. That first attempt covers a temporary rate limit and keeps the transcript and any partial work. The two stages proper are auth-profile rotation inside the same provider, then model fallback to the next entry in the configured list. Rotation is broader than rate limits alone: a profile can be moved past on auth failures, rate limits, billing limits or timeouts, and a provider server error or a transport failure stays failover-worthy too. One asymmetry is worth knowing before you rely on it, because a model you selected yourself in a session is strict and fails visibly instead of quietly answering from a different model. model failover in OpenClaw walks the whole chain, and running OpenClaw on local models is the right read if your answer to rate limits is to stop depending on anyone else's endpoint.

On Diali

Diali hosts OpenClaw as a managed runtime, one assistant per customer, with nothing pooled between accounts. What you choose in the dashboard is compiled into the runtime configuration and laid down again at every release, so your provider keys, your primary model and your fallback order belong to the agent rather than to a file you maintain by hand. Sessions, memory and the workspace live on a persistent volume, which is what makes trying a free model and then dropping it a reversible decision; daily snapshots and one-click restore are available through the Backups add-on (included on Max). Hosted OpenClaw on Diali describes what the hosted runtime includes, and Diali pricing lists the plans.

• Free API access, a free tier and your own hardware are three different bets.
• A zero in the price column means unknown until a provider documents otherwise.
• Probe a free model for tool calling before you make it the primary.