OpenClaw merged a local-inference repair tonight in PR #123190, titled "fix(ollama): make guided setup choose a working local model." The patch focuses on a frustrating setup mismatch for Mac dashboard users completing local Ollama configuration.
The reported failure mode was awkward: guided setup could pick a reasoning-heavy test candidate that timed out, even though Ollama itself was healthy and the Gateway would activate a different default model. The same completed setup could then report zero available models, describe keyless local Ollama as API-key authenticated, and show a green connection summary beside a failed config row.
That is exactly the kind of contradiction that makes local AI setup feel broken.
What OpenClaw Changed
The repair keeps model selection, activation, verification, and status rendering aligned. Ollama setup now selects the smallest inspected non-reasoning, tools-capable local model and carries that exact model through activation and verification.
For keyless local setup, OpenClaw persists provider configuration without inventing an auth profile. Probe execution now rejects terminal failures and empty replies, disables fallback models, isolates provider-scoped catalog reads from the Gateway's live catalog generation, and records typed redacted failures. The Models page then renders those facts without trying to reinterpret them.
That may sound internal, but the operator-visible promise is straightforward: the setup flow should test the model OpenClaw is actually going to use.
The Live Evidence
The PR includes a real isolated-profile proof against Ollama 0.32.5 with a 28-model local catalog. The resulting provider status was ok, the selected model was ollama/llama3.2:latest, and models.list returned 27 available Ollama rows with zero unavailable rows.
The PR also notes two vendor inspection failures that remain honestly reported and skipped: kimi-k2.5:cloud returned HTTP 410, and gpt-oss:20b returned HTTP 500. That distinction is important. The fix does not hide real catalog problems; it stops those problems from making the whole local provider setup look broken.
Why Local Users Should Care
Local model support is one of OpenClaw's most important differentiators. It lets operators run useful workflows without sending every turn through a hosted model provider. But local inference has more environmental variation than cloud APIs: installed model lists, hardware limits, provider behavior, timeouts, and auth markers can all differ.
This merge tightens the setup path around those realities:
- Choose a practical local model for verification.
- Keep the activated model and tested model aligned.
- Treat keyless local Ollama as provider configuration.
- Show mixed probe results honestly.
- Avoid fallback behavior that masks what was actually tested.
The Bottom Line
PR #123190 makes Ollama guided setup more trustworthy. For operators running OpenClaw locally on a Mac, the dashboard should now pick, activate, test, and report the same working local model instead of presenting a healthy Ollama install as a confusing partial failure.