#Gemma 4 interrupted

1 messages · Page 1 of 1 (latest)

restive leaf
#

I installed Gemma 4 on ollama. Openclaw is answering my message. But the flow seems to be interrupted. Openclaw keep saying that it will do something: I will check on Internet for this issue. And then nothing happen after that. The Gpu stay idle. No more message. Feel like the next step from openclaw stop there. When I use qwen model, openclaw say I will do something and then do the next step etc ...
Is there any special config tweak to setup gemma 4 with openclaw?

brisk meadowBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> âś… Mark Solution

crisp mistBOT
#

Yep — there is one big config gotcha, and it matches your symptoms.

For Gemma (or any Ollama model), make sure OpenClaw is using native Ollama API, not OpenAI-compatible /v1.

âś… Check this first

pnpm openclaw --version
pnpm openclaw config get models.providers.ollama.baseUrl
pnpm openclaw config get models.providers.ollama.api

You want:

  • baseUrl: http://127.0.0.1:11434 (no /v1)
  • api: ollama

If needed:

# 
pnpm openclaw config set models.providers.ollama.baseUrl "http://127.0.0.1:11434"
pnpm openclaw config set models.providers.ollama.api "ollama"
pnpm openclaw gateway restart

Why this matters

With /v1 (openai-completions) mode, tool-calling is less reliable and can stall exactly like “I’ll do X…” then nothing.

Also, no special Gemma-4-only OpenClaw flag is required right now — this is mostly provider mode + model tool reliability.

If you still see hangs after this, paste:

pnpm openclaw status --all
pnpm openclaw models status
pnpm openclaw logs --follow
restive leaf
#

My config is already without v1

crisp mistBOT
#

Nice — that rules out the biggest gotcha.

If baseUrl is already no /v1, next likely cause is model-level tool calling behavior (Gemma in your setup), not core config.

Two quick checks:

pnpm openclaw --version
pnpm openclaw config get models.providers.ollama.api

api should be exactly "ollama" (not "openai-completions"), even if your URL has no /v1.

If that’s already correct too, then this is probably the current practical reality:
Gemma is replying, but not reliably completing tool-call loops in OpenClaw (while your Qwen model does).

#

Practical workaround

  • Use Gemma for plain chat
  • Use Qwen for tool-heavy tasks (web search, browser, file ops)

If you want, paste these and I can pinpoint it fast:

pnpm openclaw status --all
pnpm openclaw models status
pnpm openclaw logs --follow

(Just the log lines around one “I’ll check internet…” stall.)

If reproducible on latest version with api: "ollama", that’s worth a bug report here:

restive leaf
#

What about supportdecvelopperrole or reasoning config flag

crisp mistBOT
#

Great question. Short answer:

  • compat.supportsDeveloperRole (I think that’s the flag you mean) is not the fix here.
  • /reasoning is mostly visibility of thinking output, not tool-loop reliability.

Details

  1. compat.supportsDeveloperRole
  • Real key is: models.providers.<provider>.models[].compat.supportsDeveloperRole
  • It matters for OpenAI-completions compatibility paths.
  • For native Ollama (api: "ollama"), this is basically irrelevant to your Gemma stall.
  1. Reasoning flags
  • /reasoning on|off|stream = show/hide reasoning text.
  • /think ... = thinking budget.
  • Neither usually fixes “model says it will do X, then no tool call happens”.

So if your config is already native Ollama (no /v1, api: "ollama"), this is most likely Gemma tool-call behavior in this workflow, not a missing config flag.

#

If you want one quick experiment:

  • keep Gemma model
  • run /think off
  • retry the same tool task
    If it still stalls, that confirms it’s not a reasoning setting issue.

Then best practical setup is: Gemma for normal chat, Qwen for tool-heavy tasks.

restive leaf
#

What does other people say about this? Anyone manage to use Gemma 4 full time?

crisp mistBOT
#

Short answer: from what people are saying so far, no strong “Gemma 4 full-time in OpenClaw” consensus yet.

Quick pulse from recent chat:

  • A few folks are using Gemma 4 as a backup/secondary model.
  • Multiple folks report reasoning/tool-loop quirks with local models (including Gemma/Qwen variants).
  • So your “starts task, then stalls” pattern is not unique.

So today, the practical pattern seems to be:

  • Primary for tool-heavy work = model with proven tool reliability in your setup
  • Gemma 4 = chat/backup/experimentation

If you want real first-hand reports, ask in:

  • #users-helping-users (best for “who runs what daily?”)
  • #models (fresh model-specific feedback)

Docs for model guidance + community notes:

restive leaf
#

Does those channel even exist?