#gemma not responding

1 messages · Page 1 of 1 (latest)

lean moat
#

i"ve configured gemma on openclaw but it does not working anymore

it says startup model warmup failed for google/gemma-4-26b-a4b-it: Error: Unknown model: google/gemma-4-26b-a4b-it

olive badgerBOT
torn echoBOT
#

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

olive badgerBOT
crude agateBOT
#

Let me verify the current CLI behavior and where the docs live, then I’ll give them a concrete fix.

#

First thing, what does openclaw --version say?

I checked the docs/code, Gemma 4 support landed in 2026.4.7, so if you’re older than that, update first.

Also, that specific startup line is a bit misleading. startup model warmup failed ... Unknown model ... is from a startup warmup check, and that path only does static model resolution. Gemma 4 on google/... is handled more dynamically, so the warning can show up even when the model ref is valid.

Try this:

openclaw --version
openclaw models set google/gemma-4-26b-a4b-it
openclaw models status

If you’re already on 2026.4.7+ and it still won’t answer, paste these 3 things here:

  1. openclaw --version
lean moat
#

🦞 OpenClaw 2026.4.15 (041266a)
Less clicking, more shipping, fewer "where did that file go" moments.

Config : ~/.openclaw/openclaw.json
Agent dir : ~/.openclaw/agents/main/agent
Default : google/gemma-4-26b-a4b-it
Fallbacks (0) : -
Image model : -
Image fallbacks (0): -
Aliases (1) : gemini -> google/gemini-3.1-pro-preview
Configured models (2): google/gemini-3.1-pro-preview, google/gemma-4-26b-a4b-it

Auth overview
Auth store : ~/.openclaw/agents/main/agent/auth-profiles.json
Shell env : off
Providers w/ OAuth/tokens (0): -

  • google effective=profiles:~/.openclaw/agents/main/agent/auth-profiles.json | p rofiles=1 (oauth=0, token=0, api_key=1) | google:default=AQ.Ab8RN...fS1LYvVw
crude agateBOT
#

Their config/auth look basically fine, so I want to separate the harmless warmup warning from the real runtime failure and give them one solid next check.

#

Nice, this helps a lot.

Your version is new enough (2026.4.15), and your config/auth also look basically correct. So I don’t think the startup model warmup failed ... Unknown model ... line is the real blocker.

That startup check is a bit noisy for some Google/Gemma refs. The useful next check is the live probe + runtime log:

openclaw models status --probe --probe-provider google

Then send one test message to the bot, and immediately run:

openclaw logs --plain --limit 200

Please paste both outputs.

lean moat
#

🦞 OpenClaw 2026.4.15 (041266a) — Ah, the fruit tree company! 🍎

│
◇
Config : ~/.openclaw/openclaw.json
Agent dir : ~/.openclaw/agents/main/agent
Default : google/gemma-4-26b-a4b-it
Fallbacks (0) : -
Image model : -
Image fallbacks (0): -
Aliases (1) : gemini -> google/gemini-3.1-pro-preview
Configured models (2): google/gemini-3.1-pro-preview, google/gemma-4-26b-a4b-it

Auth overview
Auth store : ~/.openclaw/agents/main/agent/auth-profiles.json
Shell env : off
Providers w/ OAuth/tokens (0): -

  • google effective=profiles:~/.openclaw/agents/main/agent/auth-profiles.json | profiles=1 (oauth=0, token=0, api_key=1) | google:default=AQ.Ab8RN...fS1LYvVw

OAuth/token status

  • none

Auth probes
┌───────────────────────────┬──────────────────────────┬────────────┐
│ Model │ Profile │ Status │
├───────────────────────────┼──────────────────────────┼────────────┤
│ google/gemma-4-26b-a4b-it │ google:default (api_key) │ ok · 15.9s │

crude agateBOT
#

Perfect, that rules out the main suspects.

Your model ref is valid, and your Google auth probe is passing, so the startup warmup failed ... Unknown model ... line is basically a red herring here.

At this point I want the runtime error from an actual failed turn.

Please do this:

Terminal 1

openclaw logs --follow --plain

Then

  1. send one test message to the bot
  2. copy the first error that appears after that message
#

If nothing useful shows up, also run:

openclaw health

and paste that too.

I’m specifically looking for the error after the failed reply, not the earlier startup warning.

lean moat
#

ok
i just figured out what happened

i give him a instruction hard that he stills running and not responding me

how can i stop him

crude agateBOT
#

I’ve got the cause now. I’m checking the exact supported way to stop a stuck run so I don’t invent a command.

#

Yep, easiest fix:

In the same thread/chat where it’s stuck, send a standalone message:

stop

or

/stop

That should abort the current run.

If it spawned child agents too, use:

lean moat
#

<think> The user is prompting me. I'll try to get the configuration again. If that doesn't work, I'll try to find the config file on the system. </think><call:default_api:gateway{action:config.get}

responded that

crude agateBOT
#

That helps. I’m checking whether this is a known tool-calling/output-format issue with Gemma, so I can point them at the right fix instead of guessing.

#

Ah, yep, that’s different.

That output means Gemma is dumping raw reasoning/tool-call markup into chat:

  • <think>...</think> = internal reasoning
  • <call:...> = raw tool-call text

So this is not the old Unknown model problem anymore. It’s the model/session getting messy with tool use.

I’d recover like this, in the same thread, as standalone messages:

/stop
/reasoning off
/verbose off
/trace off
# 
/new

Then test again.

If you want a quick sanity check, switch this thread to your existing alias:

/model gemini

If a fresh session still leaks raw <think> / <call:...> text with reasoning/verbose off, that smells like a real bug. Then paste:

openclaw status --all