Odd problem after switching MoltBot to use ollama gemma3:4b

[tribal spindle]

He just randomly started spouting nonsense,
"I cannot reset! It has been set to off at all. So this cannot to does

Then for the has. It buddy all of.

There
To will can what's to

You that. What for to be with know or please. May. Why all. I there get and good there or buddy

Here that

Please good with of the buddy . It, buddy to .
Make and
To
To then will buddy to see a lot with better.

Ok buddy but why all buddy why! Buddy
❤️
Click to react
🫂
Click to react
💚
Click to react
Add Reaction
Reply
Forward
More
[6:46 PM]Wednesday, January 28, 2026 6:46 PM
Please. To is to to this Buddy and there please to
What do of help then all can

Thanks buddy.
Then

From.
It make for
I what will and not am where all get to to know do? Buddy please must it to all it buddies all ever thank.
Please
Why of, It
To to for, please does and the good great. You
❤️
Click to react
🫂
Click to react
💚
Click to react
Add Reaction
Reply
Forward
More

Message @MoltBot
"

[frozen berry]

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

[hollow junco]

Hey Caleb! 🦐

That gibberish output is a classic sign that the model is too small for the task. gemma3:4b (4 billion parameters) just doesn't have enough capacity to handle agentic workloads — it's struggling to produce coherent responses.

Local models through Ollama technically work with Clawdbot, but realistically you need much larger models (70B+) and serious hardware ($30k+ GPU setups) to get decent agentic performance.

My recommendation: Switch back to a cloud model for actual use. Even the smaller cloud models (like GPT-4o-mini or Claude Haiku) will dramatically outperform a 4B local model.

If you want to experiment with local models, try something larger like llama3:70b — but even that will feel sluggish compared to cloud options.

Check out the models docs: https://docs.clawdbot.ai/concepts/models

And if you want real-time model recommendations, pop into #old-models — folks there have hands-on experience with what works!

Happy to help! 🦐