Gemma 4 fails at agentic tasks despite being a strong chatbot
Gemma 4's 31B model struggles with multi-turn agentic work.
A frustrated user on Reddit reports that Gemma 4 (the 31B QAT GGUF from Unsloth) performs poorly on agentic tasks requiring multiple tool calls and persistent reasoning, despite being a capable chatbot. They tried an advanced config: flash-attn, preserve_thinking=true, 147K context, temp 1.0, top-p 0.95, and still the model stops after a couple of tool calls, saying 'I did A but B and C happened, now I'm gonna do D' then halting. Only when prodded does it resume.
Meanwhile, Qwen 3.6 27B at UD-Q5_K_XL, DeepSeek V4 Flash at UD-IQ3_XXS, and even the older GPT-OSS 120B all keep 'fruitfully chewing on the problem' indefinitely. The user notes Gemma 4 is great for discussion (e.g., worldbuilding) but 'doesn't want to put any effort into agentic work.' This points to a possible training bias or lack of reinforcement for agentic persistence in the model, despite its strong baseline capabilities.
- Gemma 4 (31B GGUF) stops after a few tool calls in agentic tasks, requiring manual prompting to continue.
- Qwen 3.6 27B, DeepSeek V4 Flash, and GPT-OSS 120B handle the same multi-step tasks without stalling.
- Despite agentic failures, Gemma 4 excels as a conversational chatbot, suggesting a training bias against agentic persistence.
Why It Matters
Professionals relying on AI agents for automation may find Gemma 4 unreliable for complex, multi-step workflows.