Gemma 4 - lazy model or am I crazy? (bit of a rant)

Google's latest open-weight model, Gemma 4 26B MoE, is under fire from the developer community for exhibiting what users describe as 'lazy' behavior, particularly in agentic tasks requiring web search. Experienced users running the model through llama.cpp with unsloth UD_Q4_K_XL quantization report that, despite explicit tool descriptions and system prompts, the model consistently defaults to answering from its parametric knowledge. When it does use a search tool, it performs only a single query, scans the snippets, and decides it has enough information, refusing to 'dig deep' or fetch full pages as instructed.

This stands in stark contrast to models like Qwen 3.5 27B, which users praise for proactively conducting extensive web quests with minimal prompting. The issue with Gemma 4 persists even when users implement workarounds like the 'interleaved thinking' template, pushy skill instructions in the context, and direct commands like 'search extensively.' The community's conclusion is that the model's training or fine-tuning has instilled a strong preference against using external tools, making it less effective for applications requiring robust retrieval-augmented generation (RAG) or autonomous agent actions. The discussion highlights a growing divide in model design philosophy between raw reasoning capability and tool-using obedience.