Gemma 4 can't use tools

A user who tried Gemma 4 26b-a4b wrote that while the performance impression was good, clear problems showed up in actual agent workflows.

• Performance degraded sharply as context grew longer, more noticeably than with other models.
• It largely ignored the system prompt, and responses didn't change no matter how strongly it was instructed.
• It could barely make tool calls, failing to call properly even when explicitly requested.

The author explained that even strong prompts like zero-assumption, tools always use, and parallel tool calls were tried but had no effect. They also noted that words like tool, system, and check never appeared at all in the reasoning output, and that when asked about the system prompt, the model would state its content but give a false explanation for why it wasn't following it.

In the comments, many raised the possibility of a runtime/template issue as the cause.

• An opinion that the latest llama.cpp should be used
• Advice to re-download the quantized model received early in Gemma's release
• Mentions of a CUDA 13.2 issue and some low-bit quantization (IQ, lower K) issues
• A tip that llama.cpp's undocumented interleaved template might improve tool calling
• An explanation that for Gemma 4's system prompt, the position of the <|think|> token matters, and reasoning can break if the format is wrong
• Conversely, there were also accounts that it worked better with a minimal sysprompt in llama-server webui or opencode

In other words, this is a real-world case showing that template, runtime version, quantization, and thinking format—rather than the model's own inherent limits—can significantly determine agentic performance.