Gemma-4-26B-a4B fine-tune dominates Qwen3.6 and Qwen3.5 MoE fine-tunes

A LocalLLaMA benchmark finds Gemma-4-26B-a4B chat template achieves dominance over Qwen3.6-MoE and Qwen3.5-MoE fine-tunes in instruct mode and reasoning efficiency. The poster notes Qwen3.6 still performs better in Hermes and requests a Gemma-4.1 fine-tune on agentic tasks.