Gemma 4 26B, 31B diverge on coding

The interesting part here is not just speed versus quality; it’s that local runtimes may be amplifying a real behavioral gap between the MoE and dense checkpoints. Gemma’s own docs frame 26B A4B as a fast, 3.8B-active-parameter model with configurable thinking, so if it starts sounding like a philosopher, that could be a mix of model architecture, chat template handling, and sampler defaults rather than “just” parameter count.

• –Google’s official model card says the 26B A4B model is MoE with 3.8B active parameters, while the 31B is dense, so different inference dynamics are expected.
• –The model docs also call out specific sampling defaults and thinking-mode tokens, which makes prompt/template fidelity a likely culprit when behavior looks off.
• –Community reports around Gemma 4 on llama.cpp have already pointed to sampler and template quirks, so the discrepancy may be runtime-specific rather than a pure model-quality issue.
• –For a fair comparison, the clean test is the same prompt, same chat template, same thinking setting, and the same backend in vLLM or Transformers before blaming MoE itself.
• –If the 26B still over-explains after that, it may simply be the model’s preferred style under agentic prompting, not a bug.