Raising Gemma 4's token budget fixes dense model refusals

Developer discovers that insufficient token budget (max_tokens) was causing Gemma 4 dense model refusals, not architectural differences.

Published 11sem1 source

≈ 20sSpeed

The fact

Increasing token budget restores the dense model's performance across all tested scenarios.

Corrects previous analysis that blamed MoE-vs-Dense architectural divergence rather than resource constraints.

Click the link to read an article on the topic:

Explore this topic

Auto-synthesis from 1 media source · identified on May 27, 2026