Shipping on Gemma 4: chain-of-thought leakage, MoE-vs-Dense, and on-device pragmatism

This is a submission for the Gemma 4 Challenge: Write About Gemma 4

I built and shipped Curio Kid, a kid-safe multimodal Android app where my 6-year-old asks Luna (a Gemma-4-powered tutor) anything by text, voice, or camera. The product story is in my other submission. This post is the engineering writeup — three things about Gemma 4 that I had to actually work around in production, with the code and reasoning behind each fix.

If you're about to ship a Gemma 4 app, these are the three traps I'd want to know about on day one.

1. Chain-of-thought leakage is real, and it hits the user

Shipping on Gemma 4: chain-of-thought leakage, MoE-vs-Dense, and on-device pragmatism

Other newsrooms on this story

Related reading

I switched my Gemma 4 model three times in 72 hours. Here's the decision tree I…

From Fragmented Pipelines to Coherent Intelligence — Why Gemma 4 Actually…

I was trying to Learning About Gemma 4 and It was pretty good

Prompt Physics: Building a Cognitive Steering Layer for Gemma 4

What Gemma 4 Doesn't Know About Cameroon — and What That Taught Me About…

Gemma 4 on 16GB RAM: What Actually Works for Structured AI Workflows

Other newsrooms on this story

Related reading

I switched my Gemma 4 model three times in 72 hours. Here's the decision tree I…

From Fragmented Pipelines to Coherent Intelligence — Why Gemma 4 Actually…

I was trying to Learning About Gemma 4 and It was pretty good

Prompt Physics: Building a Cognitive Steering Layer for Gemma 4

What Gemma 4 Doesn't Know About Cameroon — and What That Taught Me About…

Gemma 4 on 16GB RAM: What Actually Works for Structured AI Workflows