Gemma 4 Suddenly Stopped Answering — How an External Collaboration Found the Root Cause in 24 Hours

In PROJECT JAMES’s local Graph-RAG system, some Gemma 4 stages were returning deterministic empty responses.

Cross-checks with external reports and single-variable internal tests showed the issue disappeared when max_tokens was raised.

The model was spending about 500 hidden reasoning tokens before any visible output, so the old cap left no room for a reply.

A code change increased the default token limit to 4096 across four stages, and PR #399 was merged to fix the outage quickly.