Gemma 4 loops in LM Studio

This looks like an integration bug disguised as a model failure. Gemma 4 is meant to run locally, but if the runtime is feeding it the wrong chat format or stop tokens, the model can spiral into exactly this kind of self-referential loop.

• –The output tokens shown here are from non-Gemma chat schemas, which points to a prompt/template mismatch or incorrect stop-sequence handling.
• –Vulkan plus quantized KV cache plus a MoE model is a brittle stack; any backend edge case can turn into repeated garbage generation.
• –Google positions Gemma 4 as a local-first, agentic open model family, so a failure like this is a support-gap issue that matters for real-world adoption.
• –The first things to try are disabling KV-cache quantization, verifying the Gemma 4 chat template, and testing a different backend or build.