Gemma 4 Trips in Local Runs

This is the classic first-week open-model reality: impressive launch materials, then messy local runs expose conversion, sampler, or runtime edge cases.

• –Gemma 4 is a major release with multimodal and agentic positioning, so even isolated failure reports matter to adopters benchmarking local workflows.
• –Similar community threads are already surfacing around `llama.cpp`, `MLX`, `Unsloth`, KV cache usage, and token leakage, which suggests backend mismatch can masquerade as model failure.
• –For developers, the practical lesson is to validate across runtimes before blaming the weights, especially when quantization and long-context support are involved.
• –If the underlying issue is reproducible, it is the kind of bug that will shape which inference stack people trust for production.