LiteRT conversion stumps Gemma 4 tinkerers

This reads less like a launch announcement and more like a usability check on LiteRT’s developer experience. The tooling is real, but outside the exact supported Gemma conversion path, you’re juggling export assumptions, tokenizer packaging, and memory constraints all at once.

• –Google’s docs do support safetensors-to-LiteRT for Gemma-style models via LiteRT Torch, but not as a generic “convert any safetensors checkpoint” workflow.
• –The pipeline is more than weight conversion: the final bundle also needs tokenizer assets and task metadata.
• –Kaggle free-tier compute is a weak fit for long CPU-bound export and quantization jobs.
• –For most builders, the practical path is to start from an officially supported checkpoint, or use a community format like GGUF/MLX first and only switch to LiteRT if the target platform demands it.