NVIDIA Gemma 4 NVFP4 targets Blackwell GPUs

This is less a broken model than a format/runtime mismatch. The checkpoint is optimized for NVIDIA's NVFP4 path, which points you toward vLLM and Blackwell, not a generic Ollama workflow.

• –NVIDIA's model card explicitly lists vLLM support and Blackwell hardware compatibility, so that is the intended execution path.
• –Ollama is generally centered on GGUF/llama.cpp-style workflows, so this checkpoint is unlikely to drop in cleanly. This is an inference from the model/runtime docs and the discussion, not a direct NVIDIA statement.
• –If you want local inference on consumer GPUs, a different Gemma 4 quantization or a GGUF/AWQ variant is the practical route.
• –The useful takeaway for developers is that "safetensors" alone does not guarantee broad local compatibility; quantization format and target runtime matter more than file extension.