The vision weights of google/gemma-4-12B-it, so a decisio server on Apple silicon can read images with a Gemma 4 12B model whose MLX conversion holds only the language model.
decisio loads this file by itself: start the server with --backend mlx --images on, and when the model's own files carry no vision weights, decisio downloads vision.safetensors from here once, checks its sha256 and merges it in.
To run offline, pass the file with --vision-weights PATH.
One file serves the base model and every tuned model of that base, since tuning leaves the vision weights unchanged.
vision.safetensors: 10 tensors, 99.8 MB, bf16: the patch embedder and its projection (Gemma 4 12B reads pictures without a separate vision encoder).google/gemma-4-12B-it at revision 707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7: names, shapes, dtypes and bytes unchanged.vision_config.json: the source model's vision_config and image token ids, so an offline start needs nothing else.vision_manifest.json: each tensor's shape, dtype and sha256, and the file's sha256 (112725a204111dfdc5d29d2c4456818e4b6de33109991edfaa7120c731fb06a8).Apache 2.0, the licence of the source model.
The weights are Google's, from google/gemma-4-12B-it; see its model card for its terms, intended use and limitations.
google/gemma-4-12B-it