The GGUF file won't load

New architectures land in llama.cpp before they reach the tools that wrap it, so a model published this week may need a build from this week. The other common case is a file that isn't what it appears: a shard downloaded without its siblings, an interrupted transfer, or a repository file that is a vision projector rather than a model.

1. 1
   Update the runtime

   Architecture support is added upstream first. Wrappers can lag by weeks, which is the usual explanation for a model that works for everyone else.

2. 2
   Check you have every shard

   Large quantizations are split across files named -00001-of-0000N. All of them are required; the first one alone is not a smaller version of the model.

3. 3
   Verify the file size against the published size

   Compare against the exact byte count on the model's quantization page here. A mismatch means an incomplete download, which usually presents as a parse error rather than a size warning.

4. 4
   Load the projector too, for vision models

   A vision-language model in GGUF ships its vision encoder as a separate mmproj file. Without it the model loads and then ignores your images.

Rather than guessing, pick your model and card and read the grid: every quantization at every context, with the memory each combination actually needs.