Error messages

llama_model_load: error loading model

llama_model_load: error loading model

Every model-loading failure in llama.cpp funnels through one catch block, so this line turns up whatever went wrong. Here's a real one from #22048:

llama_model_load: error loading model: missing tensor 'blk.24.attn_k.weight'
llama_model_load_from_file_impl: failed to load model
common_init_from_params: failed to load model '/path/to/model.gguf'
srv    load_model: failed to load model, '/path/to/model.gguf'

We read the loader in release v0.5.0 (September 23, 2026). The lines below the first one only repeat that loading failed, one layer further out each time. The first line carries the reason, after error loading model:, and that's the only part worth searching for. As of September 25, 2026, 306 issues in the llama.cpp repo quote this line.

Missing tensor, or the wrong number of them

A tensor is one named block of weights, and blk.24.attn_k.weight is layer 24's attention key matrix. llama.cpp knows which tensors each architecture should have. missing tensor '...' means the file lacks one it expects, and done_getting_tensors: wrong number of tensors; expected N, got M means the count doesn't match.

Both usually mean the GGUF and your build disagree about the model. In #25657 early Hy3 GGUFs named one tensor differently from llama.cpp's own converter, and it was fixed on the uploader's side when they replaced the files. In #22048, the ggml-org and lmstudio-community GGUFs of the same model loaded fine and only a modified copy failed. And in #23033 a contributor answered a missing tensor with "Normal", because support for that part of the model was still an open pull request.

Unknown architecture or pre-tokenizer

unknown model architecture: '...' means the GGUF declares a model family your build has never heard of. error loading model vocabulary: unknown pre-tokenizer type: '...' is the same problem for the tokenizer (the step that splits text into tokens). A build from before the model came out can't load it. Update, or check whether support has been merged yet.

Failed to open, or not a GGUF at all

If the reason is llama_model_loader: failed to load model from and a path, look one line up. The GGUF reader logs why first: failed to open GGUF file with the operating system's reason in brackets (a wrong path gives "No such file or directory"), failed to read magic for a file too short to have a header, or invalid magic characters ending expected 'GGUF' when the file isn't a GGUF. Either of the last two can mean the download stopped early or saved something other than the model.

What to do

• Run llama-cli --version and compare the build with the date the model was released. Most failures on brand-new models go away on a current build.
• Try the same model from a different uploader, such as the ggml-org or lmstudio-community copies, to tell a bad file from a llama.cpp problem.
• For open or magic errors, check the path and the file size against the size listed on the download page.

If a current build still fails on a well-known upload, that's worth an issue. Paste the whole first line, since it's the part the maintainers will ask for.