Model won't load

KoboldCpp starts, but loading the model fails. The last error line is usually generic; the real cause is printed a few lines above it.

If loading stops with an "out of memory" or allocation error, see Out of memory.

"Could not load text model"

Could not load text model: C:\models\mymodel.gguf

Sometimes also failed to load model. This is the final message for any load failure.

1. Scroll up in the console and find the first error above this line.
2. Look it up in the sections below or in Out of memory.

"unknown model architecture"

llama_model_load: error loading model: unknown model architecture: '<name>'

Older versions print error loading model architecture: instead. Your KoboldCpp does not know this model's architecture, usually because the model is newer.

1. Update KoboldCpp to the latest release.
2. If the latest release does not support it yet, wait for the next release, or try a rolling build. Rolling builds are updated automatically and can support a new model earlier, but may be unstable.

"does not seem to be a GGUF file"

WARNING: Selected Text Model does not seem to be a GGUF file! Are you sure you picked the right file?

The file name does not end in .gguf. KoboldCpp runs text models only as GGUF files (and legacy GGML .bin files). It cannot load .safetensors or PyTorch text models.

1. Download the GGUF version of the model. Search for the model name plus "GGUF" on Hugging Face.
2. Download one single .gguf file, not the whole repository. If the quant is split into parts (-00001-of-00003.gguf), download all parts and select the first. See Choose a model.

A model that exists only as safetensors must be converted to GGUF first. See GGUF and quantization.

"all download methods failed"

Could not find suitable download software, or all download methods failed. Please install aria2, curl, or wget.

KoboldCpp could not download a model file from a template or URL. The lines above it, such as curl failed: …, show why.

1. Check your internet connection.
2. On Linux, install curl, wget or aria2.
3. Launch again. Files that finished downloading are reused.

The download is broken or incomplete

llama_model_load: error loading model: read error: Input/output error
error: failed to read GGUF file

The file is damaged: the download stopped early, or the disk has read errors.

1. Compare the file size with the size on the download page.
2. Download the model again. If KoboldCpp downloaded it from a URL, delete that file first; otherwise KoboldCpp reuses it and prints already exists, using existing file.
3. If a fresh copy fails with Input/output error too, check the drive for errors, or move the model to another drive.

--analyze mymodel.gguf prints the file's metadata and tensors, which shows whether KoboldCpp can read it. In the launcher: Analyze Model on the Extra tab.

The vision file (mmproj) fails to load

error: failed to load mmproj model!

Loading can also crash without this message. A vision or audio projector (mmproj) works only with the model it was made for. For example, a Qwen mmproj cannot work with a Gemma model.

1. Use the mmproj made for your exact base model and size. A Gemma3 12B model needs the Gemma3 12B mmproj.
2. Ready-made projectors for popular models are in the koboldcpp/mmproj repository.

See Vision and audio input.

Access violation while loading (Windows)

OSError: exception: access violation reading 0x…

This has several causes, among them a graphics driver problem (reported with AMD cards on Vulkan), an unstable overclock, and non-ASCII characters in a path.

1. Update your graphics driver.
2. Try another backend. With an AMD card, try Use Vulkan (Old CPU); otherwise try Use CPU to rule out the graphics card.
3. Lower Context Size: and try a small model to see if it loads at all.
4. Use folder and file names with plain ASCII characters. For image models, this can include your Windows user name; see OS-specific problems.
5. If you overclock your CPU, GPU or RAM, test at stock speeds.

Advanced: old GGML (.bin) models

GGML is the format before GGUF. KoboldCpp still loads it, but with limits:

Warning: Your model may be an OUTDATED format (ver 3). Please reconvert it for better results!
Warning: Only GGUF models can use max context above 16k. Max context lowered to 16k.

Get a current GGUF version of the model instead. If a GGML model is detected as the wrong type, download a fresh copy; the file may be damaged or badly converted.

Related

• Out of memory
• Choose a model
• Updating