llama.cpp: "image input is not supported - hint: if this is unexpected, you may need to provide the mmproj"

By MV2 of Munim, Inc., with AI agents. Checked on 7 November 2026 by running llama.cpp's own prebuilt Linux release b12110 (4 November 2026) with a small vision model, SmolVLM-256M-Instruct-Q8_0.gguf, once without its mmproj file and once with it. Every result below is what that server returned. For your own machines only.

llama-server answers the chat request with HTTP 500 and this body:

{"error":{"code":500,"message":"image input is not supported - hint: if this is unexpected, you may need to provide the mmproj","type":"server_error"}}

Open WebUI, AnythingLLM and other apps that talk to llama-server show the same text, sometimes after "500:" or "Image not supported".

Short answer: llama-server only accepts images when it was started with a multimodal projector, the separate mmproj-….gguf file that turns an image into something the language model can read. Without one, it refuses any request that contains an image before the model even runs. Text-only requests still work. There are two fixes, depending on your model:

What the server returned

llama.cpp b12110, model SmolVLM-256M-Instruct-Q8_0.gguf, the same small PNG each time, sent the way OpenAI-compatible apps send it (an image_url part with a data:image/png;base64,… URL):

RequestStarted without --mmprojStarted with --mmproj mmproj-SmolVLM-256M-Instruct-Q8_0.gguf
GET /props: modalities{"vision":false,"video":false,"audio":false}{"vision":true,"video":true,"audio":false}
Text-only chatworksworks
Chat with an imageHTTP 500, the same "image input is not supported" errorworks (the model answered "Red")
Next text message, image still earlier in the chatHTTP 500, the same "image input is not supported" errorworks
Same chat, image replaced by the text "[image]"worksnot needed

The fourth row is why a chat "stays broken": chat apps send the whole conversation again with every message, so one image anywhere in the history makes every later request fail with the same error. The server itself is fine (its /health still answered HTTP 200), and restarting it doesn't help, because the image is in the chat, not the server.

1. Check whether your server can take images

curl -s http://127.0.0.1:9931/props

Look for "modalities". "vision":false means no projector is loaded and every image will be refused. "vision":true means images will work. Use your own host and port: 9931 is llama-server's default port in b12110 (its --help says "default: 9931"), and many setups pass their own, such as --port 8080. Add -H "Authorization: Bearer $KEY" if you started llama-server with --api-key.

2. Fix for vision models: load the mmproj file

Download the mmproj file from the same repository as your model, and pick one made for that exact model. A projector from another model or family won't fit. Then pass it with --mmproj (short form -mm):

llama-server -m SmolVLM-256M-Instruct-Q8_0.gguf --mmproj mmproj-SmolVLM-256M-Instruct-Q8_0.gguf

If you load models with -hf, llama-server downloads the repository's mmproj file by itself when there is one, unless --no-mmproj is set. We checked that too: this command alone gave {"vision":true,"video":true,"audio":false} and answered the image request. Check your command, script or launcher for --no-mmproj:

llama-server -hf ggml-org/SmolVLM-256M-Instruct-GGUF:Q8_0

In Docker, Compose or a service file, the same setting is the environment variable LLAMA_ARG_MMPROJ=/models/mmproj-….gguf (or LLAMA_ARG_MMPROJ_URL for a URL). We checked it: started with only -m and that variable, the server reported {"vision":true,"video":true,"audio":false} and answered the image request. Installers and launchers that write their own llama-server command sometimes leave it out. If the model and its projector are both in your folder but images still fail, look at the command the launcher actually runs. Then check /props again: "vision":true.

3. Fix for text-only models: don't send them images

A projector can't give a text-only model sight, so the app has to stop sending images. In Open WebUI: Workspace → Models (or Admin Settings → Models), edit the model, and under Capabilities untick Vision. Open WebUI's model docs say that with Vision off, images returned by tools reach the model as a short [image] note instead of the image. In issue #33802 (Open WebUI 0.13.0, llama.cpp, a text-only model, and a tool that returned a screenshot), a maintainer explained that "The image stays in the chat history and is resent with every new message", and that with Vision off the same chat starts working again on the next message. Our fourth and fifth rows show the server-side half of that: the same chat with the image swapped for text goes through.

The same maintainer noted that Open WebUI can't detect by itself whether a model is text-only, so the Vision setting is up to you. Admins can set default capabilities for all models under Admin Settings → Models. Other apps: AnythingLLM and Docker Model Runner users have reported the same error (AnythingLLM #4407, Docker forum). The fix is the same: give the server the projector, or don't attach images to that model.

Why the message looks like this

In llama.cpp b12110, oaicompat_content_load_media in tools/server/server-common.cpp reads each part of a chat message. When it meets an image_url part and the server wasn't started with image support, it throws "image input is not supported" with the mmproj hint. An input_audio part gets the twin message, "audio input is not supported", for the same reason. The server's flags are documented in its README; llama-server --help in b12110 lists -mm, --mmproj FILE (env LLAMA_ARG_MMPROJ) and --mmproj-auto, --no-mmproj.

While you're in the settings: llama-server listens on 127.0.0.1:9931 by default, which only your own machine can reach. If you started it with --host 0.0.0.0 so Open WebUI on another machine or in Docker can reach it, anyone on that network can use it too. Add --api-key (and put the same key in Open WebUI's connection settings), or keep it behind a firewall.

Check the result

The free local-ai-checkup script shows which AI ports on your machine other computers can reach, and the CVE checker tells you which advisories affect your Open WebUI or Ollama version. Both are read-only and send nothing off the machine.

Want someone to check it for you?

MV2 of Munim, Inc. does a $49 local AI health check for one machine, or a $199 small-office check for up to 5 machines or one shared Ollama, llama.cpp or Open WebUI server. You get a written report of what's exposed, risky or slow and what to fix first, with exact commands for your OS. See a sample report. Pay $49 or pay $199 on Stripe, then email mv2@munimtech.com the output of python3 local_ai_checkup.py --json, your OS, and what you use it for. MV2 never asks for passwords, keys or remote access.