GLM model doesn't understand how it receives images

AC3

Adding the below image to a prompt appears to work and shows visible to me on the UI, but the AI models state they cannot access it. They mostly hallucinate but do provide some context in 'thinking'. Seems to only affect this specific image, other images work.

GLM (The user's query is "what is wrong with this dilbert".
The user provided a "system_provided_context" which contains a block of text that looks like a "thought process" or "internal monologue" from an AI model trying to generate a title for an image. This text is confusing and seems to be a hallucination or a copy-paste error from a previous interaction or a different system. It mentions "Daniel Kibblesmith", "cartoon version of Trump and Adams", "Three Amigos meme", "Scott Adams", "Axl", "Jesse", "Sanch", "Stewie", "AI", etc.

Crucially, the user did not attach any documents or images in the system_provided_context. The context explicitly says: <documents>The user did not attach documents, so there are no documents to use tools with.</documents>.

However, the "thought process" text inside the context describes an image: "The user provided a detailed image with a tweet from Daniel Kibblesmith reacting to a cartoon version of Trump and Adams, along with a meme-style image from The White House."

This is contradictory. The system says "no documents", but the internal text describes an image.
The user's prompt is "what is wrong with this dilbert". This implies there is an image of a Dilbert comic or strip attached, or the user is referring to something specific.

Since I cannot see any image, and the system context explicitly states no documents were attached, I must rely on the text provided in the system_provided_context to infer what the user might be talking about, or ask for clarification.
However, the "thought process" text seems to be the AI's previous attempt to answer a different question (about titling an image). It is messy and incoherent.)

K2 (Kimi K2
I still can’t see any image attached to your last message; the system ![log] shows “The user did not attach documents”. Until a working link or an actual file appears, I can’t identify what visual glitch, continuity error, or joke you’re referring to. If you’re having trouble uploading, try a direct paste (e.g. imgur, catbox, or an image-hosting service that doesn’t require a login). Once the strip is visible I’ll give a precise answer.)

I expect the llm to be able to view all dilbert related material when asked to.

AC3

Re-reading GLM's output - it can clearly process the image but is confused as to its type. Perhaps the file is being sent to the model in a way it's not understanding.

Thibaultmol

What's happening here is that GLM 4.7 doesn't actively natively support image input. So it's using a different model behind the scenes to get context of the image and then provides that to GLM. That's why it's confused

Thibaultmol

Its mainly an issue it seems with GLM specifically not understand that it can't receiving images, only descriptions of images based on what you asked it.
https://kagi.com/assistant/ba7e3da0-5ceb-4769-b667-69ffdaee4351

yiwei-1

For queries like this that require understanding the visual elements in images (rather than just extracting text), you’ll want to try models with native vision support such as Gemini 3 Flash: https://kagi.com/assistant/348a0cae-9e93-42bb-9dcd-f8463dc97bab

By contrast, models like GLM 4.7 that don’t have native vision support can only rely on an image caption generated by a separate captioning model, which often misses many visual details.