GLM-5 is text-only as a model input modality: it takes text prompts (and structured tool definitions) and produces text outputs. If your product requires direct image understanding—like interpreting screenshots, diagrams, or PDFs as images—you should treat GLM-5 as the “language and reasoning” layer, and pair it with a dedicated vision model or an OCR/vision pipeline that converts visuals into text before GLM-5 sees them. In other words, GLM-5 won’t natively “look at” images; it works from text that you provide, whether that text came from users, documents, code, or extracted visual content.
In implementation terms, “text-only” does not mean “limited.” Many multimodal experiences still use a text LLM at the core, with a preprocessing step that turns non-text into structured text. A common production pattern for screenshot-heavy support tickets is: run OCR on the screenshot → extract UI strings and error messages → normalize into a structured payload → feed the payload to GLM-5. For PDFs, you can parse text directly when possible, and fall back to OCR for scanned pages. For diagrams, you can store a human-written “diagram description” alongside the image asset and feed that into the model. This design is usually easier to maintain than assuming a single model can do everything, and it gives you better control over what the model is allowed to use. It also makes your system auditable because you can log exactly what extracted text was used as input.
Text-only also pairs cleanly with retrieval. If your application is a documentation assistant, you can store your docs in a vector database such as Milvus or managed Zilliz Cloud retrieve only the relevant text chunks, and pass them to GLM-5. If your content includes images (architecture diagrams, screenshots, UI flows), store an “alt-text style” description as a sibling chunk in the same collection, with shared metadata like doc_id, section, and version. At query time, you retrieve both the surrounding text and the diagram description, then instruct GLM-5 to answer using those retrieved passages only. That way, even though GLM-5 is text-only, it can still answer questions “about images” as long as the image information is represented in text. This approach keeps latency predictable and avoids bloating the prompt with irrelevant material.