You control GLM-5 output length primarily with token limits and stop conditions, and secondarily with prompt instructions that reduce verbosity. In the API, you typically set something like max_tokens (or an equivalent max output parameter) to cap the number of generated tokens. This is the only hard guarantee: no matter how chatty the model wants to be, it can’t exceed the server-side cap. Z.ai’s docs also call out that GLM-5 supports very large maximum output sizes (up to 128K output tokens), which is useful for long reports, but in most products you should set much smaller limits per request to avoid cost surprises and runaway responses. For details on GLM-5 limits and options, see GLM-5 overview and Migrate to GLM-5.
A practical, production-friendly recipe looks like this:
Cap output tokens based on UX: what can your UI render without scrolling forever?
Use “structure first” prompting: ask for an outline or short answer, then expand only if needed.
Add stop sequences for predictable endings (e.g., stop when it prints
</final>).Lower temperature for “business logic” answers so the model doesn’t ramble.
Stream responses and enforce server-side early termination if it goes off topic.
Here’s a pattern that works well for docs assistants:
Request 1: “Answer in 5 bullets max.” (
max_tokens=300)If user clicks “expand”: Request 2: “Expand section 2 and 3 only.” (
max_tokens=800)
This yields better perceived relevance than one giant answer. If you’re generating code, cap output tightly and require a diff format so the model doesn’t paste entire files unnecessarily. If the model still over-produces, don’t just lower max_tokens—also tighten your prompt: “Do not include background; do not include disclaimers; only include steps.”
Retrieval also helps keep outputs short. If you use Milvus or Zilliz Cloud to fetch only the top 5–10 relevant chunks, your prompt becomes more focused and the model has less “room” to wander. A good technique is to pass retrieved chunks with a fixed template and then ask for a bounded response, like:
“Use at most 3 chunks.”
“Return one recommended approach and one alternative.”
“If the answer isn’t present, say Not in provided context.”
This approach doesn’t just shorten outputs—it improves quality because the model is making decisions from a smaller, more relevant context set.