GLM 5.3 Flash frequently loops/repeats in chain-of-thought since becoming unmetered

[myplxdm]

What happened

Summary

Since GLM 5.3 Flash was switched from a limited/paid tier to an unmetered, always-available default model, it has become noticeably more prone to looping — the model repeats the same reasoning step, question, or conclusion over and over instead of making progress, especially on more complex or multi-step tasks. This didn't happen nearly as often when the model was still rate-limited.

Expected Behavior

The model should either make steady progress through the task, or — if it genuinely can't complete it — stop and say so, rather than repeating the same step indefinitely.

Actual Behavior

On complex tasks, the model frequently gets stuck repeating the same reasoning/question in a loop, burning turns without making progress. This has to be manually interrupted.

Possible Contributing Factors

• The README notes that models "may still serve from a quantized (Q8_0) build." Now that GLM 5.3 Flash carries unlimited free traffic as the unmetered default, it's worth checking whether quantization is being applied more aggressively (or more often, e.g. under peak load) than before.
• There's currently no way to configure the reasoning/thinking effort level for GLM 5.3 Flash in Freebuff, even though Z.ai's own docs reference multiple reasoning tiers for this model. A fixed or lower effort setting could make looping more likely on harder tasks.
• It's unclear whether the unmetered tier is served through the same infrastructure/precision as the previous rate-limited tier, or a cheaper backend to sustain unlimited free usage.

Suggested Improvements

• Expose a way to adjust reasoning effort for GLM 5.3 Flash.
• Add a loop/repetition detector in the agent loop that interrupts and asks the model to summarize progress instead of continuing to repeat.
• Clarify whether the unmetered tier runs at different precision/quantization than before, so users know what tradeoff they're accepting.

Steps to reproduce

Steps to Reproduce

This is a general pattern observed across multiple sessions rather than a single isolated reproduction, since the exact prompts vary by project. The conditions under which it tends to occur:

1. Use GLM 5.3 Flash (default/unmetered) in an agentic session involving multiple steps — e.g. a task requiring several rounds of tool calls, file edits, or iterative reasoning.
2. Let the task run past roughly 10-20+ turns / tool calls (exact threshold varies).
3. On a subset of these longer sessions, the model gets stuck repeating the same reasoning step, question, or conclusion instead of progressing, and has to be manually interrupted.

This was rare before GLM 5.3 Flash became the unmetered default, and has become noticeably more frequent since. Happy to share a specific transcript if it would help, but wanted to flag the general trend first since it doesn't seem tied to any particular prompt or project.

Where does this happen?

CLI (terminal client)

Operating system

macOS

Version

0.0.162

Model

glm5.3 flash

Logs or screenshots