never-ending reasoning
Hi, I was using your model in a research project to evaluate some multi-hop chemistry-related questions, and I encountered a few issues. I was accessing it through OpenRouter (via the ZAI provider).
First, the model appears to exceed the maximum token limit. For example, I set the limit to 8,192 tokens, but it actually generated around 10k tokens.
The second issue is more severe: in many cases, the model produces no meaningful output at all. Instead, it seems to get stuck in an endless reasoning loop; basically, a never-ending internal conversation with itself. When I checked the reasoning stack, it was clearly stuck mid-thought.
To confirm, I tested one of the same inputs directly through the ZAI chat interface, and it ultimately generated roughly 100k tokens without providing an answer. This appears to be a consistent pattern, an infinite reasoning cycle with no final response.
At first, I thought it might be my fault, maybe I wasn’t setting parameters like top_k, top_p, or temperature correctly, but I observed the exact same behavior even on the official https://chat.z.ai/.
Query: Which type of chemical process, responsible for producing hundreds of millions of tons of chemicals each year, is catalyzed by the inorganic compound sometimes represented as CoIICoIII2O4 that has been engineered through modifications such as cation doping and oxygen vacancies for enhanced oxygen evolution reaction performance?
This is the reasoning stack from Open Router:
https://pastebin.com/isU1KZEG
And this is the chat:
https://chat.z.ai/s/3616437b-4349-46ec-b66d-26fa867e2a3e
- The thinking part is not calculated, your 10K token includes the output of the thinking part.
- Try setting the temperature to 1.0, but if you've already tried it on z.ai, it's most likely not an issue with your call, we have already recorded this issue.
im having the exact same problem, thinking just fits the whole output and gives no answer whatsoever