Thinking Nightmare
This model is thinking too much, much more than Qwen3.8 27b (reasoning effort = xhigh).
I was unable to get an answer to this simple question using a reasoning budget of 8k and reasoning effort = medium | high | xhigh:
The following text contains a structural ambiguity.
Identify it, explain the two possible interpretations, and rewrite the text in two separate versions,
each clearly disambiguating one of the two readings: "The professor called the student into his office because he was late".
I was only able to get an answer by disabling reasoning or setting reasoning effort = low
Even worse I was unable to get an answer by increasing the reasoning budget to 16k
I think that to get this model usable it will be necessary to set reasoning effort = low
P.S.
Till will be written the small sentence at the bottom: "LLMs can make mistakes. Double-check response", the LLMs will be unable to replace humans, no matter what AI fanatics say...
What quant? I’ve had this problem with unsloth models recently, I think it’s their new quantization method somehow making these models never stop thinking
Limiting models to reasoning budget doesn't work with qwen models. How many times does this have to be repeated?
Even worse I was unable to get an answer by increasing the reasoning budget to 16k
I think that to get this model usable it will be necessary to set reasoning effort = low
P.S.
Till will be written the small sentence at the bottom: "LLMs can make mistakes. Double-check response", the LLMs will be unable to replace humans, no matter what AI fanatics say...
For Unsloth we're working on improving auto compaction. May I ask which quantization were you using?
What quant? I’ve had this problem with unsloth models recently, I think it’s their new quantization method somehow making these models never stop thinking
This quant is using V2.5 actually which was from 3months ago or so
I'm using: Qwen3.8-Flash-Next-GGUF\UD-IQ3_XXS
i can't replicate. works fine for me (very short, reasonable thinking). i ran this prompt, ud q4 xs, xhigh:
That's interesting. I performed the tests above using as client Unsloth Studio Desktop (last build for Windows), and as showed in my previous screenshot I was unable to get an answer with reasoning effor equal to medium or above. But changing the client it works, for example by using the WebUI included in llama-server I get this answer (reasoning effort = medium)
to provide the answer were used only 2141 tokens (with "xhigh" were used 2575 tokens).
So it seems that the problem is limited to Unsloth Studio Desktop, probably due to the context sent to LLM before sending the question.
Nice, I will switch to "medium".
--reasoning-budget 80000 --reasoning-budget-message "Reasoning budget exhausted — answering now."
try running with this args. It worked really well on Qwen 3.8 27b when the model thinking went above the limits of its thinking 80k tokens it would start writing the response immidietly after that.
The fastest way to see if it works : --reasoning-budget 100 --reasoning-budget-message "Reasoning budget exhausted — answering now."
and just ask any question - it will start thinking for just a few seconds and than give the response. Shall work even on xhigh
--reasoning-budget 80000 --reasoning-budget-message "Reasoning budget exhausted — answering now."
try running with this args. It worked really well on Qwen 3.8 27b when the model thinking went above the limits of its thinking 80k tokens it would start writing the response immidietly after that.
The fastest way to see if it works : --reasoning-budget 100 --reasoning-budget-message "Reasoning budget exhausted — answering now."
and just ask any question - it will start thinking for just a few seconds and than give the response. Shall work even on xhigh
This is a terrible solution to stop looping. You’re basically saying stop looping when you hit 80k tokens. If you get 60 t/s generation, that’s 22 min of thinking. If you get 22 t/s, it’s 1 hour of thinking.
What ambiguity? The professor was in the switchboard office of his hovercraft and the student was at a payphone in the Matrix, the "late" part is a red herring.