Thinking?
I am looking for a thinking capable model. Qwen3.5 27B is quite good but reasoning is sometimes looping, generating errors on coding tool size.
Is this model thinking?
Is looping problem solved?
Interested in the benchmark values when they become available.
This model has Thinking. I use the Architect name for thinking models and Engineer for Instruct.
Originally the model has a looping issue. I modified the template in Architect/Engineer to support XML tools, and a few other minor fixes. This has helped lower the inference noise and confusion in the model, that seems to like XML tools better. Additionally, the Claude traces have helped stabilize the inference, and reduce the looping. It will still happen in some conditions, but is more stable compared to baseline.
The metrics of the parent model are applying here, as the only difference is in the template, so Engineer metrics would be identical
Qwen3.5-27B-Claude-4.6-OS-INSTRUCT-
mxfp8 0.675,0.827,0.900,0.750,0.496,0.800,0.721
qx86-hi 0.667,0.824,0.902,0.752,0.502,0.791,0.725
Qwen3.5-27B-Claude-4.6-OS-Auto-Variable-Thinking
mxfp8 0.485,0.566,0.875,0.746,0.408,0.789,0.730
So if you add to Architect this line in the jinja, you turn it into an Engineer with the metrics from Instruct
{%- set enable_thinking = false %}
OK, so I now understand your terminology.
Thank you for the values. May be the model card could be updated (replace "...coming soon" by the values).
As soon as I have them, tests on this model take a long time :)