Qwen

54.0 Dumber than usual experience score ▲ 0.7 Confidence: Medium

813 opinions tracked this week · 7-day rolling window

30-day experience score

Version breakdown

Only comments that explicitly name a version contribute to its score.

Qwen3.6 27B	53.2−0.9	n=288	Confidence: Medium
Qwen3.6 35B A3B	65.6+11.6	n=45	Confidence: Medium
Qwen3.5-122B-A10B	72.2+18.1	n=27	Confidence: Medium
Qwen3.5-9B	64.1+10.0	n=8	Confidence: Low
Qwen3 Coder Next	61.4+7.4	n=6	Confidence: Low

424 generic family mentions are included in the family score, not any version score.

By experience dimension

General text	57	n=228
Coding	58	n=163
Reasoning	60	n=154
Image / vision	66	n=11
Roleplay / creative	not enough data	n=3
Safety & refusals	not enough data	n=1
Speed & latency	65	n=217
Local deploy	66	n=36

Top positive drivers

speed quality Speed & latency ×109
strong reasoning Reasoning ×63
problem solved General text ×27
competitor flight General text ×20
problem solved Coding ×19

Top negative drivers

performance issue Speed & latency ×25
speed quality Speed & latency ×21
competitor flight General text ×15
tool failure General text ×15
performance issue Coding ×14

What users are saying

🟢 “I'm running it with preserve thinking now, and it is not doing stupid things, so this appears to be a huge improvement”
via reddit · text_general · en

🟢 “It's truly impressive what some people have done with comparatively tiny models like Qwen3.6-27B, just by using clever prompting to have it break up tasks and make it not rely on its built-in knowledge as much.”
via reddit · reasoning · en

🔴 “I was shocked at how stupid qwen 3.5 122b was at 4 bit compared to qwen 3.6 27b at 8 bit. Night and day.”
via reddit · text_general · en

🔴 “Yeah, without preserve thinking it forgets what it concluded earlier and usually goes down the same paths again. It's a bit odd it's not on by default.”
via reddit · text_general · en