I saw someone say recently something to the effect of: “that man is a working dog. if you don’t give him a job, he’ll tear up the furniture.” Qwen3.5 is a working dog.
I’ve been working with this model a lot recently. I’ve baked three dozen custom quantizations. I’ve used three different execution backends. Of everything I’ve learned I can at least report the following.
These models absolutely hate having no context. They are retrieval hounds. They want to know their objectives going into things. Your system prompt is 14 whole tokens? You’re going to have a bad time. 27B doesn’t even become remotely useful sub 3K tokens going into it. It will think itself raw getting to 5K tokens just to understand what it’s doing.
And I should note: this makes a lot of sense. These models, in my estimation, were trained agentic-first. Agent models want to know their environment. What tools they have. Their modality (architect, code, reviewer, etc). With no system prompt or prefill they stumble around aimlessly until they have something to grab onto. In my opinion: this is a good thing. Alibaba has bred the working dog of the open weights model. It is not a lap pet.
As you evaluate this model family, please keep in mind that the Qwen team has, very deliberately, created a model that wants a job. It does not want to hear “hi.” It wants to hear what you actually need done.
Also the 35B MoE is kinda trash. That isn’t poetic, it’s just true.
[link] [comments]
