Rendered at 23:13:24 GMT+0000 (Coordinated Universal Time) with Cloudflare Workers.
vblanco 2 days ago [-]
Insane scores for a model of this size. But it does seem to be a rather insane over-thinker with the biggest token use of any model, which combined with 256k context size it means it wont do much before filling it context
data-ottawa 1 days ago [-]
It’s definitely a heavy thinker, like most Qwens.
I see strings like “write, now.” In the thinning traces then it goes on to think for a lot longer, so it’s kind of weird.
I haven’t figured out hope to use this effectively yet on my strix halo.
cyanydeez 13 hours ago [-]
With llamacpp --reasoning-budget truncates thinking and makes it a usefulagain.
Its a per client request setting also.
Using qwen 3.8 27b, in open code, its unnoticeable.
So more a issue for server/harness design than model.
We don't actually know how much thinking GPT and Opus do, the labs won't show us anymore. And they certainly take their time before starting to answer.
kasperni 2 days ago [-]
Maybe a bit of context for this post? Some people have a life outside of AI.
toshinoriyagi 1 days ago [-]
This model is a preview of Qwen's upcoming Qwen4 architecture. It is a 125B-A6B MoE model, meaning it has 125B total parameters with 6B active at a time, but it also has a 51B parameter engram with it. The engram is basically a lookup table for tokens to my understanding. It allows the model to have access to a much larger amount of info if utilized well.
They said the model is intentionally under-trained since it is mainly for R&D purposes of proving the new architecture. Many people are excited for models in this range as they are a step above the common ~27B models, while not requiring exorbitant sums of money to run like much larger models.
glimshe 2 days ago [-]
I got downvoted yesterday for complaining about the name/brand confusion from all these Chinese models with similar names all claiming they are the best. While I'm an AI enthusiast, it's being hard to keep track.
I see strings like “write, now.” In the thinning traces then it goes on to think for a lot longer, so it’s kind of weird.
I haven’t figured out hope to use this effectively yet on my strix halo.
Its a per client request setting also.
Using qwen 3.8 27b, in open code, its unnoticeable.
So more a issue for server/harness design than model.
Also, there might be a real bug in the models parameters, like this for 3.8-27b: https://huggingface.co/grimoni/Qwen3.8-27B-SSMFIX-UD-Q4_K_XL...
They said the model is intentionally under-trained since it is mainly for R&D purposes of proving the new architecture. Many people are excited for models in this range as they are a step above the common ~27B models, while not requiring exorbitant sums of money to run like much larger models.