#Qwen3.7 Plus/Max
sand sage
glad timber
"recreate the classic arcade game Shinobi in html'
https://codepen.io/Madvulcan/pen/QwGpWWZ
1458 lines of code in single shot. The player can't jump high enough to reach platforms, otherwise flawless and can be iterated away. Nice single shot result
glad timber
Yeah, fixed with one prompt. And it didn't rewrite the whole thing from scratch like some models do. https://codepen.io/Madvulcan/pen/ZYBeYWB
tawny viper
stable night
wait WHAT?!??!?! HOW DID I NOT SEE THIS COMING!?!?
there is not even a blog post about the model!
the plus model seems good at generating.... SVG mockups.
look at this! this looks... almost fine.
like- this isnt html.
opal talon
In case anyone was curious...
stable night
#6 in text, huh?... interesting.
tough plank
i don't trust qwen with the bench, but they certainly are releasing less and less open weight models as of recently
stable night
i don't trust qwen with the bench, but they certainly are releasing less and les...
why trust anyone with benches?....
peeps tested it, and it appears less.... "Pretty putpzr" tuned so far. 3D stuff is somewhat meh, more plain colors, not sure.
tough plank
why trust anyone with benches?....
peeps tested it, and it appears less.... "Pr...
no i mean they historically benchmaxx, but i would admit as of recently qwen finally has some substance and can do real shit for it is size.
last rose
bench results are VERY promising so far
I expect the ranking to drop as more results come in though
a notable result is that it dethroned mimo pro in non-hallucination rate
it's also tied with it (and grok 4.1 lol) for most faithful instruction following
(it's actually not that great in omniscience, though still upper tier)
(it's 4th in AA-Intelligence and Coding, just after the 3 SOTAs)
the performance on its own wouldn't be that exciting, but combined with the honesty/instruction following it's suddenly quite attractive
bright palm
last rose
lmao
I was JUST about to post WHERE IS QWEN OR
I like it already
oh god please do not ask it math/geometry questions on high or xhigh though
it'll drain your wallet on something that other models figure out in 5k toks
it's not wrong or anything it's just totally unnecessary levels of rumination
mortal arrow
Input $2.5, Explicit Cache Creation $3.125/M
hate this so much
mortal arrow
interestingly enough, prices are ~0.66x in non-singapore regions in alibaba cloud
jaunty nova
I've been trying to use Qwen 3.6 Flash but keep getting errors - wonder if it has anything to do with this recent launch
floral fox
$2.5/$7.5 💀
I haven't seen Chinese inflation this bad since before I started paying for Fansly.
floral fox
Classic Qwen
last rose
I think it's particularly bad at "visualizing" geometry, or trusting its visualization rather
probably why it does quite good on math benches but ouch
earnest gate
omg thank you alibaba 😭 🙏
twilit eagle
Is this owl alpha?
last rose
I must say, I'm not super amazed at the hallucination rate in practice, despite the very promising bench results.
Maybe it's because I mostly trip it with pop culture questions and it reserves its accuracy for "serious" tasks like STEM or coding.
This in contrast with other reputed honest models that come flat out and admit "sorry, this is probably too obscure for me to answer"
vernal pasture
I must say, I'm not super amazed at the hallucination rate in practice, despite ...
afaik those tests don't measure intrinsic knowledge hallucination rate, they measure it over a given document
also until they fix this cache this literally more expensive than opus
brave token
wants me to drive to car wash 👍
split sphinx
So, deepseek still is the winner, specially with the pricing
jaunty nova
So, deepseek still is the winner, specially with the pricing
it'll be very difficult for Google and Qwen to match the pricing of DeepSeek. It's so so so affordable it's insane.
neat basalt
also until they fix this cache this literally more expensive than opus
dujde what the FUCK is wrong with them
jaunty nova
has anyone been using Qwen3.7 Max recently - is it any good? I'm told it's primarily meant for autonomous agent work and it's abit meh for simple chat workflows primarily because of how much reasoning tokens it generates.
floral fox
It benchmarks like an absolute beast
But the history of Qwen is also benchmaxxing, terrible EQ, and sucking down tokens like there's no tomorrow.
stable night
It benchmarks like an absolute beast
previous qwens yes, but ive seen peeps saying its preddy good.
tough plank
qwen 3.7 max is certainly an interesting model, even by closed bench standards. will wait for their open weight models.
split sphinx
They reduce the price of this model
That's a W, hope they keep the price at that rate or even make it more cheaper like deepseek
mortal arrow
@split sphinx it's a one month long promotion: https://www.facebook.com/alibabacloud/photos/qwen37-max-now-live-on-model-studio-50-off-limited-time-offer-built-for-the-agen/1431601325678622/
but if the dominos keep falling (e.g. xiaomi) then surely it's in their interest to adjust price
stable night
https://qwen.ai/blog?id=qwen3.7-plus
this is old news by now >v<
sand sage
this is old news by now >v<
They announced it an hour ago on xitter
stable night
They announced it an hour ago on xitter
the model has been out for a while i believe....
oh! nope im wrong, that was the 3.7 pereview. my bad---
sand sage
the model has been out for a while i believe....
oh! nope im wrong, that was the...
We had plus preview?
stable night
We had plus preview?
yes, on qwen chat for a while by now.
earnest gate
good ad for opus
surreal olive
models have jst been regressing
neat basalt
what im most interested in is the price/performance pareto frontier and the mimo v2.5 price cut was goated but i think alibaba dgaf abt thje pricing no more