Qwen 3.7 Max
Posted by Sicarius_The_First@reddit | LocalLLaMA | View on Reddit | 23 comments
Qwen 3.7 looks pretty impressive.
I think we've reached to the point if Chinese labs catching up with the western frontier labs.
The question is, will the weights be available for download?
natermer@reddit
Anything named "Max" is probably something far too big to be ran on anything I will have access to locally.
wgaca2@reddit
Wait until the Pro Max drops
juaps@reddit
Very local, wait until Ultra
Sicarius_The_First@reddit (OP)
tbh qwen 80b next is very much locally runnable. 120b too.
hell, i even run minimax 230b on my LAPTOP at acceptable speeds. (not a mac).
mouseofcatofschrodi@reddit
that seems to be a good laptop... yisus
Sicarius_The_First@reddit (OP)
nvidia 5080 and soldered LPDDR5X 64gb.
expensive, but not macbook expensive.
wgaca2@reddit
where do you even get your hands on ram chips to upgrade it?
Possible-Pirate9097@reddit
Very contrained, wait until heretic.
hackiv@reddit
How? 3.6 just released, how is this possible? It basically beats Opus 4.6
Sicarius_The_First@reddit (OP)
It beats it on specific benchmarks, frontier models like Opus are still better in edge cases and multi lingual, but these are impressive scores nonetheless.
FullOf_Bad_Ideas@reddit
It's noteworthy that it also outputs about 30% less reasoning tokens than Opus 4.6 in the suite of benchmarks ran by ArtificialAnalysis, while having higher composite scores.
I hope this will translate to solid open weight models and practical usage.
Sicarius_The_First@reddit (OP)
This is very nice for larger moes, those 30% less reasoning tokens matter a lot at scale, with larger models. Some of us use RAM offload to run these big bois, so this translates roughly to 30% more speed in a way hehe.
VoiceApprehensive893@reddit
amazing model,had a really positive experience using it to vibe code some small ~1000 line apps
Sicarius_The_First@reddit (OP)
Yeah the long reasoning is a quirk worth addressing.
It's somewhat 'forgivable' on smaller models (like qwen 35b), but shorter reasoning is very much appreciated on larger models, like this one.
ridablellama@reddit
wowie those bench scores are nuts. has anyone tried it out yet?
Better-Struggle9958@reddit
why MAX?
VoiceApprehensive893@reddit
qwen 3.x max was always a big ass qwen model, bigger than plus(qwen3.5 and qwen3 plus were the biggest open models)
dryadofelysium@reddit
> weights be available for download
they never release Max weights
Dany0@reddit
Man holy shit how are they delivering like this despite losing their best talent wtf š
The day Q3.7 open weights drop it's gonna be mayhem here
DataGOGO@reddit
They are distilling Opus and OpenAIĀ
Dany0@reddit
Bhahahahahahahahahahahhahahahahahhahahahaa hahhahhHAHHHAHAHAHAHAHAHHAHAHAHAHAHAHAHHAHAHAHAHAHAHHAHAHAHA
Get a load of this guy
Sicarius_The_First@reddit (OP)
I really hope this will be open weights, not even because these benchmark scores look impressive (they are), but because qwen 3.6 was amazing.
Dany0@reddit
Max will never be open weights unless it leaks. They are a public company so their promise makes them legally liable to their investors. Your hope is misplaced