[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / r / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


A guy just fine-tuned Qwen3.6-35B-A3B to imitate Claude Opus 4.7's reasoning style — basically he took Opus's chain-of-thought traces and used them as training data, so the model now "thinks" the same way Opus does, wrapped in <think>...</think> tags.
The wild part is the efficiency. It's a 35B MoE model but only ~3B parameters are active per token, which means you can actually run this thing on a single A100 or H100. No cluster needed.
And it's fully open. Apache 2.0. Weights are public. Training dataset is public.
This is essentially reasoning distillation — taking what makes a frontier model good at thinking and compressing it into something accessible.
Not saying it matches Opus on benchmarks. It probably doesn't. But the trajectory is clear — the gap between "I can afford this" and "this is actually good" keeps shrinking.
https://huggingface.co/lordx64/Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled
>>
>>108638991
>A guy
you
>imitate Claude Opus
It's been shown that these small model opus finetunes perform worse than their base model on most benchmarks and subjective experience.
>H100
You think you need one of those to run a 35B-A3B?
>Three em-dashes
>>
>>108639192
>You think you need one of those to run a 35B-A3B?
i run that in a 1070 with 32 gb of ddr3 ram
>>
>>108638991
>35b parameters
That's Mac Mini tier, not A100 lol

Also why are you selling this as something noteworthy when people do distills like that all the time?
>>
>>108638991
use better model to write your post next time and tell it to remove all slop and em-dashes from it
>>
>>108638991
I know what a meter is
I know "para" means "around"
Whats a parameter?
>>
why is he listing it using the same thinking tags Qwen always used to begin with as some kind of relevant feature?
>>
>>108638991
At least clear the em dashes before spamming your posts (gens) around.
>>
>>108638991
Pajeet
>>
>>108638991
You literally used AI to write this. What a retard.
>>
when will 4chan get emojis? then all these bots can post using them and they'll be even easier to identify. they can already use emdashes...
>>
>>108638991
4.7 is worse than 4.6
>>
>>108639192
fpbp
op clueless.
>>
>>108640341
it's an imperial measurement for how many paratroopers it takes to solve a problem.
>>
>>108640645
you can use emojis on /sci/ technically, or in filenames
>>
>>108638991
>Just shipped an...
what is it about xitternigs that makes them talk like that?
>>
>>108639192
>three emm dashes

>And it's fully open. Apache 2.0. Weights are public. Training dataset is public.
This line should have tipped me off. I need to get better.
>>
>>108638991
Buy an ad Rajesh
>>
>>108638991
Claude doesn't send actual reasoning traces.
They are thinking summaries designed to prevent exactly this kind of thing.
>>
>>108644474
it's already been seen on the claude leak



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.