Video wird geladen...
Video konnte nicht geladen werden
Deepseek vs Fable 5.2 Both were given the same prompt, only difference was harness. Both came out great! If I had to guess Fable 5.2 is trillions of parameters larger than Deepseek, and MUCH more expensive. With all of that said Deepseek is king here.
36,215 Aufrufe • vor 8 Tagen •via X (Twitter)
38 Kommentare

Deepseek is doing well

Yeah I have another example coming a bit later, Deepseek can't be counted out yet!

Is it wrong to say I'm MUCH MUCH more hyped for qwen 4? Like Fable is great, but local models... we need more.

No I'm also hyped for Qwen 4! I want to test it more but I'm trying to save tokens lmao

I just want to run qwen 4 27B and qwen 4 flash next locally :(

我很震惊,为什么你发布的项目效果都如此好,就像官方宣传自己模型效果时候的发布片。美术有用到模型生成或者图片生成模型吗?

haha appreciate that! I don't use any reference images or anything, it's a lot of tweaking I use other quick models (mostly Deepseek now) to sketch stuff out basically and dial in a prompt then have other models test it. For everything I post theres at least 5-10 failures lol.

@yngdl1789479 this was over API?

@yngdl1789479 Yea

It's like what ? 100x cheaper also

That's a pretty well made ship by deepseek!

the cheap one winning is the best part of this whole story

fable water is good,Deepseek architecture and atomosphere is good

DeepSeek Flash:现在都轮到我和最新的fable系列对打了吗🤡🤡🤡

I'm pretty sure I played this game before

DeepSeek is amazing. I find PI + DeepSeek works well in CLI environment.

same prompt but two different harnesses means two variables changed, so deepseek winning might just be the deepseek harness handling this task better. i run a fleet daily and the same model in two harnesses swings wider than two models in one harness. retries and error reading are doing most of the work.

Totally on point, tackling this specific friction point makes a huge difference in practice.

永远可以信赖Deepseek

Fable5.2 这个水面很真实,就跟真的一样了

DeepSeek wins!

Share the prompt plase

哇哦,你一定写了一个非常优秀的提示词。

Fable is cooking ☠️

deepseek的表现也不差啊

Prompt

harness being the only difference is the interesting bit did you log tool errors? that's usually where the gap shows up

Same prompt, different harness. You compared harnesses, not models.

is it a full three.js implementation or is there more?

同样的图,DeepSeek没输

This is sick man

prompt?

The anthropic models are still based on mythos architecture huge dense models. Dense model sure keep reasoning accurate but at humongous infra costs and power consumption. So if looking at the reasoning/energy consumption ratio (and so cost) chinese models definitely are on top.

Yeah Fable 5.2 is better but 200X better? Because it would cost that many times in this test LOL

I think we are getting back to the point of the models being only as good as the prompts . Always better prompts will gov eyou better builds

Look what I made! Look what I made!

run it again with the setups swapped, deepseek in fable's and fable in deepseek's

fable win in all ways
