This is Deepseek V4.1 Flash, are they doing a stealth upgrade behind the scenes? 1 prompt, 0 revisions, no harness straight to the API. This is the kind of output I expect and hope for from Anth/OpenAi models, NOT from an open weight Flash model. This is insane!

looks great you should bring that back

Oh man I spent A LOT of time with this one on the SNES actually

No that means your oneshot from the API was routed to fable 5.1 behind the scenes and your prompt was used for training the next checkpoint

😂

What type of prompt did you do for this

If you're able to run local version from open weights and compare on the same prompts that'd be really interesting.

I'm rerunning prompts and will post some comparison videos I have all my stuff saved these are just DS4 demos 😂

Sure seems like they are doing behind the scenes work

Maybe you're routed to some mysterious model 🧐

This is definitely not the same model as before!

I've noticed quite a bit of improvement, but this is even further.

Are you running full weights or is this quantized any?

This is their official API

Amazing. DS41F is my daily driver so any improvements are very welcome

ya, that's pretty insane, i want to run this model!

yes deepseek's flash is a beast with reasoning. it's almost free you can run 50 concurrently with their api key

DeepSeek improving is utmost important to civilization, but at the same time geopolitical location is troubling AND Chinese are so often copycats, while we’re sending them all our prompts. I think we should make them their own sovereignty, land power and all. Or help them distribute their compute across the land so it’s both not targetable, AND if the people stop liking it, they can vote it off by making a personal decision to power it.

If Flash is this good, what are the closed models even selling

at this point open weights ARE the frontier

Hmm, interesting! Do you have v4.1 Flash running locally as well? Even if it is a quant or from another non-DeepSeek API it would be interesting to run it and see the difference.

No I only have my previous runs like 500+ that I'm going to go comparison test.

i didn't use 4.1 Flash for a week or two and thought the same thing after using it last night & this morning definitely better than it was last month

They should be conducting a random canary test; the routing is via ds v4.1pro.

Big if true, will they update the weights then as well

I did notice weird with deepseek. I just got a new broswer and in helium in the deepseek site it says search and the title is "what can i do for you" and the other is called smart searcg and searchs different and says "Where would you like to begin?" Which is kinda weird

this is pretty, i like it :)

@JoeGuglielmucci V4.5? Would be amazing

WOW!! This is next level. What a concept as well Which API is this from Loktar? OpenROuter

Official API

Curious whether it's still that good in a few days. Silent bumps are great until the model you tested isn't the one answering anymore.

That's incredible. Three.js?

Yeah, the internet runs on it now lol

I thought it got quicker and a little better a few days ago. But didn't test extensively.

看起来效果很棒

Wait, local weights?

What do you mean “no harness, straight to the API?”

Blender i take it?

No just three.js, single prompt

Oh wow. Very very nice.

deepseek patch notes: "bug fixes and performance improvements" 😭

the flash-model tier quietly becoming the frontier is the story of the year. if the cheap open models do 90% of the job, the $200 subs start looking like luxury goods

Have anyone ever seen something actually useful created with AI?

被路由到 fable5.5 flash 了！

WHAT THE FUCK IS THATTTT