Ligtr

...

[Ligtr @kire0n]

@ArtificialAnlys @deepseek_ai Redo it, it is significantly better than flash, you were using flash most likely as the deployment had issues😢😢

[Ligtr @kire0n]

@aisearchio Don't test v4 pro yet, something is up with it's backend, it's performance is inconsistent similar to preview, I have a feeling that it hasn't been deployed fully in the API, best to test it when the announcement comes out👍👍

[Ligtr @kire0n]

@notjazii v4 pro!! 80 cents per million output token for a model that nears or is on par with kimi k3 is going to be crazy!

[Ligtr @kire0n]

@teortaxesTex Still INSANEEEE😧😧

[Ligtr @kire0n]

@teortaxesTex OUR WHALE IS SURFACING!!🐳🐳

[Ligtr @kire0n]

@Artedeingenio I do think that it'll be slightly better than SD 2.0 in its final post training, considering it hasn't hit a ceiling yet and is already competitive against SD 2.0, high motion stuff is still rough, but that's probably what they are going to improve on in it's post training.

[Ligtr @kire0n]

@Artedeingenio Flux 3 is going to be open weights, that's the most exciting point, and it's currently an early candidate, which means it's not the final checkpoint yet, they also said they expect further improvements during the early access phase, which imply they haven't hit a ceiling yet🤔🤔

[Ligtr @kire0n]

@thalwain @junwatu Flux 3 is currently only an early candidate, meaning it's not even the final checkpoint, they've also said that they expect further improvements during the early access phase according to their blog, 'further improvements' imply they haven't hit a ceiling yet🤔🤔

[Ligtr @kire0n]

@RijnHartman "we expect further improvements during the early access phase."

This is from their blog post, it's pretty good right now, they also said "preliminary evaluation of an early FLUX 3 candidate", implying it's not a final checkpoint either, exciting👀👀

[Ligtr @kire0n]

@Cephas_Badger @cocktailpeanut Historically, they usually either release at launch altogether, or a few weeks to 1 month max (FLUX.1 Kontext took only a month i believe), I don't think multiple months is on the table, at least historical wise👀👀

[Ligtr @kire0n]

@teortaxesTex Our whale is surfacing!🐳🐳🐳

[Ligtr @kire0n]

@Elaina43114880 Our whale is surfacing...🐳🐳

[Ligtr @kire0n]

@Waguri_Kaoruko8 Woah! Go whale!🐳🐳

[Ligtr @kire0n]

@teortaxesTex It'll land a 55-57 on AA, I believe in our whale!!🐳🐳

[Ligtr @kire0n]

@7ocemCh @NierYearn @LuminaXspace It'll match opus, a 1.6t parameter model that even very undercooked in RL, was able to land a 44 in AA, it's headroom is big and considering deepseeks reputation with iterating/post training, they'll do fine, and also the fact that they had 2 months to post train v4, let's hope🤔

[Ligtr @kire0n]

@dummerspast39 @teortaxesTex The Moon can stay with Terra and Sol in space, we need something from earth that allows us to do our work without burning up our credits too quickly, that's where the whale is at, on earth with 0.80$ per million token, for all, not out in space with 15$ per million output token!

[Ligtr @kire0n]

@Mikethejohnson @mattschrader @claudeai It is frontier tho, can't deny that

[Ligtr @kire0n]

@teortaxesTex Deepseek instant mode is really token efficient right now, prompts that would take 200+ seconds consistently are now under 100 seconds consistently, excited!!

[Ligtr @kire0n]

@KuittinenPetri @teortaxesTex Luckily deepseek has some focus on RP which is good for creative writing, they have 'Emotional Intelligence Data Product Manager' positions, labs are focusing on EQ almost as much as IQ now, so creative writing will improve regardless👀👀

[Ligtr @kire0n]

@Elaina43114880 Only flash or both pro and flash?