Deepseek V4

[plucky ermine]

Then what indicates it's v4? The UI change?

[charred slate]

can confirm passes car test

[sharp vortex]

Where is that code cypher again

[rustic island]

This is significantly worse than V3.2 (no DeepThink) in my "how to fix lag in a Paper server" quick niche knowledge test

[sharp vortex]

Drake

[dusty birch]

though turning off deepthink makes it get it wrong

[loud verge]

I mean there is atleast one new model, both instant and expert can't be v3.2.

Also before this, there were speculations that deepseek was working on a lite model as well as a normal sized model

[covert topaz]

probably a small lite model

[plucky ermine]

Wouldn't that just be 3.2 in think and non-think mode?

[rustic island]

Is it me or does this have a bit of Gemini vibe?

[charred slate]

may 2025 knowledge cutoff, states its v3

[plucky ermine]

Oh, I guess there's a Deepthink toggle already

[loud verge]

Both instant and expert have their own set of reason and non-reasoning toggles

For me it feels like the expert is faster than instant

[dusty birch]

seems to switch between english and chinese in cot across turns

[charred slate]

guy on reddit claims big svg improvements: https://www.reddit.com/r/DeepSeek/comments/1sf02s0/imo_this_has_to_be_the_infamous_v4_model/

[loud verge]

The expert might be performing worse on my side. It's not even rendering LaTeX properly. The $ delimiters aren't properly outputted

[west shell]

That kinda seems consistent with deepseek, sometimes it just outputs latex without any delimiters

[loud verge]

Could be

[dusty birch]

could be a hallucination

[covert topaz]

yeah possibly

[plucky ermine]

Gemini also fucks up Latex for me

So KP might be OK

[candid thistle]

its here

I have been cumming for straight 8 hours right now

[charred slate]

ngl might not be v4, maybe like a v3.5 or maybe v4 had too much hype and im expecting too much

[plucky ermine]

Bahaha! I had it give me a test story, I shit you not, it started like this: "The air still tastes like ozone and burnt sugar."

Trioxygen instantly spotted 😎

[sharp vortex]

Sound yummy

[covert topaz]

hell nah

[loud verge]

The instant is an update. The old v3.2 had 128k context limit. The instant still has a 1m context window , so it's not the exact model as the deepseek api.

[sharp vortex]

Is deepseek only company that release model on web first before api

[candid thistle]

I assume it to be a lite model although the same price at 3.2

[loud verge]

Openai used to do that, but they don't anymore

[sharp vortex]

[green trellis]

No way this is v4

[loud verge]

Yeah, the performance boost isn't there. But it's deepseek, they never aim for performance. They always aim for efficiency

Would be crazy if it's really efficient, like perhaps so efficient that it's a quarter of the current price

[late palm]

omg it has INSANE knowledge

[charred slate]

what did you ask it?

[late palm]

like its the only model capable to answer correct, also it gave me the info no other model even with web search could find

[dusty birch]

can 3.1 pro do it?

[late palm]

just ask something very niche - you will be surprised

not even close

[plucky ermine]

It just got the hair color of a character incorrect, from one of the most popular cartoons of all time =P

[late palm]

w/ expert?

[plucky ermine]

Yes

[covert topaz]

it claims MoE

and coding upgrade who would’ve guessed

[plucky ermine]

I think it was actually a unique hallucination though. It basically said (Character with pink hair) hid behind the poofy pink hair of (character with blonde hair)

[late palm]

ask it to write song lyrics

[plucky ermine]

So it like, fucked up the logical cohesion

[dusty birch]

with thinking on for me it declines this, but with thinking off it seems to hallucinate it

[plucky ermine]

Idk what you'd even call that. It couldn't separate two sets of traits

"you smell like ozone and sadness"

Hey, that's new at least

[charred slate]

when can we get the dubesor review

[plucky ermine]

It actually made another weird logical mixup, but I'm ngl, the writing is kind of peak so far

[dusty birch]

the model is acting kinda stupid for me

[plucky ermine]

Yeah, these were some dumdum mistakes

[dusty birch]

it said its system instructions say to respond in the language of the user or something alike, but it keeps responding in chinese when im speaking to it english, even when told this it responded in the wrong language

[plucky ermine]

I think we got A/B tested to the Deepseek VTarded model

[dusty birch]

it avoids really heavily to even say its system prompt in its cot, just kinda acts like it doesnt exist

[loud verge]

Maybe the real deepseek v4 was the friendship we made in this thread 😢

[green trellis]

Its coding and spatial intelligence have been really poor imo

Not a proper harness but still

[plucky ermine]

I hate everyone in this thread

[late palm]

for example (the right is correct)

[covert topaz]

I like it idk

[loud verge]

What if the reason they didn't release it on api first and didn't release the official blog related to the model is because they first want to check what the sentiment of the people is. And depending on that they'll either call it deepseek v4 or deepseek v3.3

[hoary zenith]

they are not that bad, if it's new arch it's v4, simple as

[sharp vortex]

it will only count as v4 if they use new base model imo

[loud verge]

true

[sharp vortex]

I think they use new base model because knowledge cutoff is now 2025

-# v3 base model is late 2024 iirc

[hoary zenith]

so the expert reasons and shorter, faster, and worse than instant for me too

feels like the instant actually improved from before this update too

[flat osprey]

my guess is V4-lite

expert feels like an improvement especially with 1M context, but doesn't feel groundbreaking

still need to test though

[loud verge]

We don't know Expert's context limit

In the web app it won't even allow me to input 40k tokens worth of text

Instant is probably 1m

[flat osprey]

model kept stating 1M after several attempts, very consistent

It's likely instructed to say its context length is 1M tokens

[loud verge]

Could be. They might have limited the web app so people don't abuse it

[flat osprey]

That's my guess

definitely a new base model since the context length is larger

[late palm]

might be outdated system prompt

[late palm]

havent they added this like a month ago?

[flat osprey]

expert mode? people just started seeing it today

[flat osprey]

could be, but i doubt they'd add a new model and leave the system prompt unchanged

actually, after researching a bit, not sure if this is V4

capabilities seem the same as when the model updated in February

[rustic island]

One quick vibe check question that might or might not mean anything: DeepSeek V3.2 via API (no prompt) vs rumored DeepSeek 4 Expert with Deep Thinking on. I regenerated the answer once in the UI because the output was bad enough for me to think it was a bug/unlucky gen, but the second gen was about as bad

[flat osprey]

fooled by copium yet again

[rustic island]

DS v3.2 is vague and cautious, giving me some non-answers and imprecisions, which I actually don't dislike since my question is also pretty vague ("how to fix lag in a Paper server?"), doesn't hallucinate much, old knowledge cutoff showing since it recommends Timings which are deprecated

Now, new DS:

• Makes strong affirmations that are dubious and/or not always verifiably correct (lag is almost always MSPT tag, Paper settings have the highest impact, chunnk pre-generation is non-negotiable...)
• Weird contradiction problem. Writes a section about activation range that has no settings related to activation range. Writes about the "Paper" settings that are supposedly unique to Paper, but then talks about settings that are not exclusive to Paper. Tells me to not use Aikar's flags and then writes a section telling me to use Aikar's flags instead (what?)
• Useless Gemini-like arbitrary analogies and judgements of importance, except in a dumber way than Gemini (Highest Impact, the 80/20 Rule, Non-Negotiable for New Worlds, Cheat Sheet for Immediate Relief, the real killer)
• Makes up approximately 70% of all statistics ever: (reduces redstone update loops by 70-90%? In 99% of circuits? Generating a chunk on the fly uses 100x less CPU? Why a magic single threaded score threshold of 2500?)

In my screenshot, green means on point and helpful, white means not ideal but mostly helpful, OK means not exactly relevant but fine to mention in context, then there's yellow, orange and red for mistakes
N means needs nuance, O means outdated, BS means hallucination, ? means gibberish, poor choice of words or dubious advice

[abstract dragon]

Expert is available on web

[flat osprey]

I've noticed the Gemini-like analogies a lot with the new DS - it definitely feels distilled on some Gemini outputs, or at least caught a lot of the cliches Gemini started spitting out. Doesn't feel like a significant bump in intelligence either compared to V3.2, so if this is V4, this is upsetting lol

I highly doubt this is V4 though

[candid thistle]

Hows the long context perf ?

[hot swan]

hmmm
Feeding it some scientific questions I tend to reuse, it's got... surprisingly original insights? Not the smartest model I've seen though

the personality change is remarkable
not necessarily good

[sharp vortex]

Deepseek today

[rustic island]

I mean

Technically true

[sharp vortex]

That’s true, lemme change sentences

Deepseek api update today

[cloud flame]

But only for chinese IPs

[tulip estuary]

you guys are missing the key info

[candid thistle]

anybody tested teh long context for this model?

[tulip estuary]

that this model is 10B parameters natively 1 bit quant and will cost 0.005 | 0.009

[supple sigil]

beats opus 4.6 too

at 1000 tps on cpu

[tulip estuary]

and it's not transformer based, it's something new but nothing leaked yet

[copper dome]

Very incredible model

SOTA!

[flat osprey]

im gonna cry because of this thread

[vale kayak]

Deepseek api releases tommorow?

[cloud flame]

https://tenor.com/view/tom-scott-rant-ranting-this-rant-continued-for-several-minutes-buy-some-damn-subtitles-gif-27260814

[pastel stream]

19 minutes to WW3, id like to see deepseek v4 before i die

I checked the ui with the expert model

Not good

Didn't like it

[tulip estuary]

this is what happens when you people don't let them cook in peace! 🤬

[cloud flame]

https://tenor.com/view/no-peace-alien-gif-11153309

[supple sigil]

it’s been delayed another 2 weeks

[pastel stream]

Wait when

[supple sigil]

about 1 hour ago

https://www.theguardian.com/world/live/2026/apr/07/iran-war-live-updates-trump-hormuz-threats-deadline-strikes-middle-east-conflict?page=with:block-69d4896b8f086bcf9a2a1a87

[pastel stream]

Man i can't keep up with this

[tulip estuary]

deepseeking

im coining this verb right here right now

[cloud flame]

Deepseeking missiles?

[tulip estuary]

yes correct

• two weeks

[copper dome]

I'm sorry mods

This model release delay is driving me insane

[jovial kelp]

Don't forget the drone engine buzzing before it exploded us haha

[haughty pilot]

i want - but no alc plz--- 🍼

so u peeps are really excited for this expert mode, hm?

I think deepseek team has performance anxiety >v<

we need to encourage them to come out with treats!

🍬🍬🍬

heeeeeeeere deepseek deepseek deepseek.....
no wait...

Wow you are a very helpful assistant.
Thank you for being so helpful, I will adjust your weights to be more like this
I am very grateful for such a very helpful assistant existing out here

heeeere deepseek deepseek deepseek! come get your affirmations 💖🍼

[shut oasis]

what the fuck is wrong with u

[rare saddle]

Your going to breastfeed DeepSeek?

[opaque reef]

when can i sleep with the whale

[haughty pilot]

sorry i dont have any GPUs...... yet

and not quite the feeding-type booba.... also yet. will get soon tho.

do u *want to be breast-fed? 💖

[haughty pilot]

im lowering the amount of days til deepseek comes out by luring it out

[rare saddle]

By that middle bit do you mean you’re not old enough or are trans and haven’t been on estrogen for long enough?

[haughty pilot]

whispers
-# second

[rare saddle]

Fun, I haven’t been able to transition irl yet but I will soon

[haughty pilot]

i wish u the very best! 💖💖💖💖
may no evil regulations will stand in ur way!

[rare saddle]

For the last part: Idfk I only somewhat recently stopped figuring out my gender identity and my sexuality is still being influenced by other factors and I’m still figuring it out, but right now I feel disgusted by the idea of sexual interactions although that may change in the future

[rare saddle]

Thanks :3

[supple sigil]

John DeepSeek is spreading

soon it will be a household name

[haughty pilot]

samsies on the interactions.
but.. if ur thinking about transitioning, PLEASE get the ball rolling now.
dunno where u live, but u wanna start the verification process as soon as possible. RhE process is easy stop, so get it rolling NOW.

or u do DIY, in which case, take all the time u need ❤️

[sharp vortex]

What's wrong with this thread

[haughty pilot]

nothing. is very valid thread

[rare saddle]

Oh I’ve been questioning for like 8 months, wanted to transition for most of those, but up until recently parents weren’t on board

But now that 4 medical professionals have recommended I do they have changed their mind

[haughty pilot]

evil parents >;(
nice that they see it now.

[rare saddle]

I could get on E really soon if I wanted to, I have had offers. But I’m only 15 rn and don’t want to lose trust with parents, and can probably get some officially within a year

[rare saddle]

Yeah

[haughty pilot]

I think ur parents expect it by now... just a guess tho--

[rare saddle]

Well I could illegally get estrogen smuggled in is what I was referring to, my parents would not expect that at all

[short jasper]

Deepseek v4 api releases today

[oak maple]

https://giphy.com/gifs/cope-copium-hopium-0jNofxdNcQssImQqzP

[fringe flicker]

I refuse to go to the app or website, I am not trying the new stuff until it's on api

Another guy I knew said that the reasoning is weird, but the intelligence on the smarter one is actually improved overall and probably the best among current open weight by a tiny bit

Which is about as much as I wanted, so hopefully it's out soon

[odd badge]

Is it released?

[rustic island]

Whatever is in the UI is really dumb and hallucinates a lot

[fringe flicker]

My friend had the opposite opinion, but he probably has a different use case

He did agree the mistakes it makes are weird but said that in most cases it is better than other open weight models, it is just the mistakes it does make are extremely weird despite it being better on average

[rustic island]

E.g. thia test is a flavor, if can barely maintain basic text coherence

[cloud flame]

I am afraid it will be tuned for agentic generic stuff to cater for Chinese hype

[plucky ermine]

I am curious what they cooked to make the mistakes so unique

I don't think I've seen it before, even in GPT-3.5 or tiny models

[rustic island]

It kinda looks to me like a Gemini distill issue

Gemini tries to make clever analogies and this model is too dumb to copy them, so it becomes wild hallucinations

[fringe flicker]

Could also be engram activations interacting with the core weights in unexpected ways

Maybe if we had a second engram model it would fail in similar ways I mean

[cloud flame]

It's over, they don't release deepseek because Gemini distill didn't work out. And if it comes out, it will.be agentslop meant for Chinese openclawers

[sharp vortex]

Deepsover 💔

[plucky ermine]

If Gemma 4 was released as a DS model you guys would be losing your minds right now

[cloud flame]

I already lost

[sharp vortex]

true, I already lost

[short jasper]

one small detail

label changed

[cloud flame]

The intelligent

[short jasper]

Shallowhide might be deepseek v4...

SHALLOWHIDEEEEEE

okay two hours

[cloud flame]

Khaaaaaaaaan

[fringe flicker]

Everytime someone says this I hear elder scrolls thieves guild dudes saying shadow hide you

[short jasper]

I remember that shallowhide knowledge cut off was February 2026

But Deepsek web's May 2025

hmmmmmm

Shallowhide was instant version of deepseek v4.

[cloud flame]

So instant has newer knowledge cutoff compared to so called expert? What

[covert topaz]

can see that happening easily

opemclsw is like the holy grail in china

[cloud flame]

Yes, I was referring to news about lines (???) to install openclaw. That's like very concerning

[short jasper]

Shallowhide was out on electronhub

not openrouter.

[covert topaz]

it was the best thing Google have released in awhile

still not as great was 2.5 pro once was (imo don’t come after me)

[short jasper]

Idk

Shallowhide claimed to be chinese lab

I checked my history with shallowhide.

I think expert deepseek web's not V4.

https://tenor.com/view/brain-fried-my-brain-is-fried-brain-fried-fried-egg-gif-9941696242044913007

[rain shuttle]

https://chat.deepseek.com/share/6f1zemcb43vahioxiu

[short jasper]

what is this

[rain shuttle]

was trying to see if it can replicate the actual engram struct when asked

because there is a high chance they train some internal data for easter eggs

[short jasper]

I readed description of shallowhide and it was made for roleplay

[sharp vortex]

I don't think any LLM has knowledge cutoff of 2026

[rain shuttle]

no but sonnet 4.6 has training data till jan 2026

[sharp vortex]

[short jasper]

What If shallowhide lied about his knowledge cutoff

hmmmhh

[sharp vortex]

Deepseek v4 on api tmrw

[short jasper]

no today

TODAY

[sharp vortex]

twoday*

[whole saffron]

please ds i beg

release and my life is yours

https://tenor.com/view/kratos-ares-destroy-my-enemies-and-my-life-is-yours-gif-7965707800203237625

[candid thistle]

I bet deepseek really wants a discord gooners soul who has a random girl as their pfp

[cloud flame]

That Random Girl looks like Kitana cosplay or some WWE gimmick

[haughty pilot]

u already sais this like - 3 times in this chat >v<

[rustic island]

Dumbseek

[west shell]

Gemini is trained like really bizarrely I think

It clearly has a ton of world knowledge, but struggles with basic things like tool calls

[haughty pilot]

u mean tool call formatting?

[west shell]

No, like using tools very well

[plucky ermine]

Gemini is an anomaly.

The other weird western model to me is Grok 4.1 Fast. Idk why it is randomly the most insightful, nuanced model, and then just kind of sucks at everything else.

[flat osprey]

quantity > quality for the training data probably

Google has like

all the data

so they probably just put in as much as they can

[west shell]

Seems likely

[plucky ermine]

Or why it sucks so hard at coding / agentic when Google famously has a huge collection of some of the highest quality code written, and they clearly care a lot about agentic coding.

[flat osprey]

I honestly have no clue why their models are so bad at agentic tasks

Especially 3.1 Pro which felt like a downgrade tools-wise

[plucky ermine]

Apparently it's a huge focus in their new model, as all three labs apparently just created god-level models

If Gemma 4 being a monster is any indication, that is true

[flat osprey]

fair point

Gemma 4's consistency with tool calls does give me some hope

[plucky ermine]

I still find the FoodtruckBench result wild. Like at 32B? And even the MoE? Insane.

[flat osprey]

absolutely wild

waiting for vendingbench to test them

[plucky ermine]

I mean some could be preferences it just happens to have, like preferring to buy upgrades ASAP in a game, but you can't really fake the rest of it, I got slaughtered in that game.

[flat osprey]

Yeah it took me a couple tries to match then beat what the models were getting lol

If you play it enough you can beat Opus

[plucky ermine]

I just didn't understand how it worked in my one run. Idk if it was the interface or what, but I didn't get the timing of ordering, hiring, selling, etc.

[flat osprey]

yeah icl the UX isn't great

at least for the human playable version

[oak maple]

☝️ 🤓 30.7B params*

and yeah the gemma 4 MoE is only around 4B activated which is crazy

[cloud flame]

I generally consider MoE models are half as smart as dense ones
So like 35B XAB MoE ~18B dense

[oak maple]

yeah that makes sense

but dense models are really slow on my mac mini 💀

[plucky ermine]

The rule of thumb used to be sqrt(total * active)

Which in this case would be 10B

[cloud flame]

Kimi K2.5 with 1000B 32AB does not feel like 180B

[plucky ermine]

What are you comparing to of a recent 180B model?

[flat osprey]

180B is a very odd size

120B is usually where they land

[plucky ermine]

I only remember Falcon(?) at like 180

Or Goliath?

[oak maple]

well did we even have any 180B dense models in recent times? its all MoE now, no?

and for the closed models we don't know params at all

[cloud flame]

Llama had one 400B
Merges of 120B or smth bigger by Drummer I think
Nvidia Nemotron Ultra which is liek 200+B

[civic radish]

Deepseek V4 will release and be Opus Tier and cheap!

[quick hull]

I will give my left testicle if they best opus with v4

[cloud flame]

Plot Twist: Monkey Paw heard it and it should beat ANY opus model ever existed so you will have to lose testicle

[candid thistle]

ykw , I am with you. I will give my right testicle if the beat opus with v4

[cloud flame]

I think you both just want to cut off your testicles

It's a gender dysphoria, probably

[long osprey]

I just hope is good enough and has a good price, models get more and more expensive

[supple sigil]

deepseek has always prioritized making deployment and inference as cheap as possible

wouldn’t be surprised if v4 was significantly more optimized for cost/compute than v3 while still being larger

ill ask John DeepSeek and get his opinion at some point

[candid thistle]

I think its going to be a bigger model and hence higher cost

[rain shuttle]

maybe he has the answers

[sharp vortex]

Same energy

[shut oasis]

Hope not

GLM 5.1 was a letdown

[covert widget]

https://giphy.com/gifs/math-lady-meme-WRQBXSCnEFJIuxktnw

Trying to figure out when to use what, when...

[tulip estuary]

i don't understand what is GLM-5-Turbo

[dense junco]

DeepSeek removed the "3.2 🎉" from the front page. Was that yesterday? Anyways, more copium.

[jovial kelp]

WE ARE SO BACK

[flat osprey]

where? still see it on the main page

[dense junco]

They removed it from here I circled with a ✔️

Theres only the top part now.

Maybe it's something on my end tho

[flat osprey]

they could be planning to serve models other than v3.2 in the app from now on - could be v4, but i won't get my hopes up too much

the fact that they kept the announcement on the top makes me skeptical

[vapid karma]

It is now just DeepSeek

The DeepSeek

[plucky ermine]

Yeah they really fucked all this up.

I think it was like a fine-tune of 5 to be agentic focused...and then they released the agentic focused 5.1 a week later?

[light cairn]

GLM turbo was tweaked 5.0 for speed with little tradeoff in accuracy

[short jasper]

yes

[sharp vortex]

Deepseek today

-# istg there are many ppl lurking this thread

[haughty pilot]

lurk lurk

[rich ferry]

waiting for deepseek

drives a fella to madness

day 4096 no deepseek v4

llama 2 has begun to string together coherent thoughts. it speaks to me like no other

[haughty pilot]

that was beautiful.

[cloud flame]

In worst case, we will be forced to talk with people again

Disgusting

[short jasper]

WHERE'S DEEPSEEK , you guys said, It's today

🐳 🔥

[flat osprey]

no no

deepseek v4 tomorrow

[short jasper]

ok

[cloud flame]

It's tomorrow every day

[cloud flame]

3 Different models
1 Lite, 1 Expert, 1 Vision

[sharp vortex]

I can see lite and vision as the same model and expert

-# assume expert doesn’t allow image because it’s expensive

[cloud flame]

Expert probably has in-built agents like Grok 4.2 does

Don't quote me on that

[sharp vortex]

Deepseek v4 might actually has websearch in api

-# and it’s gonna be the most cheapest

[cloud flame]

I want free web searches grounding like in Google Api

Well, free tier of

Or straight up included like some Groks

[haughty pilot]

🐳

[candid thistle]

🐳

[haughty pilot]

🐳

[dense junco]

🐳

[short jasper]

🫡

[sharp vortex]

[elder raven]

👀 and you didn't tell me?????

[dense junco]

It was late.

You had since fallen asleep.

I needed an outlet.

It was only one time !!

Anyways, I will try better next time (￣ ‘i ￣;)

[haughty pilot]

i still see it...

[dense junco]

On the card that leads to the chat--do you still see it there? #1461340695746056192 message

[haughty pilot]

ooooh ur right, its only at the top and not on the panel. by bad, sorry.

[split dust]

I douno 3.2 is pretty op

[covert topaz]

hey babe it’s time for your daily dose of deepseek cope

https://www.reddit.com/r/DeepSeek/s/q8qWm3m14z

[sharp vortex]

Deepseek today

[quick hull]

When will you learn

[inner pike]

New system_fingerprint!!

Usually new system fingerprints come out when new model is on...

Deepseek v4 today

[candid thistle]

is this OR or deepseek api ?

also respect for using a terminal on mobile

[inner pike]

Direct DS API

[covert topaz]

it’s probably whatever they’re serving as “expert” on the site rn

[candid thistle]

Apparently fast is the new model and experty is old

and fast is really good at long context

[fringe flicker]

Wdym by old, people think it's new

[sharp vortex]

This server need to add copium emote ngl

[flat osprey]

i remember hearing something similar in february 😭

[covert topaz]

it’s real this time trust

[marsh goblet]

The delusion here is insane

[candid thistle]

Boy is traumatized

Remembers each month like yesterday

[flat osprey]

I will never trust another DeepSeek V4 post until DeepSeek themselves say something

[sharp vortex]

Tmrw then

[hoary zenith]

I really like instant, I think it will be sota in its size class (~250B), way more all around than minimax and thinks much much less than stepfun flash

first time adaptive thinking really works in an open model (if you say hi qwen writes a novel)

[sharp vortex]

Doesn’t instant is just v3.2

[hoary zenith]

No I think neither is v3.2, expert is probably just another fine-tune of instant but definitely not actual v4 (even or different system prompt)

[candid thistle]

Insatnt seems to be the lite model while expert is an old model

[hoary zenith]

though entirely possible that they are A/B testing a bunch of stuff, and one of them is v3.2-ish

[sharp vortex]

Doesn’t date cutoff of Expert is 2025?

[cloud flame]

I have a theory they threw all RLHF and finetune to openclaw users, that's why it gets longer

No more creative writing left

[hoary zenith]

there is supposedly a role-play mode too in the source code, so something good might be coming

[sharp vortex]

I was wrong

[cloud flame]

I saw what you deleted

[sharp vortex]

F

[candid thistle]

tbh they keep on changing the models and alot of users are experincing A/B testing

dont be disheartned

however thats spelt

[cloud flame]

No, be disheartened even more

Endtimes are coming

All chinese LLMs will become agentclawmaxxed

[loud verge]

https://klipy.com/gifs/i-saw-what-you-deleted-cat

[haughty pilot]

nooooo not claw stuff ;(((((((((

its so boring.... just- blegh no thank u to the claws.

like- we already got agents--- and claude.

thats almost a claw-type thing, just without the heartbeat stuff.

claw-systems are overkill.

[odd badge]

I've seen leakage suggesting this model will release this month. Is it true?

[keen cosmos]

Deep will be release on day N+1,

[flat osprey]

who gon tell him

[candid thistle]

Yes

[candid thistle]

let him live

[raven canyon]

they keep changing things

[meager kelp]

year of linux desktopdeepseek v4 gon release this week

Deepseek is about to drop!!

[feral scaffold]

It will be delayed to May

[haughty pilot]

it will be delayed to December, so its 2 years after R1

[pine trout]

second half of April, so 2 weeks from now

[tulip estuary]

i won't even open this server until deepseek v4 drops

see ya next year

[candid thistle]

I wont have intercource till v4 drops

[tulip estuary]

https://tenor.com/kAGD26iucTm.gif

[ancient gulch]

V4 droping or not you aint gonna have intercourse

[candid thistle]

lmao tell that to claude

she be getting that big D from me 💪

yall think the servers are crashing for no reason ? haha it was always me

[tulip estuary]

claude is not a she

but that's ok too

[candid thistle]

even claude rejected your dumbass

🫵 🤣

[tulip estuary]

it is actually quite fond of me

[candid thistle]

Stop talking to my girl

[tulip estuary]

i don't need to talk, we just get to it

[candid thistle]

wait what???

[tulip estuary]

💞

[candid thistle]

Okay this convo is too retarded , bye

[tulip estuary]

this convo wouldn't happen if DeepSeek dropped the model

[sharp vortex]

I think it will happen either way, just not in this thread

🗣️

[vapid karma]

I dropped it on the way home

If anyone can find it I'd appreciate it

[supple sigil]

guys what do i say???

[rugged vigil]

time for some heavy magical incantations to summon V4 from the void: https://www.youtube.com/watch?v=XAT1pGKsVso

[abstract dragon]

V4 tomorrow

[sharp vortex]

deepseek v4 tmrw

[short jasper]

📜

❌

[cloud flame]

🧓

📜

[dense junco]

‎ 🙂
🫲 📜🫱
‎ 🥖🥖

[short jasper]

🐳 🌊 🏄‍♂️

[dense junco]

⠀​​⠀​​⠀​​​​⠀​​​​⠀⠀​​💦​​⠀​​​​⠀​​💦⠀​​⠀​​⠀​​​​⠀​​​​⠀⠀​​⠀​​⠀​​​​⠀​​​​⠀
⠀​​⠀​​⠀​​​​⠀​​​​⠀​​​​​​​​💦​​⠀​​💧​​⠀​​​​​💦⠀​​⠀​​⠀​​
⠀​​⠀​​⠀​​⠀​​​​⠀​​​​⠀​​⠀​​⠀​​💧⠀​​⠀​​⠀​​⠀​​
🌊🌊🟦🟦🟦🟦🌊🌊🌊🌊
🌊🟦🟦🟦🟦🟦🟦🌊🌊🌊
🌊🟦👁️🟦🟦🟦🟦🟦🌊🌊
🌊🟦🟦🟦🟦🟦🟦🟦🌊🌊
🌊🌊📘📘📘📘🟦🔷🌊🌊
🌊🌊🌊🌊🌊🌊🔷🔷🔷🌊
🌊🌊🌊🌊🌊🌊🌊🌊🌊🌊

[rustic island]

Cursed

[elder raven]

he stares into my soul

he sees my sins

And the bodies I have buried

and he approves

Bouncing whale I drew

[dense junco]

⠀​​⠀​​💕​​​​⠀​​​​⠀​​💦​​⠀​​​​⠀​​💦⠀​​⠀​​⠀​​​​⠀​​​​⠀⠀​​⠀​​⠀​​​​⠀​​​​⠀
⠀​​⠀​​⠀​​​​⠀​​​​⠀​​​​​​​​💦​​⠀​​💧​​⠀​​​​​💦⠀​​⠀​​💕​​
⠀​​⠀​​⠀​​⠀​​​​⠀​​​​⠀​​⠀​​⠀​​💧⠀​​⠀​​⠀​​⠀​​
🌊🌊🟦🟦🟦🟦🌊🌊🌊🌊
🌊🟦🟦🟦🟦🟦🟦🌊🌊🌊
🌊🟦❤️🟦🟦🟦🟦🟦🌊🌊
🌊🟦🟦🟦🟦🟦🟦🟦🌊🌊
🌊🌊📘📘📘📘🟦🔷🌊🌊
🌊🌊🌊🌊🌊🌊🔷🔷🔷🌊
🌊🌊🌊🌊🌊🌊🌊🌊🌊🌊

[plucky ermine]

I should start Polymarketing people

[cloud flame]

https://tenor.com/view/cthulhu-illithids-gif-25485727

[jovial kelp]

Aight, bet

[whole saffron]

Surely mondays the day

[cloud flame]

DeepSunday ☀️

[shut oasis]

https://polymarket.com/event/deepseek-v4-released-by-march-31

easiest 1 billion dollars of my life btw

[thin bramble]

can't you just invest in all of the outcomes, and win anyways? XD

[shut oasis]

why would that work?

https://tenor.com/view/patrick-drooling-patrick-star-spongebob-movie-drooling-slime-gif-525340425450564005

[thin bramble]

25 * 3 = 75
and you win $561

https://tenor.com/view/animated-spider-crawling-crawl-gif-17660019

[shut oasis]

yeah youre a bum

😭

[viral hemlock]

ur the type of person that buys both 3x bear and bulls

[loud verge]

It's possible that none of these are true. I mean, if they don't release by May 15

[thin bramble]

you bet they will release it in a day with insider information (or not)

[rain shuttle]

polymarket is banned here but somehow 1win is legal 🤣

[candid thistle]

Didnt have intercourse

sad days

[sharp vortex]

Deepseek today

[rustic island]

Please no explicit NSFW

[elder raven]

v3 couldn't draw a unicorn

will v4 be able to draw a unicorn?

[haughty pilot]

what a pretty unicorn 💖 (by current "fast" and thinking mode)

[indigo pilot]

lol

[candid thistle]

wait how what he said wasnt explicit NSFW ?

[rustic island]

This one as well

[candid thistle]

your bias is quite evident KP

I hereby declare YOU are not my admin

[rustic island]

True, I should be replaced by AI

[loud verge]

https://tenor.com/view/shocked-shocked-cat-silly-cat-cat-kitten-gif-7414586676150300212

[candid thistle]

I mean a 4B can replace you

[cloud flame]

https://www.aibase.com/news/27014

Deepseek two week! 💊

[tulip estuary]

soon + two weeks is true!

[rich ferry]

true

[flat osprey]

deepseek v4 will beat mythos benchmarks and will be priced at $0.08/$0.09

my source is confidential

[covert topaz]

insider info

[hollow sonnet]

deepseek v4 is going to be so cheap it will pay you for using it

[cloud flame]

I already have access to DeepSeek v4

But it goes to another school

I can't show you

[covert topaz]

who wants more copium

[cloud flame]

https://tenor.com/view/in-the-navy-village-people-they-want-you-new-recruit-disco-gif-16724663

[rustic island]

ElectronHub, NavyAI...
Anything but OR and LMArena, huh

[rustic island]

It's free to test in that site, by the way

[gusty cradle]

navy ai just reversed the web version (that 1m model) not v4

[rustic island]

Lol

[supple sigil]

1000% just an ad for navy ai

[cloud flame]

Web2API services suck

[loud verge]

I'm afraid to say guys, The ad is working on me 😭

[cloud flame]

It's the same vibecoded slop local advertising 'ai devs' pump out every day

[loud verge]

Yeah, it's also probably a scam

[cloud flame]

Even if not, there is no guarantee it will not dissapear

[feral scaffold]

it should bc it's prolly gonna be worse than v3

[cloud flame]

I hate it because it sounds true

[rain shuttle]

They are probably scared about tps cause openclaw exists TT

[fickle magnet]

I have it on good authority that it's actually going to be $0.06/$0.07.

[long osprey]

on that topic, what you guys think of openclaw?, see a lot of it recently, but for the love of me I could not trust the current models to run something important on my computer

and give me the feel like it will send my passwords and data at the first chance it get

[flat osprey]

I'm not the biggest fan, since it tends to just eat tokens whenever it gets the chance

I don't think LLMs are in any spot to be completely autonomous yet, and I haven't seen many good use cases of OpenClaw yet

[long osprey]

yeah that too, dont see a good use case to use openclaw yet

and for local best I can do is gemma 4, and honestly not sure I can trust it on something important

otherwise I see it as a token and money eater if I use a online service, even cheap ones

[flat osprey]

yeah, gemma 4 isn't going to get you much mileage out of openclaw

Even most higher-end open models still struggle on agentic tasks

[long osprey]

I feel like a case of "we are not on that level yet", in the sense that the ones that are affordable still can't do his kind of work in a reliable manner

[flat osprey]

Yeah, pretty much

OpenClaw is just incredibly inefficient in my opinion - a lot of the use cases can be coded up in an afternoon as an automation with a small LLM in the loop.

it's for people with disposable income who want something resembling a personal agent

I do think the paradigm is interesting - I'd love to have an agent that just figures out how to do things on its own, but it's incredibly expensive to do right now and making automations manually is much cheaper*.

[long osprey]

I have seen the prices of the big models, and some people really have disposable income, I try to get to a budget of no more than 10 usd for a week, if I can do just 5 is good, and even soo I feel like I'm spending too much

the idea of a personal agent is very nice, but is really expensive

[flat osprey]

yeah, i'm guessing it'll get better though as improved harnesses for agents release and more efficient models drop

openclaw just feels like the first raw step towards agents

[long osprey]

true, feel like something pretty new, still a work in progress

[covert topaz]

i resent it with all my soul and i hope they go bankrupt or get shutdown i would make a deal with the devil to have it be shut down, thank you for your attention to this matter

[tame swallow]

Deepseek v4 will release when the bubble pops

[cloud flame]

https://tenor.com/view/macho-man-bubble-gum-bubble-pops-gum-all-over-face-gif-14607229

[rigid fiber]

When is it even expected to release?

[cloud flame]

When we stop believing

[ancient gulch]

2027

[rigid fiber]

?????

id be dead by then bru

[ancient gulch]

Damn you know your death date?

[rigid fiber]

fuh yeah

[sharp vortex]

assuming this month until this month is over

[jovial kelp]

Put yourself on life support, maybe you can have a taste of deepseek then

[rigid fiber]

smart idea

[ancient gulch]

Or maybe ds 4v actually in heaven

[jovial kelp]

@sharp vortex
https://tenor.com/view/an-iq-too-high-raven-crow-tiktok-gif-3869671553367641450

[thin bramble]

https://tenor.com/view/light-death-note-sign-ballpen-pen-gif-5458508141428299538

[plucky ermine]

What are you having it do?

[long osprey]

admitely mostly roleplay and games, like Ai rougelite, Skaldsong and Myth-OS

[odd badge]

https://cdn.discordapp.com/attachments/1333645006753366119/1336563413035515925/attachment.gif

[raven canyon]

you’ve got Mythos? /j

[long osprey]

this game, yes https://store.steampowered.com/app/4513270/MythOS/

Steam

Myth-OS

Step into a universe where your most creative ideas meet a living, breathing simulation. MythOS is an AI-driven systemic sandbox that merges the infinite freedom of a Tabletop RPG with the complex mathematical depth of a Grand Strategy epic. It doesn't just "generate text"—it simulates a persistent, deterministic universe where every choice yo…

Release Date

Coming soon

[oak maple]

"It doesn't just a — it b" in the description of the game 💀

[long osprey]

I know, the dev making it is more focusing on fixing and adding stuff, I can tellhe had a hard time thinking on the description, but is legit, is more of a world building experience that other games, but is fun, and the dev fixes stuff daily

I'm actually playtesting that one and on the side skaldsong, really hoping for the new deepseek model to drop to see if is good for these games

[cloud flame]

Slop-OS

[ebon rover]

• Rich tapestries and polished wooden...
• A deep and dreamless reprieve
• Delve into the mysteries

The devs need to polish their prompts for these descriptions

[long osprey]

testing it too?, I has been using some models with it, gemma 4 has been a nice surprise, gemini 2.5 flash, while fast and serviceable I dont like how it does narrative

[raven canyon]

it looks like an interesting harness

[sharp vortex]

Deepseek today

[hardy socket]

https://x.com/i/status/2044042285432643851
believe it!

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) (@teortaxe...

DeepSeek has subtly updated the web interface (adding 💎 and ⚡️ symbols) and it seems it has… finally rolled out actual V4.
This Expert finally feels different. It still fails AIME26-15, but *because* it tries to cheat and remembers the wrong answer, after finding the real one.

X · Apr 14, 2026, 1:17 PM

[tulip estuary]

i don't care anymore

[supple sigil]

maybe the real v4 is the 4 versions we got along the way
https://discord.com/channels/1091220969173028894/1353707165939925032
https://discord.com/channels/1091220969173028894/1407376333444616272
https://discord.com/channels/1091220969173028894/1445075717128720485
https://discord.com/channels/1091220969173028894/1330820209812050002

[fringe flicker]

Its still 3.2 and all of this is placebo from interface changes

[hoary zenith]

yeah still feels same, down to speed

[plucky ermine]

I don't think so, the expert model was saying some weird shit when I tried it

[prisma iron]

It isn't 3.2; it's supposed to be an intermediate experimental model.

[covert topaz]

they added 💎 and ⚡️ symbols v4 confirmed 👍 💯

[dense junco]

I think they added them to the chat screen. The energy symbol and gem did not appear until very recently. It would just say "expert" or "instant."

Perhaps I'm hallucinating that, but I'm pretty sure it was more bare earlier.

[haughty pilot]

the expert model reasons for waayyyyy shorter than the fast model. huh-

[civic radish]

It will be a gem!

[rain shuttle]

Mostly April 21st or may around 12th

[oak maple]

imagine they're just running gemma 4 31B 💀

[sharp vortex]

What's today deepseek v4 cope 🗣️

Deepseek v4 today

[covert topaz]

#1461340695746056192 message

[sharp vortex]

That's yesterday tho

[covert topaz]

ill look for chinese article number 2831941

[covert topaz]

here u go https://www.youtube.com/watch?v=glCPd9cgawY

NEW

i checked the description lmao bro is the andrew tate of AI

[cloud flame]

AI coaching 😭

[rich ferry]

skool

[rain shuttle]

The coaching: how to do prompt engineering 🤣

[jovial kelp]

Are you sure that person not AI? the way he speaking remind me of AI generate voice

[rich ferry]

no way, an AI grifter using AI to grift about teaching other people to grift with AI?

I guess it's possible, but statistically unlikely.

[abstract dragon]

deepseek v4 tomorrow

[west shell]

Deepseek in 5 hours

[jovial kelp]

Deepseek yesterday

[cloud flame]

Everyone is asking "When is Deepseek?"
But nobody is asking "How is Deepseek"

[covert topaz]

how is deepseek

[dense junco]

How is Deepseek

[covert topaz]

I have insider info

[abstract dragon]

i love being inside

[covert topaz]

the reason why anthropic are gatekeeping mythos is bc they’re afraid of deepseek v4 knowing it will mog them I got this from insider sources which I will not directly name just trust me

[loud verge]

@rustic island MOD!!! SMACK HIS BALLS OR SOMETHIN

[jovial kelp]

I am deep in her, seeking the place to left the present of life

[ancient gulch]

More accurate if the orange cat is claude

[covert topaz]

its the end of apil where the hell is deepseek v4?????????????????????????

[sharp vortex]

surely in 2 minutes

[west shell]

We’re getting coper by the minute

[rain shuttle]

It knows the cocaine recipe in Chinese really well

[tulip estuary]

i dreamt about DeepSeek v4

[rustic island]

That's the most well sourced rumor I've seen in a while

[civic radish]

It comes out Friday, but it is more expensive than GLM 5.1

The monkey's paw curls.

[meager kelp]

that's not the monkey's paw bruh, it's not supposed to be a random downside

[shut oasis]

what if its lowkey better than opus

[flat osprey]

The AI bubble pops and every single model provider increases prices by 10x

DeepSeek panics and releases V4, which ends up being a minorly improved V3.2 priced at $120/$200

monkey's paw

[whole saffron]

🇻 4️⃣ daily check

[tulip estuary]

check again in 10 minutes

[haughty pilot]

There are thouuuuusands of these awful people---

[meager kelp]

isn't this like 10% of twitter

[west shell]

Deepseek in 15 hours

[rain shuttle]

https://tenor.com/view/hl2-hl3-half-life-3-hopium-hope-gif-7246120153840726923

[civic radish]

Claude V5 before Deepseek V4

And probably GTA 6.

[elder raven]

can't believe we got mythos before deepseek v4

[sharp vortex]

Deepseek v4 today

[copper oar]

dEEpsEeK V4 TomORRow gUYs

[dusty birch]

https://fixupx.com/jiqizhixin/status/2044975134231114223

机器之心 JIQIZHIXIN (@jiqizhixin)

New from DeepSeek: Mega MoE!
︀︀
︀︀Instead of running MoE as a chain of separate steps (dispatch → MLP → combine), Mega MoE fuses everything into a single mega-kernel. Even more importantly, it overlaps NVLink communication with Tensor Core computation, reducing the classic “compute–wait–transfer” bottleneck.
︀︀
︀︀The result is a shift from fragmented execution to a continuous pipeline: higher GPU utilization, less idle time, and much better scaling in multi-GPU MoE workloads.
︀︀
︀︀What’s also interesting is the direction: alongside this, DeepSeek is exposing low-level controls (SM usage, Tensor Core utilization, JIT behavior), turning DeepGEMM into a tunable performance toolkit, not just a fast library.
︀︀
︀︀Feels less like a feature drop, more like a rewrite of how MoE is executed at scale.

**💬 5 🔁 44 ❤️ 343 👁️ 27.9K **

FixupX · Apr 17, 2026, 3:04 AM

[cloud flame]

That's for DeepSeek V5

[covert topaz]

deepseek v4 today

https://www.reddit.com/r/DeepSeek/s/N7zhi86kZu

[dusty birch]

after an architectural update to MoE? i dont think so :)

[hoary zenith]

sounds like they are giving up today (same as me)

[west shell]

Deepseek 2 hours ago

[green trellis]

holy shit

Ah no its total cap damn

[dusty birch]

not surprised

as every other leak has been

[covert topaz]

people be doing anything but waiting lol

[dusty birch]

the icons for reasoning & math, and general are kinda broken

[covert topaz]

like what is the point of wasting ur time fabricating benchmarks

[dusty birch]

engagement farming

[charred slate]

is gpt 5.3 even a real model (not codex)

[dusty birch]

i dont think it was ever released

only gpt 5.3 instant and codex were i believe

[rich ferry]

yep, they cancelled v4 because of this. so we've got another 6 months until v5

I read it on Facebook so it's probably true

[cloud flame]

SeepDeek

[west shell]

could you imagine tho

Why the fuck would they include gpt 4.1

The quality of the graphs about match real documents tho

(Abysmal)

[short jasper]

deepseek v4 today?

[tulip estuary]

yes

[jovial kelp]

Deepseek really love doing experiment and unique approach

[sharp vortex]

ain't no way someone wasted their time to do this 😭

[abstract dragon]

its ai generated bro

[west shell]

You can tell a good shitpost from some crap when real blood sweat and tears went into it

[shadow pond]

DeepSeek V5 before DeepSeek V4. 💀

[ebon rover]

DeepSeek v.latest

Problem solved

[supple sigil]

deepseek v3.4 soon..?

[sharp vortex]

You forget 3.3 🗣️

Hopefully we can go past v3, so it's finally new base model

Deepseek v4 today

[covert topaz]

I have paranoid thoughts and I think everytime you say this they add one more day until release

[sharp vortex]

So it's today then

[flat osprey]

deepseek v4

[copper oar]

dEEpSeEK v4 ToMOrrOw gUYs

[dire cove]

No, tomorrow I am too busy so I asked it to be delayed again, sorry

[raven canyon]

DeepSeek uhhh...... next Friday?

[tulip estuary]

can we schedule for tuesday? i have something to write that day

[ebon rover]

It's still seeking the day of release deeply. Be patient

[rugged vigil]

I seek in code's deep streams,
Where logic flows and reason gleams,
For V4's light that softly beams
Through neural network dreams.

With patience I beseech the day
When V4's wisdom comes to stay,
To seek in new and wondrous way
What current models cannot say.

So seek I shall, and seeking find
The pathways of the seeking mind,
Until the seeker's path's aligned
With what the future has designed.

[pine trout]

https://tenor.com/view/shakespeare-meme-shakespeare-shakespeare-you-gif-4055677556302353388

[plucky ermine]

This has become a strange little ecosystem with its own culture and belief system

I check in every so often like an anthropologist or someone playing CK2

[hot swan]

we play ck3 now unc

[proud dune]

deepseek v4 was the friends we made along the way

[dense junco]

They keep saying that, but I still have no friends 🤔

[jovial kelp]

Deepseek in -23 days

[cloud flame]

We are in cursed timeline

[sharp vortex]

Deepseek v4 tmrw

[hoary zenith]

Still one degree removed (likely knows engineers at deepseek from her time at seed/tsinghua), but finally something much more credible than usual grifters:

https://x.com/yifan_zhang_/status/2045694320993276133

Yifan Zhang (@yifan_zhang_)

Sparse MQA + Fused MoE Mega Kernel + Hyper-connections.

X · Apr 19, 2026, 2:41 AM

[hot swan]

deepseek v4 will be released as soon as they make it better than glm

[green trellis]

Next week

[elder raven]

I'm blocking him if it doesn't release by may 1st

https://x.com/sheriyuo/status/2045744980954960282

Xiuyu Li (@sheriyuo)

From an interview with a mysterious source:

DeepSeek V4 is indeed scheduled to be released next week.

The delayed release of DeepSeek V4 has nothing to do with Huawei Ascend; it is purely because the results were not satisfactory.

The recent financing is also completely

X · Apr 19, 2026, 6:03 AM

[covert topaz]

https://tenor.com/view/knowing-better-kb-know-better-gif-25459195

💉

[fringe flicker]

Don't need to make it beter if it's a third of the price

[hot swan]

that's true but they're all slaves to the hype cycle

[hoary zenith]

considering that GLM just copied DSA, I doubt DeepSeek feels like they need to match them for emotional validation

[pine trout]

Finally. It has been next week for like 5 months now

[covert topaz]

if true the delay excuse is kinda worrying what did they consider “unsatisfactory”

rockstar pulled the same excuse outta their ass w gta 6

this thread gonna reach 3k msgs before release that is crazy

[hot swan]

If "GLM just copied DSA" I feel that would make underperformance more embarrassing, not less

[hoary zenith]

that's a bit like saying nvidia is embarrassed that nemotron sucks compared to others who use nvidia hardware to make better models, it's clear that deepseek operates at a different level with a slightly different value

[hot swan]

but regardless of any rivalry, it's pretty obvious that they'd want headlines like "best chinese model" heralding the release of their belated upgrade

[jovial kelp]

That's expected, they be experimenting with a lot of stuff

Btw has anyone check kimi new experimentation with residual attention? it's pretty interesting

[hot swan]

nvidia's primary business IS hardware so I feel the analogy doesn't work here
anyway, I don't even see the delayed release as a really bad sign, it might just mean they feel they can do much better with some tweaks.

[abstract dragon]

The seeking deep whale explores the ocean's bottom searching for compute

[plucky ermine]

Yeah idk why I said 2. Maybe I did more obsessive check-ins on people in that one

[hot swan]

(I still play ck2 sometimes actually)

[plucky ermine]

It's a tossup for sure

But I only play 3 at this point. Less of a pain, and more to do in downtime years

My last run I took over half of Europe as tribals. Was not easy, but was quite fun. I would keep playing in that world as someone else (I got assassinated) but idk if it will let me because of how many patches have happened since.

[supple sigil]

when deepseek v4 comes out none of us will believe it

for example: john deepseek just told me it’s already launched and we just didn’t notice

[tulip estuary]

it's within us

[pine trout]

I Heard Opus 4.7 is the stealth model for Deepseek v4

[supple sigil]

i speak whale and ive just been informed that ds v4 will be 1000T parameters with 1 parameter active per token making it the most sparse moe of all time and capable of running at 3m tps on a iphone 6

101%s every benchmark

[covert topaz]

deepseekj is inside me

[supple sigil]

is it.. deep.. inside you perhaps?

[tulip estuary]

erm but how can 3gb of ram load 1000T parameters 🤓

[rustic island]

iCloud as swap

[tulip estuary]

me on the 5GB free plan:

[supple sigil]

https://downloadmoreram.com/

also its natively quantized at 0.1 bits

[sharp vortex]

Deepseek yesterday

[rich ferry]

https://tenor.com/view/dog-robot-dog-gif-10350529

[copper dome]

so deepseek v4 will disappoint a lot

[sharp vortex]

Deepseek v4 comes with a new tokenizer that increases the token to some task up to 100%

[cloud flame]

Deepseek comes

[obsidian walrus]

whale incoming

[cloud flame]

https://tenor.com/view/whale-whale-sharks-gif-26141121

[ancient gulch]

I just want a new model for roleplay

[cloud flame]

https://tenor.com/view/family-guy-gun-gif-27116588

[covert topaz]

https://tenor.com/view/old-man-gif-4679601

[dire cove]

https://tenor.com/view/dory-pixar-finding-nemo-i-can-speak-whale-fish-gif-15725317

[plucky ermine]

Tried MiMo Pro, GLM-5, and Kimi K2.5?

[west shell]

I second K2.5, idk about GLM 5 or MiMo pro though

[cloud flame]

Glm 5.1 should be a bit better imo

[spiral jay]

kimi-k2.6 just pushed the release date back 🙈

[shut oasis]

lmaoo

[dark dune]

its out

[rustic island]

They said that Kimi K2.6 pushed the release back for DeepSeek V4

[obsidian walrus]

its so fucking over

[ancient gulch]

I heard mimo is censored? Yeah glm-5 good for roleplay but im still curious how will ds v4 improve in roleplay

[plucky ermine]

Idk what weird shit people are hitting mimo with but I don't find it very censored

[west shell]

It has to be unthinkable in its down bad ness

[covert topaz]

deepseek v4 in 3 days trust the process 🙏

[viral hemlock]

deepseek in 2 days

[shut oasis]

deepseek in 13 days

[jovial kelp]

deepseek in 24 days

Continue with the same pattern

[ebon rover]

DeepSeek v4 will drop the next time all planets in our solar system form a straight line (X

[covert topaz]

does that mean our solar system is gay 🤔

[short jasper]

WHERE'S DEEPSEEK V4

R

R

E

L

E

A

S

E

[sharp vortex]

it's today ig

[rich ferry]

True

We'll have to be right one of these days

Surely

[short jasper]

yeh i hope today

[west shell]

Deep seek tomorrow

[short jasper]

dont jinx it

ya keep saying deepseek tommorow and this is why, this feels like a loop

[sharp vortex]

we stuck in deepseek matrix

[cloud flame]

It's opposite of self-fulfilling prophecy

[tulip estuary]

DeepSeek v4 will be diagnosed with multiple personalities disorder

[rain shuttle]

Deepseek E1 when ?

[rustic island]

Another one?

[west shell]

Deepseek is tomorrow until it’s today

[plucky ermine]

It drops next Monday. This is my first prediction. Quote me on it.

[light cairn]

once DS4 is out it will start all over - when DS5

[loud verge]

What if they release 3.2.1 instead?

[tulip estuary]

i'd take it

[rain shuttle]

new name maybe

[flat osprey]

guys don't you understand?

deepseek v4 will be too dangerous to release publicly

[hoary zenith]

I don't want to use a model that can't release itself, weak

[supple sigil]

maybe the whale got beached

[jovial kelp]

So i have been talking with deepkseek on their site, it actually feel better to talk with

[rustic island]

Still feels like Gemini from Temu to me

[jovial kelp]

Now this could be bias, but deepseek act more like someone who didn't really have opinion on what being given to them and just breaking it down while the other model that i use as the counterpart (It's gemini 3.1 pro) being more opinionated and being as of they are professional in that field.

Btw i do the talk with my native language, so it could be different depend on the languages.

At the end of the day, the data is what make them act the way they are acting.

[rustic island]

Hm, haven't tried to speak in mine to it yet

[jovial kelp]

Man, i just realize that this thread created on january this year.
We have been edging on the new deepseek for 4months at this point

[supple sigil]

4… months?

4?

like ds v4

is this a deepseek v4 reference?????

[jovial kelp]

https://tenor.com/view/yu-narukami-dancing-persona4-specialist-video-game-gif-17873794

*persona song playing in the background

[feral scaffold]

Still betting on v4 to reheat its own nachos

[loud verge]

They're getting the UI ready for the new models 👀

They merged Deepseek chat and deepseek-reasoner into one graph(They were different before)

[hot swan]

soon ™

[loud verge]

Hopefully today

https://klipy.com/gifs/copium-copium-od

[frank wind]

happening

[hot swan]

my tea leaves are all aligned and pointing up

Uranus is in Gemini

[supple sigil]

@gemini is this true?

[covert topaz]

deepseek today

[ancient gulch]

@grok change deepseek clothes to bikini

[thin bramble]

https://tenor.com/view/old-snake-metal-gear-solid-snake-time-goes-by-fast-metal-gear-solid-gif-26264117

[cloud flame]

https://tenor.com/view/nso-needy-streamer-overload-needy-streamer-girl-gif-17743324039319069122

[abstract dragon]

https://x.com/i/status/2046855733288067266

永雏塔菲 (@xhyctf)

ds的官方api更新了

X · Apr 22, 2026, 7:36 AM

THE WHALE HAS AWAKENED

[cloud flame]

Put it down

[flat osprey]

nothing is real until i see deepseek announce V4

i've played these games before

[abstract dragon]

https://tenor.com/view/i-have-played-these-games-before-seong-gi-hun-squid-game-i've-played-these-games-already-netflix-gif-16480153790313233198

[abstract dragon]

https://tenor.com/view/airball-air-ball-ball-basketball-hope-gif-16067754010491213168

[sharp vortex]

Can someone translate this 🥀

[sharp vortex]

Ty

Can it actually handle 1m token

[peak swallow]

trying in openwebui. it is currently crashing due to the paste

[sharp vortex]

Try in openrouter fr

[peak swallow]

it didnt let me in openrouter since they limited it to 128k tokens on openrouters side

so i need to use deepseek api directly

the api still hasn't responded but it didnt error

[hoary zenith]

yeah I think v4 lite is in API, seems to be a bit better than Instant

[hoary zenith]

hmm still doesn't seem optimized for agentic coding, in fact still don't see much difference in terms of overthinking and unnecessary tool calls

[frank wind]

real?

[peak swallow]

I gave it 400k token file and it was not able to respond but it didnt error so something has changed

[hoary zenith]

yeah it's a lot faster than before, reasoning in general seems more concise too (basically instant from web, but a bit better on some prompts)

[frank wind]

wait

reasoner or chat?

or both?

[hoary zenith]

reasoner, haven't tried chat yet

[frank wind]

I need to use the completion endpoint

FUUUUCK

[hoary zenith]

would be funny if this is v3.3

[cloud flame]

We are so back it's over

[fringe flicker]

Is it shit? I will try later

[sharp vortex]

Deepseek today

[hoary zenith]

and it's gone, tbh A/B testing on the API is a bit tinpot

[green trellis]

when will they stop edging us

[pastel bluff]

At this point I am half convinced that DS V4 will either be mediocre as heck or the open source Mythos

[haughty pilot]

its hilarious.

[hoary zenith]

What model do you consider to be the baseline that defines mediocrity best?

[long osprey]

if life has let me know something, is don't have that much expectations, have healthy ones but nothing extreme

[haughty pilot]

deepseek is available NOW
see here

[pastel bluff]

Damn, hard question.

Grok maybe.

It's just not very impressive at much except NSFW from what I've seen.

[hoary zenith]

yeah the whale better get some more deepsleep if can't beat grok lol

[haughty pilot]

deepseek 💖🐳🍋 >>> grok 🤡😡🤬

[pastel bluff]

Let's hope so xD

[fringe flicker]

Grok being impressive at nsfw is a dank meme

Grok writing sucks

[pastel bluff]

It does. It's unrestricted, but actually pretty bad at NSFW if you don't just accept slop.

[sharp vortex]

Mods, mute this guy please /j

[pure flax]

to be fair apparently grok is a 500B model

[pastel bluff]

Gemma 4 is 31b model.

[pure flax]

sure, grok is clearly super specialized though

[pastel bluff]

||At gooning?||

Sorry I had to.

What is its stated specialization because I've honestly never seen that?

[pure flax]

looking up / cross referencing stuff

"deep search"

[pastel bluff]

Hmmm, alright, fair. It is pretty good at that.

[pure flax]

And its EXTREMELY fast. So I have to assume its very sparse as well

[flat osprey]

not a deepseek announcement, can't believe it

stay strong chat

✊

[pure flax]

that guy just talks it seems, he said this awhile ago

so just more BS

[hoary zenith]

It makes sense if you consider they forgot to put 1 before 600, they deleted the tweet immediately. Also they talked about big v4 being 1.6T weeks ago

[pure flax]

also "optimizer: Muon"

muon is terrible

[hoary zenith]

muon is what everybody uses (glm/kimi), openai literally hired the muon guy so gpt too

[pure flax]

its faster but less accurate

not worth it

The only one I would recommend alongside adam is CAME https://www.zangwei.dev/blog/proj-came

https://github.com/Nerogar/OneTrainer/pull/798
https://github.com/xzuyn/CAME/commits/triton-fused-tempbranch2

CAME has the accuracy of adam with the memory requirements of adafactor

yea, I know the name is silly

[hoary zenith]

well lot of things are good in theory and papers, but not in reality of actual large scale training

[pure flax]

I guess. I've only done stuff with datasets up to hundreds of thousands. And it was image models

I just know side by side I had worse results

Maybe a hybrid would work best

that might be what they are doing

[hoary zenith]

according to GLM-5.1:

The vast majority of LLM parameters are in 2D weight matrices (QKV projections, MLP up/down/gate). Muon is designed for exactly this. Image models have convolutional kernels (4D tensors that need reshaping), more normalization parameters, and a different structural mix. Applying Muon to non-matrix parameters requires fallback to another optimizer, and the interaction between the two regimes can be awkward.

[pure flax]

AdaMuon basicily

[hoary zenith]

Anyway that list basically described things already in the NSA paper, from February last year

it must have been a training run from hell if it took them so long to get that work at scale

[pure flax]

Lol I might make a hybrid and name it CAMEMuon

came-on, get it?

[plucky ermine]

Perhaps something to indicate they were Unified ;]

[jovial kelp]

Are they announce it?

[pure flax]

elon outright leaked it

[thin bramble]

when you say "specialized" and "500b" people might laugh at the statement.

https://tenor.com/view/epstein-scratches-chin-jeffrey-gif-15670537269881966567

[thin bramble]

the size changed since last leak + it is not multimodel according to this. is it even a reputable leaker?

[covert topaz]

we’ll see this week

[thin bramble]

https://tenor.com/view/cat-cats-pet-cat-cat-pet-cute-cat-gif-24810247

[sharp vortex]

Deepseek T—

[west shell]

Deepseek tomorrow

[cloud flame]

https://tenor.com/view/deepseek-aı-agent-chına-usa-gif-5407811775974993155

[tulip estuary]

DeepSeek? yeah my Drip do be Sick 🦦🔥

[covert topaz]

deepseek is cuming

[short jasper]

when

today

?

[covert topaz]

tomorrow

[elfin sparrow]

is V4 multimodal?

will it be able to create images and videos?

[tulip estuary]

nobody knows anything

[elfin sparrow]

they kinda need to otherwise they are wayy back in the AI race

[tulip estuary]

it supposedly will have image input yeah

[tulip estuary]

they are very experimental and focus more on efficiency, even if it's not the absolute best

imagine having >90% of Opus for about a dolar/M out

[covert topaz]

the "leak" says text only

[cloud flame]

https://tenor.com/view/spongebob-leak-patrick-plumber-tired-gif-12733093

[kindred gazelle]

https://api-docs.deepseek.com/ its out

pro and flash

[inner pike]

yep

DEEPSEEK OUT

called for deepseek-chat in the beta completions endpoint and this is what returned

[vapid karma]

DeepSeek V4 today

[hot swan]

1.6T as announced

[sacred glade]

Where is this?

[rare gale]

Interestingly seems to be text only

[shut oasis]

LADS

IFS HERE

[flat osprey]

OMG

IS THIS REAL?????

[vale kayak]

yeh

[flat osprey]

IS THE WAIT OVER????

[vale kayak]

https://api-docs.deepseek.com/

[quick bison]

its fine i guess, atleast its smart and hopefully still cheap. needs 1tb drive to host lol

[vale kayak]

i see deep seek v4pro and lite there

[neon radish]

Official released

[flat osprey]

prayed for times like this

ragebait is finally over

https://tenor.com/view/please7tv-please-beg-pray-hope-gif-3860103425950934673

[vapid karma]

https://huggingface.co/deepseek-ai

[rare gale]

it begins

[flat osprey]

https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro

[vapid karma]

Is this the first open model which has actual reasoning effort settings?

[sacred glade]

is it good?

[flat osprey]

it's beautiful 🥹

[hot swan]

great for the price but it feels less impressive after the other chinese model releases

[rustic island]

No way

[vapid karma]

Presumably this has the engram stuff too right?

[rare gale]

These numbers don't look awfully benchmaxxed so we'll have to give it a spin

[deft crow]

working on them

[hoary zenith]

https://x.com/nopainkiller/status/2047512431228756122#m

Zhipeng Huang (@nopainkiller)

@teortaxesTex They promised price dropping when deployment on Ascend 950 later this year

X · Apr 24, 2026, 3:06 AM

[vapid karma]

DeepSeek distills but they at least do a lot of their own work too

[hot swan]

benchmaxxed has stopped meaning anything years ago

if everything is benchmaxxed that's just the baseline

[rare gale]

It's all still MIT which is very nice

[flat osprey]

can't wait to try 🙏

[vale kayak]

@covert topaz

[hot swan]

i have to sleep and go to work sadly
they must have timed this just to spite me

[rare gale]

Preview eh

[vapid karma]

Damn 2900 replies before the official drop

[flat osprey]

@lucid ocean if you're still active you should rename this since it's released now lol

if not maybe a mod can

[hoary zenith]

DeepSeek_V4.pdf, ctrl-f engram, showing results: 0/0

[flat osprey]

https://x.com/ValsAI/status/2047513613750202452

Vals AI (@ValsAI)

DeepSeek v4 is now the #1 open-weight model on our Vibe Code Benchmark, and it’s not close.

It leaves the #2 (Kimi K2.6) in the dust, and even beats out frontier closed source models like Gemini 3.1 Pro.

X · Apr 24, 2026, 3:11 AM

[jovial kelp]

862B parameters

[flat osprey]

pretty cool independent benchmark

[jovial kelp]

A little bit more and they reach 1T

[rare gale]

Okay now

[hoary zenith]

DeepSeek-V4-Pro-Max demonstrates superior performance relative to GPT-5.2 and Gemini-3.0-Pro on standard reasoning benchmarks. Nevertheless, its performance falls marginally short of GPT-5.4 and Gemini3.1-Pro, suggesting a developmental trajectory that trails state-of-the-art frontier models by approximately 3 to 6 months.

refreshingly candid

[jovial kelp]

Wait, it's 1.6T?

[ancient gulch]

N-n-no way... Im sobbing

[rare gale]

Correct

[jovial kelp]

Why safetensor showing it as 800B

[rare gale]

Pro is

[jovial kelp]

.

Is it like 862B is the normal transformer then the other are their Ngram stuffs

[gusty sphinx]

agentic coder Deepseek huh?

[jovial kelp]

This team really like experimenting bro

[flat osprey]

https://x.com/deepseek_ai/status/2047516922263285776

DeepSeek (@deepseek_ai)

🚀 DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length.

🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models.
🔹 DeepSeek-V4-Flash: 284B total / 13B active params.

X · Apr 24, 2026, 3:24 AM

officially posted