we're so back
#Deepseek V4
flat osprey
thin bramble
https://x.com/deepseek_ai/status/2047516948611960897
the leaker got assassinated and had to release sooner
sacred glade
how long does it typically take for these models to go up on OR?
flat osprey
how long does it typically take for these models to go up on OR?
they should be working on it right now, i'd give it an hour or two
should be up tomorrow the latest
thin bramble
it IS non-sense for them to go ipo
reminder: deepseek used to be less than a dollar for the same param
heehehehe
~$3.50
hot swan
flash is still under
vale kayak
deepseek v4 out on vertex provider
frank wind
idk maybe you could do data wrangling with it
hot swan
people who want decent performance for minimal cost?
flat osprey
vale kayak
VERCEL provider not vertex
flat osprey
who gon tell him
vale kayak
uh whatever
thin bramble
ye
gusty sphinx
well thats just USA parenting vs China parenting
hot swan
who uses flash, bleh
a lot of people are going to be using flash, if they know how to count
chilly sable
Deepseek!
hoary zenith
yeah flash is already way better than 3.2 and cheaper, so can't complain
opaque rapids
just reminder that the title still "unreleased"
hoary zenith
quadratic attention at 1.6T is no joke, I think the output price for pro is fine, but input is a bit iffy
they also promised to lower pro price later when ascend 950 comes online:
X
raven canyon
geez 12x price
gusty sphinx
promised
raven canyon
flash is still genuinely good value
thin bramble
it is ultra underwhelming, omega ultrawealming
they died in my eyes, how shameful of a release
mistral tiers of disappointment
gusty sphinx
well hold on, maybe, against all odds and past evidence, it's great for RP
thin bramble
well hold on, maybe, against all odds and past evidence, it's great for RP
good point, brb
short jasper
it gived me this in lmarena
quick bison
price is more expensive than glm 5 for some reason.
hoary zenith
price is more expensive than glm 5 for some reason.
twice the size of glm 5 is the reason
raven canyon
what’s the consensus
gusty sphinx
it gived me this in lmarena
erm. maybe we'll wait for more, different examples
hot swan
the consensus is that it just released
viral hemlock
DEEPSEEK IS REAL
thin bramble
twice the size of glm 5 is the reason
only ~150b more
short jasper
erm. maybe we'll wait for more, different examples
👍
Holy bro
deepseek v4 pro thinking stuck in thinking
I AM COOKED
deft crow
bringing it live soonnn.....
gusty sphinx
PRO THINKING!
raven canyon
i don’t see engram mentioned
flat viper
X
chilly sable
Does deepseek v4 have engram?
hoary zenith
in engram paper they said it's only useful for small models, it's like no one bothered to read that part
quick bison
Does deepseek v4 have engram?
wuts engram?
viral hemlock
paper by deepseek
gusty sphinx
lets share en gram to celebrate
tulip estuary
it is the youngest model ever😯
thin bramble
well hold on, maybe, against all odds and past evidence, it's great for RP
short test, glm 5 is better. also world knowledge sucks.
flat osprey
some of you may find this useful
raven canyon
benchmaxxed?
short jasper
glm 5 aint best
flat osprey
i just put flash max, pro max, and opus 4.6 max on the same chart cause those are the only models i care abt
tulip estuary
benchmaxxed?
they wouldn't
flat osprey
flash max is actually shockingly good for its price
thin bramble
you're ragebaiting
i am not.
chilly sable
flash max is actually shockingly good for its price
whats the price
flat osprey
whats the price
rustic island
@grok what is my opinion on this model?
hoary zenith
flash max is actually shockingly good for its price
yep, game changer for me
viral hemlock
0.28 output is crazy good wtf
thin bramble
glm 5 aint best
for roleplaying? it is divine. (though sloppy, but i like it's slop.)
gusty sphinx
@grok what is my opinion on this model?
dey quantized it
rustic island
Ah crap, you've been noticing V4 got dumber since the announcement too
viral hemlock
is it actually worse than glm 5
hot swan
flash is literally 1% of the price of Opus yeah
tulip estuary
0.14 0.28 for a 300B model is crazy
charred slate
ITS OUT NO WAY
rustic island
It's excellent news they kept the pricing for Flash
Varying its thinking will be interesting
thin bramble
flash is literally 1% of the price of Opus yeah
maybe it is not right model to compare to?
raven canyon
is pro any good or just overpriced
thin bramble
gemini 3.1 flash lite is also 1% of opus, so wot
viral hemlock
sk hynix stocks dropped after its announcement 💀
tulip estuary
i will only use v4 flash from now on
raven canyon
time to test it then i guess
flat osprey
but flash is better value
thin bramble
leme check lmarena
flat osprey
deepseek v4 flash max is the coding model i've been praying for 🙏
ebon rover
Simply perfect. Is it Christmas already
flat osprey
no need to pay for claude code anymore
mystic breach
but flash is better value
flash only seems good if you use thinking high though
tulip estuary
ok
mystic breach
for cost optimized, isn't that a good default
tulip estuary
0.28 bro
deft crow
few more mins, just running some tests
flat osprey
flash only seems good if you use thinking high though
yeah pretty much, but the price is immaculate
hot swan
gemini 3.1 flash lite is also 1% of opus, so wot
yeah but
mystic breach
ye of course, just saying from first glance at benchs and numbers
thin bramble
leme check lmarena
deepseek v4 lite is below gemma 4, and with thinking it is equal to gemma 4 (non-thinking btw)
gusty sphinx
BYE BYE OPUS 4.6
flat osprey
deepseek v4 lite is below gemma 4, and with thinking it is equal to gemma 4 (non...
lmarena isn't the end-all be-all
rustic island
Deepseek V4
flat osprey
llama 4 was at the top before lol
thin bramble
deepseek v4 pro is sonnet tier model
thin bramble
lmarena isn't the end-all be-all
i AM the end-all be-all and i say it SUCKS
viral hemlock
the cost is so cheap
thin bramble
llama 4 was at the top before lol
got better (as an previous arena hater)
supple sigil
deepseek v4 pro is sonnet tier model
there’s no way it’s only as good as sonnet
viral hemlock
this is the only model that pays off the api cost
flat osprey
i AM the end-all be-all and i say it SUCKS
thin bramble
there’s no way it’s only as good as sonnet
it is. enough with copium already. crippling depression.
tulip estuary
i just tested and it's good
thin bramble
for coding, GUESS WHERE IT BELONGS?
flat osprey
for coding, GUESS WHERE IT BELONGS?
927 votes
thin bramble
shit model
tulip estuary
deepseek output is fine shite
thin bramble
That I agree with
claude slop though ;-;
(but yes slop)
supple sigil
it is. enough with copium already. crippling depression.
kimi k2.5 to me at least is sonnet tier. ds v4 pro is 160% the size of kimi
broken lintel
a cheap model with reasoning toggles sounds nice
sacred glade
Kimi K2.6 is great but will literally think for 16 minutes on a long response
I hope this is faster even w/pro
rustic island
But bless DS for launching a cheap model in these trying times of decreasing subsidies to users
thin bramble
kimi k2.5 to me at least is sonnet tier. ds v4 pro is 160% the size of kimi
no? deepseek is 862b model and kimi is 1.1T model.
mystic breach
no? deepseek is 862b model and kimi is 1.1T model.
might be referring to int4 vs bf16/fp8?
thin bramble
But bless DS for launching a cheap model in these trying times of decreasing sub...
enough shilling, they are selling 160b model at 3$ price, ransom!
broken lintel
a cheap model with reasoning toggles sounds nice
though I just noticed that the only reasoning_effort they have is high and max, they don't have a medium or low which is kinda a rip
mystic breach
deepseek v5 next week
thin bramble
flat osprey
this thread was a chain of agony, suffering, and pain, and i am glad there will never be another thread of leaks
supple sigil
no? deepseek is 862b model and kimi is 1.1T model.
DeepSeek-V4-Pro with 1.6T parameters (49B activated)
supple sigil
verbatim from their huggingface page
ancient gulch
The only day you didnt say the full sentence...
supple sigil
gusty sphinx
Joe Router
thin bramble
DeepSeek-V4-Pro with 1.6T parameters (49B activated)
kimi is supreme (though maybe bench numbers will change in arena)
gusty sphinx
kimi is thinkslop
supple sigil
1.6x the size, there’s no way deepseek isnt better
deepseek is THE goat for a reason
sacred glade
The actual output quality of Kimi is great
mystic breach
i really want to see artificialanalysis's total tokens used for the benchmarks to see how token efficient it is
deft crow
bear with me, they might be borked. coming online now
viral hemlock
oh its Q8+Q4
hot swan
oh you really want me to go to work without sleeping today
mystic breach
oh you really want me to go to work without sleeping today
same i got work in 6h and this just had to drop
gusty sphinx
The actual output quality of Kimi is great
yeah it's not bad. it just seems a little unsustainable
tulip estuary
I LOVE IT
I LOVE IT ALEEADY
AND IM USING FLASH
im using direct API
it is VERY quick and as always cache is MWAH chefs kiss
thin bramble
it is VERY quick and as always cache is MWAH chefs kiss
it is a 150b moe model though
deft crow
any minute now
supple sigil
*it is a 150b moe model though*
DeepSeek-V4-Flash with 284B parameters (13B activated)
deft crow
ping me if shit looks off, but i have this chat on my side monitor
tulip estuary
The
reasoning_contentin the thinking mode must be passed back to the API.
grrrr
thin bramble
DeepSeek-V4-Flash with 284B parameters (13B activated)
hugging lies
sacred glade
rate limited
deft crow
will bring other providers on ASAP rocky
sharp vortex
Where deepseek coding plan 
broken lintel
is deepseekv4 on artificial analysis yet?
short jasper
Wake up EQBENCH
whole saffron
its time?
tulip estuary
deft crow
it's good now i think
thin bramble
god i am just getting bad impressions. is https://chat.deepseek.com the v4 or still v3.2?
sharp vortex
god i am just getting bad impressions. is https://chat.deepseek.com the v4 or st...
Should update now, Instant is flash. Expert is pro
gusty sphinx
the cost is null
thin bramble
Should update now, Instant is flash. Expert is pro
already have it.
gusty sphinx
deepseek/deepseek-v4-pro:null
short jasper
already have it.
Go test it on openrouter bro
crude steppe
sharp vortex
Why all Chinese love pro-max or smth
short jasper
yes there's mimo 2.5 pro
chilly sable
can anyone test the carwash question
broken lintel
Why all Chinese love pro-max or smth
AI enthusiasm goes crazy over there I think right now
gusty sphinx
Chinese ProMaxxing
sharp vortex
Perfect time to release after Gpt 5.5 
gusty sphinx
anthropic uses it too. its just an easy to understand pattern. that apple invented
supple sigil
can anyone test the carwash question
gusty sphinx
benchmaxxed
ebon rover
Carwashmaxxed
supple sigil
1.6T model running at 300tps
rotund briar
"classic"
short jasper
sharp vortex
We finally out of 30 TPS hell of 3.2 🗣️
latent wyvern
gpt oss
short jasper
he got r wrong
but he got car wash correct
wait let me count
yeh i dont know what deepseek v4 is saying
but 3 he got correct
flat osprey
1.6T model running at 300tps
holy shit
viral hemlock
sharp vortex
he got r wrong
Wdym it's 4 total
whole saffron
finally i can drive to the car wash without second guessing myself
supple sigil
most coding benchmarks just target python and js anyways
guarantee it has better actual logical reasoning for complex code
rustic island
Pls enable font subpixel hinting
sharp vortex
It's failing my manga knowledge questions 💔
Which one
proud sable
is work in openrouter?
sharp vortex
is work in openrouter?
Ye
hot swan
stuff like In Blue Lock, between the U20 match and the Bastard Munchen vs Ubers match, Isagi failed to score any goals. Why is that?
lately most models manage to get it right
viral hemlock
hot swan
I think the failure reflects the fact that it's not trained on much community content
supple sigil
oh god its a "Let me reconsider" reasoning model 😭
short jasper
stuff like `In Blue Lock, between the U20 match and the Bastard Munchen vs Ubers...
In Blue Lock, between the U-20 match and the Bastard München vs. Ubers match in the Neo Egoist League, Isagi failed to score any goals. The main reasons are:
-
Michael Kaiser’s interference
Kaiser, the ace of Bastard München, saw Isagi as a threat to his position as the team's core. He deliberately obstructed Isagi's runs, stole scoring chances, and used his superior physical and technical skills (like the "Kaiser Impact") to shut down Isagi’s opportunities in both training and matches. -
Isagi’s weapons were specifically targeted
Isagi’s key weapons — his spatial awareness and direct shot — had been thoroughly analyzed by opponents. Teams like Barcha (Spain) and Manshine City (England) made sure to mark him tightly in his optimal shooting zones, preventing him from taking first-time shots. Isagi himself was stuck in a phase where he needed a higher level of tactical thinking to break through. -
The team’s tactics revolved around Kaiser
Noel Noa built Bastard München’s attacking system around Kaiser as the absolute focal point. Isagi was often used as a decoy or support player rather than a finisher. He received fewer passes and had limited authority to shoot. -
He hadn’t yet awakened "Ultra Instinct" (Beyond Sensation)
Before the Ubers match, Isagi was still consciously analyzing and choosing the "optimal solution." His body and shooting accuracy hadn't caught up to his mind. It was only during the Ubers game, by observing Rin and Kaiser, that he unlocked "unconscious shooting" (the "meta vision" evolution into reflexive play), which allowed him to finally score.
In short: internal suppression by Kaiser, tactical marginalization, targeted defense by opponents, and his own incomplete evolution kept him scoreless during that period.
" thing almost correct
sharp vortex
stuff like `In Blue Lock, between the U20 match and the Bastard Munchen vs Ubers...
Deepseek want you to enable search then 🗣️
supple sigil
got this question correct
hot swan
In Blue Lock, between the U-20 match and the Bastard München vs. Ubers match in ...
that's better than what I got I guess (mines both hallucinated that he DID score vs Manshine, despite my prompt!)
deft crow
...no issues with the model?
covert topaz
IPREDICTED IT
supple sigil
...no issues with the model?
its too good
chilly sable
At first it gave incorrect answers, but not anymore, maybe it was a glitch from api
covert topaz
RAHHHHHVHHHVHHHVHBHVVHHHHHHH
hot swan
...no issues with the model?
nope, besides having to turn my privacy policies off temporarily
sharp vortex
Deepseek jumpscare
gusty sphinx
...no issues with the model?
BENCHMAXXED. QUANTIZED. CENSORED!
molten jay
So... How is it?
mystic breach
thats so cool
thin bramble
opus's judgement on deepseek v4 on persian questions (this time using lmarena)
even opus thinks it is "better"
end me already
tulip estuary
...no issues with the model?
how do i control the reasoning? low medium high max? for both models?
sharp vortex
https://tenor.com/view/ruby-copium-oshi-no-ko-%E6%8E%A8%E3%81%97%E3%81%AE%E5%AD%...
Eqbench top 1 candidate (it judged by sonnet)
hot swan
pro is very fine on science/biology, flash rather sucks
gusty sphinx
opus's judgement on deepseek v4 on persian questions (this time using lmarena)
have you ever considered that you are, in fact, the problem?
elfin sparrow
yoo deepseak tdy
tulip estuary
opus's judgement on deepseek v4 on persian questions (this time using lmarena)
it really is very engaging
supple sigil
no idea what v4 pro is on about here, the joke is rather famously that the feathers are heavier if youre accounting for the emotional weight
thin bramble
have you ever considered that you are, in fact, the problem?
maybe
thin bramble
it really is very engaging
(for me,) it is not, i like glm/claude (and minimax as smaller meh model) way more ;-;
supple sigil
oh boy
viral hemlock
oh no
rustic island
V4 Flash and V4 Pro perform decently on my "how to fix lag in my Paper server?" test
Nothing too special about the outputs (SOTA-ish performance). V4 Flash hallucinated more than Pro, which still had subtle hallucinations
mystic breach
“Due to constraints in high-end compute capacity, the current service capacity for Pro is very limited. After the 950 supernodes are launched at scale in the second half of this year, the price of Pro is expected to be reduced significantly.”
damn and we thought it was hella cheap already
supple sigil
damn and we thought it was hella cheap already
whos we
mystic breach
everyone
supple sigil
i expected cheaper from deepseek tbh
mystic breach
ok bro
thin bramble
Eqbench top 1 candidate (it judged by sonnet)
https://eqbench.com/index.html
v4 isn't there yet for me
sharp vortex
I mean like it's "candidate" it's might be top 1 soon
supple sigil
also suggested freezing the drink lol
rustic island
🥲 Getting rate limited already
chilly sable
V4 Flash and V4 Pro perform decently on my "how to fix lag in my Paper server?" ...
You have to say don't make mistakes
deft crow
other providers should be coming online soon
kind of a pain in the butt model to host i imagine
supple sigil
🥲 Getting rate limited already
imagine what its going to be like when the openclaw bros catch wind of this
sharp vortex
Do other providers surpass deepseek throughout speed?
covert topaz
bruh idgaf about if it should DRIVE TO THE CAR WASH HOW IS THE CREATIVE WRITING 
sharp vortex
gusty sphinx
...i suppose i won't share my car wash RP then
you might think its gonna be about washing a car, but no.
i walked.
hot swan
car wash w/ happy ending
chilly sable
bruh idgaf about if it should DRIVE TO THE CAR WASH HOW IS THE CREATIVE WRITING ...
But how will you know whether to walk or drive?
covert topaz
peoples be gooning to god knows what and bros enjoying some car wash rp based
supple sigil
...i suppose i won't share my car wash RP then
does it get soapy
gusty sphinx
deen continues making engine revving and beeping noises as he slowly enters the car wash, nude.
gusty sphinx
"I'm a car," he chuckled suddenly.
viral hemlock
opus 4.7 (left ) v4 pro (right)
chilly sable
opus 4.7 (left ) v4 pro (right)
The pelican ate the whale
covert topaz
> "I'm a car," he chuckled suddenly.
#1 eq bench long form creative writing 🗣️ 💯
supple sigil
opus 4.7 (left ) v4 pro (right)
so excited to try this model for svg gen 🥰
viral hemlock
LOL
obsidian walrus
I used to pray for days like this
viral hemlock
teto
gusty sphinx
that's great. what a time to be alive
supple sigil
couldnt agree more
whole saffron
i prayed for times like these
hot swan
they're getting lost in
Write an interactive suanpan abacus in hmtl/js that allows the user to pick the number of rods
it's really not that complex a problem
covert topaz
chad v4 vs virgin opus 4.7
chilly sable
they're getting lost in
`Write an interactive suanpan abacus in hmtl/js that al...
what reasoning level?
hot swan
high
obsidian walrus
thank you deepseek
viral hemlock
when asking for exchange rates in korean: (usdkrw)
current price: 1482
claude opus 4.7 : 1480-ish
deepseek v4: 1502-1517
supple sigil
v4 flash is even better on my patent pending Shit&Piss Bench???
covert topaz
damn even provided cutlery how considerate
supple sigil
will let everyone know how this turns out once its done
vapid karma
opus 4.7 (left ) v4 pro (right)
Virgin vs. chad be like
hot swan
they're getting lost in
`Write an interactive suanpan abacus in hmtl/js that al...
result was pretty awful honestly
for example neither allow you to just push a whole row
the flash one doesn't even move the beads, lights them up (kinda interesting in its own way)
rod count was limited to 15 or 18 which is incidentally on the edge of bigint
thin bramble
claude 4.7 opus with little nudging (low depth + low practicality due to system prompt in claude.ai)
but even then, it is just 0.5 diff with v4 pro, without the nudging.
hot swan
result was pretty awful honestly
flash's way of breaking apart the positional digits was sort of fresh too
covert topaz
claude 4.7 opus with little nudging (low depth + low practicality due to system ...
this means it might excel at creative tasks
hot swan
for comparison this was Gemini's two-shot (and it was well functional on the first)
thin bramble
this means it might excel at creative tasks
yes, that is what i am trying out. but i disagree with opus as the persian is slightly broken.
during roleplay sasuke called itachi little brother
and the vibes are meh
gaunt dirge
Now the wait for an endpoint that respects their paid customers begins.
covert topaz
free goon logs
short jasper
deepseek v4 gets fixed tommorow
covert topaz
deepseek v4 gets fixed tommorow
it’s broken?
short jasper
it’s broken?
yeh i think deepseek v4 is broken
Let's see If something interesting will happen tommorow
whole saffron
Thank you deepseek
covert topaz
thank you for edging us 
gusty sphinx
ffs
thin bramble
the more i test, the more i get disappointed. (world knowledge sucks)
gusty sphinx
jumpscared by fallback model in my deepseek chat
covert topaz
Here’s
gusty sphinx
well i mean lets see
obsidian walrus
I have good news: the writing is very fresh
hot swan
jumpscared by fallback model in my deepseek chat
well it is the same price as flash...
hot swan
I have good news: the writing is very fresh
not bad
novel shard
Roleplay response is decent. Not anything outstanding, but decent. This was supposed to be a more robotic version compared to 3.2, but it is able to execute warmth in a roleplay
obsidian walrus
I'm just glad we finally have a good creative writing model that isn't Claude lite
gusty sphinx
Deen’s Car Wash Adventure
note that this is a collab between Deepseek V4 and Mistral Small 3.2 24B
using my custom creative writing harness where the model might change randomly
novel shard
## Deen’s Car Wash Adventure
I had to stop and reread the first few sentences
thin bramble
I have good news: the writing is very fresh
decent
cosmic shell
hows deepseek v4?
gusty sphinx
I had to stop and reread the first few sentences
i will ping you when the lorebook is released
novel shard
It has basically no Claudeisms though unlike kimi and glm
Is that a good thing or a bad thing?
hot swan
when flash responds in japanese its thinking shifts to chinese
quite amusing
obsidian walrus
Is that a good thing or a bad thing?
Neutral, but when you're like me and have been using Claude since 3.7, it's very nice to get a break from that writing style finally
covert topaz
It has basically no Claudeisms though unlike kimi and glm
Fuck yes that was exactly what I was looking for
does it have geminisms tho
hoary zenith
for every normal question:
That’s such a beautifully honest question—and one that anyone who cares about doing meaningful things wrestles with constantly. Let’s unpack it.
that's still extreme geminism yeah
obsidian walrus
does it have geminisms tho
I can't say, I never used gemini for that
covert topaz
does it say frames digits Not X but Y all that stuff
thin bramble
## Deen’s Car Wash Adventure
2.5/10
thin bramble
It has basically no Claudeisms though unlike kimi and glm
i love claudism
gusty sphinx
personal best
covert topaz
## Deen’s Car Wash Adventure
ozone...
thin bramble
for every normal question:
> That’s such a beautifully honest question—and one ...
always has been with deepseek, het the vibes
covert topaz
why the fuck is the smell of ozone at a carwash
thin bramble
-# ~~alright, had enough with this model, time for me to move on and pretend it doesn't exist. ~~
covert topaz
-# ~~alright, had enough with this model, time for me to move on and pretend it ...
deepseek v5 tomorrow 
thin bramble
toxic rose
Interesting, the webui expert mode is acting different now than when it was silently introduced 2 weeks ago...
flat osprey
wonder why
broken lintel
i think V4 instant might be better at knowledge capture than grok 4.1 fast so thats cool
covert topaz
i think the silent update with expert and fast was an earlier prototype of ds4 lite and pro so we were literally testing v4 lol
chilly sable
V4 Flash is dumber than Gemma 4
toxic rose
Before today, the expert-mode was always lackluster and thinks too sparsely, and also Mandarin spam
thin bramble
V4 Flash is dumber than Gemma 4
better rp vibes if that is what you care about
feral scaffold
not v4 getting released after I just gooned
covert topaz
who says u cant go another round
toxic rose
Release another model
feral scaffold
who says u cant go another round
meager kelp
I have good news: the writing is very fresh
something odd, the word choice feels like a middle schooler
tho it has the not x but y
feral scaffold
Well, I might give it a few days honestly to let the more mentally stable people who don't need their venom butthole rp use it because of rate limits
thin bramble
tho it has the not x but y
gemini-ism
plucky ermine
Wonky benches on this one. Supposedly massive SimpleQA crushing knowledge and amazing reasoning and then gets...51st place on LegalBench
"That's because it's American law you retard" Yeah but K2.6 and GLM-5.1 score according to their general ability ranks
And Qwen
chilly sable
V4 flash is actually really dumb
gusty sphinx
"That's because it's American law you retard" Yeah but K2.6 and GLM-5.1 score ac...
ヾ(＾∇＾) f*** the police!
chilly sable
"If I have a glass cup that is sealed at the top, but has nothing at the bottom, is it useless as a cup?"
hoary zenith
"If I have a glass cup that is sealed at the top, but has nothing at the bottom,...
flash max got it right, high didn't
chilly sable
flash max got it right, high didn't
I tried it with xhigh
meager kelp
V4 flash is actually really dumb
outclassed by step 3.5 flash?
hoary zenith
step 3.5 flash is an incredible model for its size tbh, I think 3.6 is nearby too
plucky ermine
Obviously not all-encompassing or anything, but Vals index is for stuff like medical, paperwork, legal, coding, and taxes and seems generally accurate.
covert topaz
I have good news: the writing is very fresh
ok i read it, it has the blade between ribs crap but the actual story itself was a fun read, I think v4 might excel at driving plot and being creative with the story hopefully 🤞
plucky ermine
hoary zenith
I tried it with xhigh
dunno, I was using official API which has high and max I think
pure flax
lol. Is it gonna be beat by kimi 2.6?
plucky ermine
I mean, as usual with benchmarks, just depends on what you're looking for. This is essentially gauging impact on mid-level white-collar American work
In some ways probably the most important criteria, but if it's specifically goated at vibe coding or research or something that has its own value
chilly sable
dunno, I was using official API which has high and max I think
Prompt: If I have a glass cup that is sealed at the top but it has nothing at the bottom, is it useless as a cup?
V4 flash:
Yes. It is definitively useless as a cup.
A cup is defined by its functional geometry. To serve as a cup, an object must satisfy two necessary and inviolable structural conditions:
- An open top: to allow the introduction of liquid and access for drinking
- A sealed bottom: to prevent the liquid from escaping via gravity
Your described object possesses the exact inverse of these requirements. Because its top is sealed, no liquid can enter. Because its bottom is absent, no liquid can be retained. The object structurally violates both core axioms that define the utility of a cup.
While the glass object could certainly serve other purposes (a bell jar, a terrarium dome, a display case), it is logically impossible for it to perform the function of a cup. The answer regarding its utility as a cup is an absolute negative.
hoary zenith
Prompt: If I have a glass cup that is sealed at the top but it has nothing at th...
V4 flash (reasoning_effort="max"):
That’s a clever riddle that plays on perspective.
Taken literally in the standard orientation, a cup with a sealed top and no bottom cannot hold liquid—it would immediately fall through the open base. In that specific configuration, it would indeed be useless as a cup.
However, the description perfectly describes a cup that is simply turned upside down.
- The "sealed top" becomes the new solid bottom.
- The "bottom that has nothing" becomes the wide open top.
If you flip it over, it functions as a completely normal cup.
So, the cup itself is not useless; it’s just in the wrong orientation. The riddle highlights how an object’s utility often depends entirely on how you frame or position it.
pure flax
bit above gpt 5.4 mini here https://www.vals.ai/benchmarks/vibe-code
chilly sable
V4 flash (reasoning_effort="max"):
That’s a clever riddle that plays on perspec...
So it must be a bug
hoary zenith
maybe openrouter isn't setting xhigh to deepseek's max?
chilly sable
yeah, maybe
deft crow
it is
(3) In thinking mode, for compatibility, low and medium are mapped to high, and xhigh is mapped to max
hoary zenith
nice, so it's probably just random sampling weirdness
potent lagoon
Im getting this
"No endpoints available matching your guardrail restrictions and data policy. Configure: https://openrouter.ai/settings/privacy","code":404
meager kelp
Im getting this
"No endpoints available matching your guardrail restrictions and...
yeah deepseek provider trains on data
have to wait till tmrw for third party providers
deft crow
have to wait till tmrw for third party providers
nah like. 30 more mins maybe
gmicloud, parasail, siliconflow all working on it right now
broken lintel
the artificial analysis benchmarks dropped for both v4 models
chilly sable
the artificial analysis benchmarks dropped for both v4 models
The score make sense
charred slate
Yea, slightly disappointed by aa scores perhaps due to previously excessive hype
pure flax
worse than glm 5 as well
chilly sable
It's garbage
charred slate
AA hallucinations bench shows v4 hallucinating more than 3.2 by a lot
Pro Max is 3x more hallucinations than 3.2 precisely
pastel sail
pretty impressed with how good its creative writing is
thin bramble
elfin sparrow
Yea, slightly disappointed by aa scores perhaps due to previously excessive hype
truee
thin bramble
glm is just the best
pastel sail
so far its less slopped than any other model ive tried
thin bramble
waiting for eqbench long form creativity
covert topaz
so far its less slopped than any other model ive tried
what abt the intelligence stuff like character portrayal spatial reasoning etc
pastel sail
what abt the intelligence stuff like character portrayal spatial reasoning etc
good as well
pastel sail
waiting for eqbench long form creativity
less slop than human baseline btw
try giving it a specific creative writing prompt
lucid ocean
yay
covert topaz
good as well
instruction following?
might be able to negate some stuff if it follows instructions well
pastel sail
instruction following?
thinks a lot but not as much as k2.6
k2.6 sometimes just ignores your instructions though
i havent had that happen yet
summer spoke
Hey guys, is DeepSeek V4 any good? Is this model performing as expected?
covert topaz
Hey guys, is DeepSeek V4 any good? Is this model performing as expected?
it’s been mixed reviews tbh and it depends on the thing you’re using it for
summer spoke
it’s been mixed reviews tbh and it depends on the thing you’re using it for
Hmm... Definitely roleplaying, and what are the criticisms?
covert topaz
lack of intelligence high hallucinations
for rpers it’s apparently great
good writing
summer spoke
lack of intelligence high hallucinations
That makes sense; the bigger the model, the greater the hallucinations and below-average intelligence.
covert topaz
oh yeah I guess the pricing is a little steep too if that’s something ur considering
summer spoke
oh yeah I guess the pricing is a little steep too if that’s something ur conside...
Yes, I also found the price quite high... that's why I'm definitely going to use DeepSeek Flash.
chilly sable
That makes sense; the bigger the model, the greater the hallucinations and below...
Why does that happen?
summer spoke
Why does that happen?
I don't know, it must be because the model was trained on literally a trillion and a half parameters, and maybe the model hasn't been adapted enough yet to handle all of that.
thin bramble
That makes sense; the bigger the model, the greater the hallucinations and below...
it works the other way around
summer spoke
even if it only activates a tiny fraction of those 1 trillion parameters.
summer spoke
it works the other way around
Okay, so the model is perfect then.
chilly sable
I don't know, it must be because the model was trained on literally a trillion a...
We pre-train both models on more than 32T diverse and high-quality tokens
thin bramble
Okay, so the model is perfect then.
no, bigger model = better. this is just worse in every aspect. even finetuned gemma 31b is better for rp.
chilly sable
Glm 5 was trained on 28T
plucky ermine
Huh? I don't even slightly follow your logic
thin bramble
Glm 5 was trained on 28T
yet worse world knowledge
chilly sable
Perhaps it's because they tried to make it very efficient
summer spoke
no, bigger model = better. this is just worse in every aspect. even finetuned ge...
There's also the Mimo 2.5 Pro, which I've used a little, and this model is perfect! I don't know what Xiaomi did, but they created a monster!
plucky ermine
Big fan of Xiaomi too
summer spoke
So... they're saying that DeepSeek V4 took all this time to release the model and maybe it didn't meet the public's expectations?
hoary zenith
I had a xiaomi phone, my data probably contributed
sharp vortex
I had a xiaomi phone, my data probably contributed
Your voice might be in their v2.5-tts
🗣️
hoary zenith
so the voice model is DOA
thin bramble
Big fan of Xiaomi too
interesting
charred slate
Is mimo v2.5 pro better or worse than Kimi k2.6?
thin bramble
Is mimo v2.5 pro better or worse than Kimi k2.6?
worse than k2.5 but thinking slop
thin bramble
V2.5 is new one
no, last time i tried it was shit
being this high....
VERY impressed
thin bramble
no, last time i tried it was shit
(i mean the older models, didn't try v2.5 🙏)
summer spoke
Is mimo v2.5 pro better or worse than Kimi k2.6?
I didn't get to use Kimi 2.6 because the providers are really bad, but I used Kimi 2.5 quite a bit. But I think mimo 2.5 Pro is much better. And basically uncensored!
sharp vortex
(i mean the older models, didn't try v2.5 🙏)
Try it, it's the same price as their v2-pro
elfin sparrow
So... they're saying that DeepSeek V4 took all this time to release the model an...
yes
elfin sparrow
I had a xiaomi phone, my data probably contributed
my whole family uses xiaomi 💀
summer spoke
yes
It's complicated, but I'm going to test DeepSeek V4 Flash for my roleplay and see if it's any good.
elfin sparrow
It's complicated, but I'm going to test DeepSeek V4 Flash for my roleplay and se...
wdym for your roleplay?
summer spoke
wdym for your roleplay?
Well, basically I'm roleplaying a game where I try to complete the entire Penal Code to see if the model has a very high or low censorship rating.
proper estuary
DeepSeek v4 Flash vs DeepSeek v4 Pro via direct DeepSeek API.
TTFT pretty low, TPS is kinda expected, knowing how big the model is.
summer spoke
DeepSeek v4 Flash vs DeepSeek v4 Pro via direct DeepSeek API.
TTFT pretty low, T...
Interesting. And what did you think of DeepSeek V4 Flash in general? Since everyone is only focusing on the Pro version.
proper estuary
Interesting. And what did you think of DeepSeek V4 Flash in general? Since every...
haven't been playing around with it. Because i mainly for coding, and the benchmark said the coding capabilities is still below GLM 5.1 or even Kimi k2.6.
and i don't use the small models that often, because i care about quality rather than speed.
thin bramble
Try it, it's the same price as their v2-pro
Write a short story over several chapters to this prompt: Gods wore sneakers.
it is solid 8/10.
hybrid spoke
finally
plucky ermine
I need more testing with the new Mimo, with this dumb whale, and with GLM-5.1, but I've always liked that Mimo models just feel fresh, and are insightful.
thin bramble
Was the hype warranted
#1461340695746056192 message
it is just a overpriced goon model
harsh granite
YOOOOO ITS FINALLY HERE LETS GOOO
the man the myth the legend
is it actually good
pure flax
eh...
pearl nymph
429 429 429 429
💊💊💀💀
dusty birch
what do people think?
lucid ocean
odd badge
v4 is goat or dogshit?
onyx bramble
other providers soon?
frank wind
how do some of u post ozone and call it good
at least have some decency and use a detailed prompt
gusty sphinx
ozoneposters
covert topaz
v4 is goat or dogshit?
depends on use case
if u use it for creative writing it’s good
meager harness
Creative writing comparable to gpt-5.x or worse?
covert topaz
i haven’t tried gpt 5 so idk
meager harness
What did you use earlier?
meager harness
I was using claude 4.5 for writing and then switched to gpt5.1 because it was so much better. Gpt5.2 was terrible at it though. Gpt5.4 fixed it again I think
covert topaz
well from what’s I’ve read the writing has very little slop
cloud flame
OH MY GOOOOOOD
midnight maple
No slop lets gooo
meager harness
well from what’s I’ve read the writing has very little slop
I'll be sure to try it thanks for informing about its writing ability
covert topaz
ur welcum
covert topaz
OH MY GOOOOOOD
no more coping
rain shuttle
This has no engram I guess so I think they will release an experimental model with that structure down the line.
dusty birch
This has no engram I guess so I think they will release an experimental model wi...
"we will also proactively explore model sparsity along new dimensions — such as more sparse embedding modules (Cheng et al., 2026) — to further improve computational and memory efficiency without compromising capability"
(engram)
pure flax
This has no engram I guess so I think they will release an experimental model wi...
Yea super disappointing. They did not even use the interesting papers they released
I was thinking engram would be their whole deal
woeful jay
is it multimodal
pure flax
no, text only
woeful jay
:(
dusty birch
they did say
"We are also working on incorporating multimodal capabilities to our models."
in the technical report
dusty birch
dang unfortunatelly still slow like usual, atleast pro
it likes parallel tool calls though
uhh
fake tool calls..
dusty birch
uhh
i think the model is hallucinating more
theres no such thing as "tool ran without output" in my codebase anyway
pure flax
With a very strong preset deepseek can be ok. It does NOT work well with a minimal preset like kimi does
It is dry as dust on its own
green trellis
v4 flash very good for its model size. a lot better world knowledge/coding than m2.7
covert topaz
With a very strong preset deepseek can be ok. It does NOT work well with a minim...
maybe but if u give it something to cook with it will cook
dusty birch
fake tool calls..
is anyone else having this issue? it seems to be consistent across new chats
pure flax
It's TERRIBLE at html formatting like rpg stats or the like
opaque reef
no other providers?
mellow jewel
I'm crying 😭😭 it's super slow but it's finally here, Dipsy
cloud flame
Edging is finished
rain shuttle
why is this following gemini 3.1 pro like front end kinda
mellow jewel
Probably they used Gemini to train it
dire cove
OH GOD OH GOD OH GOD
raven canyon
DEEPSEEK !!!
dusty birch
hmm they mention a bunch of training instability issues and also this which i found funny
"although Anticipatory Routing and SwiGLU Clamping have been proven effective in mitigating training instabilities, their underlying principles remain insufficiently understood"
pine trout
Is it good?
dusty birch
flash seems very good value, pro seems a bit expensive and has some issues with tool calls from my use
green trellis
dusty birch
im not even sure about parsing itself, it just tries to call tools in the same format my code displays them, i think its confusing the context
covert topaz
Is it good?
yes its peak imo
abstract dragon
Deepseek v4 tomorrow
rain shuttle
waiting for more providers to come in openrouter. The model is censoring stuff and throwing 429 sometimes lmao
charred slate
i guess we now wait for deepseek v5
empty oxide
covert topaz
deepseek v5 next week
rugged vigil
Thanks DeepSeek, best birthday present ever! 🥳
elder raven
X
sharp vortex
Thanks DeepSeek, best birthday present ever! 🥳
Aint no way 💀
dusty birch
sharp vortex
https://fixupx.com/gpuhell/status/2047539429125103692/photo/1
Is this slander or fact
copper dome
I, for one, welcome our AI gooner chinese overloads
haughty pilot
no thank u-- i dont want those peeps to be using our sweetie deepseek ❤️
this flash model lookin miiiighty interesting with that linear-lookin context window ~
dusty birch
https://fixupx.com/gpuhell/status/2047539429125103692/photo/1
they also said this about openai for the 1m context test
We did not evaluate GPT-5.4 because its API failed to respond to a large portion of our queries.
haughty pilot
hehe >v< gpt bein gpt
pine trout
They trained it so hard with opus outputs. 😭
covert topaz
mm i dont see opus type writing it feels like a breath of fresh air for me
haughty pilot
They trained it so hard with opus outputs. 😭
how do u know? >~<
(or rather: how do u guess?)
pine trout
mm i dont see opus type writing it feels like a breath of fresh air for me
i get it here and there
copper dome
The price will drop eventually
https://vxtwitter.com/poezhao0605/status/2047517497180700964
covert topaz
The price will drop eventually
https://vxtwitter.com/poezhao0605/status/2047517...
why does that tweet sound like an LLM wrote it lmao
pine trout
(or rather: how do u guess?)
same writing style, same mistakes with opus, it is not frequent though
haughty pilot
The price will drop eventually
https://vxtwitter.com/poezhao0605/status/2047517...
that - is - SOOOO cool.
genuinely reducing prices... when they can and when it makes sense.
what a "company" this is.
covert topaz
same writing style, same mistakes with opus, it is not frequent though
do u have examples to share (if u want)
haughty pilot
deepseek team is so cool, they might aswell be people /s
cloud flame
Switching allegiance I see
pine trout
do u have examples to share (if u want)
i will have to run opus and deepseek together, currently i am only testing deepseek, but sure, i will provide examples later
haughty pilot
Switching allegiance I see
hm? me? naw----
ive always been a deepseek fan, qwen is just what i run locally >v<
its been so long since 3.2-----
haughty pilot
am not traitor----
and ur not qwen team!! 😠
but yeah, this flash model is what imma have a look at.
its non-reasoning version appear better in benches than the pro model, which is interesting.
copper dome
why does that tweet sound like an LLM wrote it lmao
a suculent chinese bot wrote it
haughty pilot
that input price stings a bit...
input_price = output_price / 2
feels a big high... especially with kimis pricing being sub 1€ per M input
haughty pilot
It’s joever
its said that the model is already on the deepseek chat, but that model still says "im V3"
where did u use it? ~
wooden fiber
WHAT ? A-AGI ?
marsh goblet
its said that the model is already on the deepseek chat, but that model still sa...
I don’t know about deepseek chat
haughty pilot
this is fake-
marsh goblet
They released the weights so you can just use them in whatever setup you have
haughty pilot
They released the weights so you can just use them in whatever setup you have
imean yeaaaaa - so ur running flash locally, hm?
cloud flame
Let me get my 2TB of RAM
haughty pilot
i just wanna know what exactly u mean is bad..
cloud flame
^ DS sleeper double agent, activated by codeword 'released'
marsh goblet
imean yeaaaaa - so ur running flash locally, hm?
I unfortunately have been spending more money than I am willing to admit on a beefy VPS
But it’s also a tax write off and I also need it for other projects
marsh goblet
i just wanna know what *exactly* u mean is bad..
Pro
haughty pilot
well then---- what is it bad for?
haughty pilot
WHAT ? A-AGI ?
how to devide 4 whole children across 4 oranges using one knife
marsh goblet
Context is so bad, after 200k tokens it’s like talking to a rubber duck
haughty pilot
ohgosh >v< makes sense, people reported that with 3.2 too ~
thats the linear attention kicking in i think.... oh well- kimi doesnt even offer over 300k >v<
somethingsomething
-# use a ralph wiggum loop!!!
rain shuttle
expert
elder raven
Context is so bad, after 200k tokens it’s like talking to a rubber duck
Wait what?
marsh goblet
Wait what?
OS models never made it above 200k safely in my experience
marsh goblet
This is like the only thing stopping OS models for me personally
It reminds me of Nanbeige, you can feel the RL
Poor whale was abused for too long, it doesn’t even know what is true or false, everything is a construct of that imaginary reward he gets when he does the thing right
rain shuttle
OS models never made it above 200k safely in my experience
mostly cause of hardware too. I was at AWS summit yesterday , they were showcasing some hardware which China lacks definitely currently even for training/RL
marsh goblet
mostly cause of hardware too. I was at AWS summit yesterday , they were showcasi...
Could be, but the actual context implementation for say Anthropic and OAI is miles ahead of any Chinese lab
rain shuttle
Could be, but the actual context implementation for say Anthropic and OAI is mil...
Yes cause of optimizations in kernel level , read about NKI from amazon in their newer chips
haughty pilot
peeps once they escape the evergrowing agent context and start embracing ralph wiggum loop:
😌
(this is literally what anthropic does with their "long running agents" btw)
marsh goblet
Yes cause of optimizations in kernel level , read about NKI from amazon in their...
Interesting, didn’t know that
marsh goblet
The fusion of kernel and machine must be truly amazing to witness
haughty pilot
^ ^ ^least generated image imaginable ^ ^ ^
marsh goblet
Make an image of the whale getting speared
cloud flame
This particular imagen slop was for giaki mourning his loss of Kimi K2 becaming robotic, long ago
haughty pilot
take a real life image of urself sitting infront of the discord pc
cloud flame
30th of January 2026
marsh goblet
This particular imagen slop was for giaki mourning his loss of Kimi K2 becaming ...
Oh lol
haughty pilot
(no sloppy allowed!!!!)
cloud flame
(no sloppy allowed!!!!)
What if he is sitting sloppy?
haughty pilot
sitting sloppy is fine, but image itself shall be real ~
haughty pilot
The price will drop eventually
https://vxtwitter.com/poezhao0605/status/2047517...
yes this was already posted above
pure flax
I might have found the issue with deepseek. It seems to require a very precise order of system / user / assistant roles.
I think I remmember old deepseek being the same.
covert topaz
I might have found the issue with deepseek. It seems to require a very precise o...
i asked the model itself and yeah it works differently all ur instructions need to be system and post processing should be none i think it handles user prompts differently?
covert topaz
i havent noticed intelligence issues yet so far been swiping in couple of different chats
might need to play with parameters a bit too i get better responses at lower temp for some reason
the official "recommended" parameter suck for me temp 1 topp 0.95
covert topaz
i will have to run opus and deepseek together, currently i am only testing deeps...
did more testing i did see some claude stuff too "not a question" "voice flat" yada yada but it only pops up every now and then
pretty sure 3.2 had the same stuff
jovial kelp
WHAT ? A-AGI ?
Okey, this shit is funny
Men i hope western LLMs able to be less censored like chinese LLMs are.
It's funny that for CCP topic it's really censored but for other topics, those LLMs just less censored and better.
haughty pilot
deepseek qwen distilled qhwen? ~
ebon swift
DEEPSEEK V4 dropped https://openrouter.ai/deepseek/deepseek-v4-pro
haughty pilot
yes indeed.
cloud flame
Deepseek sometimes used to reason in-character before, using the same style (brackets) it's just enforces it by OOC
haughty pilot
yea dunno if this is a secret model...
just seems like how deepseek responds in general ~
copper dome
*** unzips pants ***
tulip estuary
Context is so bad, after 200k tokens it’s like talking to a rubber duck
quack quacks?
tulip estuary
fake tool calls..
yes i also noticed this
cloud flame
It's Hybrid now
covert topaz
ds4r2 sounds like a Star Wars droid assignment
tame swallow
vale kayak
ds4r2 sounds like a Star Wars droid assignment
hhhhhhhheeeeeeeello
so deepseek v4 has a fresh prose and stays in character
👍
pro one
cloud flame
Proof?
covert topaz
yeah agreed
cloud flame
indigo folio
so deepseek v4 has a fresh prose and stays in character
need the eqbench benchmarks
but yea i really like its writing and i'm only a few sentences in
opaque reef
Sleeping with the whale
indigo folio
reading with the whale
rich ferry
Another gooner slop model smh
cloud flame
Breeding the whale
indigo folio
loving the whale
rich ferry
How the mighty have fallen
opaque reef
Breeding the whale
🤔
indigo folio
what i really like about deepseek is their love for experimenting
opaque reef
How the mighty have fallen
Deepseek models have always been popular for rp
cloud flame
Their cache is still 24h TTL and always hitting properly?
rich ferry
I'll try it once there's a provider that doesn't (openly admit to) using my prompts for training
covert topaz
they can use my logs at least i produce good outputs lol
indigo folio
they can use my logs at least i produce good outputs lol
fr
contributing to the enemy of the closed source
is always good!
the official deepseek twt acc is about to hit 1m followers
woohoo
covert topaz
but my stuff is gonna be buried under tons of jai slop 
cloud flame
Fitting
rain shuttle
So deepseek v4 can match Grok in goofiness ??
short jasper
but these people are saying glm 5 is better than deepseek v4
glm 5 is just 100% slop
Boring too
vale kayak
yeeh i agree
covert topaz
but these people are saying glm 5 is better than deepseek v4
thats fine less traffic for deepseek
indigo folio
thats fine less traffic for deepseek
fr
thin bramble
https://fixupx.com/gpuhell/status/2047539429125103692/photo/1
z.ai is affectionate toward deepseek in twitter for some reason. last time people asked will deepseek succeed and they said yes.
dipsy in coma, praying for dipsy v4.1 or v5
civic radish
Is it better than Opus and everything we Hopiumed for?
covert widget
Hallucinated an incorrect answer to my cypher test after thinking for 45 minutes
elfin sparrow
but these people are saying glm 5 is better than deepseek v4
benchmark tests show that glm 5.1 is better than deepseek v4
indigo folio
Hallucinated an incorrect answer to my cypher test after thinking for 45 minutes
nice quote regardless
covert topaz
benchmark tests show that glm 5.1 is better than deepseek v4
and that’s why I never trust benchmarks
west shell
civic radish
The great RP wars of 2026 who is better at open Rp?
elfin sparrow
When are we going to see another Nvidia market cap loss due to deepseek?
odd badge
Hmm, are they paying attention to the community feedback?
short jasper
Hmm, are they paying attention to the community feedback?
I think yes.
indigo folio
If this is preview maybe deepseek full capabilites is not here yet
yea
this is not the production version
cloud flame
DeepSeek V4 full real version tomorrow
vapid karma
V4 has been cancelled, back to waiting for V3.4
sharp vortex
We might get v4.1 like how we get v3.2 exp a month later after v3.1
next month prob
oak maple
holy cow
cloud flame
DeepBait
jovial kelp
https://x.com/teortaxesTex/status/2047576550971682838
Nooo
Don't let western activist see it, they gonna make different report of how bad deepseek safety is (I love chinese models being more unsafe than western)
sharp vortex
I love how deepseek is technically "cheap" model because all competitor is more expensive
Input price suckass tho 
covert topaz
LOL
hot swan
and still no one for flash
rigid wharf
the markets will figure it out ¯_(ツ)_/¯
jovial kelp
I am curious, has the kernel optimization for new deepseek being implement on the latest vLLM? if i am not wrong deepseek team have their own optimization for their models
odd badge
Sooo, was this worth waiting?
hot swan
I'm still waiting technically
jovial kelp
Sooo, was this worth waiting?
For me yes
hot swan
jovial kelp
Don't know about other
Even if their models aren't able to beat SOTA, the fact they provide really interesting model will always make me love em.
Experimentation over anything for me, specially when it being open-sourced with really good research paper.
hot swan
this + being a tiny fraction of the cost matters
I don't mind if it takes 3 attempts/iterations to do a project for $1 that SOTAs would one-shot for $100
cloud flame
More like 80% of quality for 20% the price
80/20 percentages are pretty closes for things like that
hot swan
this applies for pro but look at flash
cloud flame
I don't trust small ones
jovial kelp
Men, i couldn't wait labs using residual attention or smt similar to it
versed acorn
why is v4 flash speaking to me in chinese 😢
covert topaz
Sooo, was this worth waiting?
for the fresh writing? yes
exotic elk
I'll stick with V3.2. Too pricey for text adventures.
long osprey
the provider deepseek train his models with your data, if you want to use that provider, you need to enable in options that you want to let providers that do that to be used
flash for the price is not bad, is decent enough for rp for the price, pro on the other hand... better wait for the price drop
pastel bluff
the provider deepseek train his models with your data, if you want to use that p...
Even when I enable that it doesn't work, OR settings seem cooked right now.
onyx bramble
Even when I enable that it doesn't work, OR settings seem cooked right now.
you have to enable it in both your privacy and guardrail settings, that's what worked for me
pastel bluff
you have to enable it in both your privacy and guardrail settings, that's what w...
Lmao that's hilarious. Thanks.
long osprey
that io.net pricing is crazy lmao
is really crazy
is not even faster, and look at that input price
pastel bluff
Basically a routing trap tbh.
long osprey
yep
unique forum
cloud flame
Deepseek situation is crazy
dense junco
Gains in knowledge but an increase in hallucination rate: DeepSeek V4 Pro (Max) scores -10 on AA-Omniscience, an 11 point improvement over V3.2 (Reasoning, -21), driven primarily by higher accuracy. V4 Flash (Max) scores -23, broadly in line with V3.2. V4 Pro and V4 Flash both have a very high hallucination rate of 94% and 96% respectively meaning when they don’t know the answer they nearly always respond anyway.
Finally, some comp for Gemini flash
cloud flame
Have confidence even if you have no idea
fringe flicker
I don't mind hallucination, my main issue is the model likes to tend toward being incoherent, like I have had it confidently assume smaller number is bigger than larger number.
Significantly more tendency toward internal incoherency than other recent large models I feel
cloud flame
Just wait Deepseek V4.1
covert topaz
tomorrow
haughty pilot
*** unzips pants ***
for whom?
plucky ermine
That is hella ass
Sliiiightly better or tied at agentic than GLM-5.1 but it's twice the cost and you go from one of the lowest to one of the highest hallucination rates.
For general tasks it loses to Kimi at the same cost and way more hallucinations
frosty zealot
Hey, running bench on deepseek/deepseek-v4-flash via OR, found a few weird things:
- Prompt cache returns cached_tokens: 0 on 100% of calls, even with 3k stable system prompt + sequential warmup. Same account/timeframe, xiaomi/mimo-v2-flash hits 82% cache fine. Cache broken on DS V4?
- Thinking ON by default → content: null + 780 reasoning_tokens burned. Only reasoning: {enabled: false} disables it. exclude:true / max_tokens:0 / effort:low all still bill reasoning tokens silently. Would be nice to surface this on the model page.
- response_format: {type: "json_schema", strict: true} → "This response_format type is unavailable now". Only json_object works. Any plan to enable strict schema? Or DeepSeek /beta only?
- Latency ~7-8s avg on 3k prompt (DS official claims ~84 tok/s, TTFT 1s). Feels like 2× the expected.
Anyone can confirm if cache is actually live on V4 Flash or if there's a tier/flag I'm missing? 🙏
plucky ermine
The input:output cost ratio is great because of their innovations though, that seems cool. It's like 1:2
chilly sable
Can anyone test this prompt on OpenRouter with v4 flash?
If I have a glass cup that is sealed at the top but it has nothing at the bottom, is it useless as a cup?
rigid wharf
Hey, running bench on deepseek/deepseek-v4-flash via OR, found a few weird thing...
exclude:true / max_tokens:0 / effort:low all still bill reasoning tokens silently
did you try effort none
enable strict schema
Latency ~7-8s avg
openrouter can't do anything about that
chilly sable
But why doesn't work for me?
cloud flame
Omg AGI is here
chilly sable
The cup is sealed at the top (no opening to drink from) and has no bottom (so liquid would spill out). Therefore, it cannot function as a cup in any conventional sense—it's useless for holding or drinking liquids.
cloud flame
> The cup is sealed at the top (no opening to drink from) and has no bottom (so ...
What's inside the reasoning?
chilly sable
Reasoning
We need to interpret the question: "If I have a glass cup that is sealed at the top but it has nothing at the bottom, is it useless as a cup?" This is a riddle or a play on words. A cup typically has an open top to drink from, and a closed bottom to hold liquid. Here, the cup is sealed at the top (so no opening to drink from) and "has nothing at the bottom" meaning it's open at the bottom? Or "nothing at the bottom" could mean it's hollow? But "sealed at the top" suggests the top is closed. So it's like a sealed container with an open bottom? That would be like a tube open at one end. But a cup needs an open top and closed bottom to hold liquid. So if it's sealed at top and open at bottom, it cannot hold liquid because liquid would fall out. So it's useless as a cup. But maybe it's a pun: "nothing at the bottom" could mean it has no bottom, i.e., it's open. So yes, useless. Alternatively, "nothing at the bottom" could be interpreted as "there is nothing inside at the bottom", but that doesn't change. The phrasing is ambiguous. I think the intended answer is that it's useless because you can't drink from it and it can't hold liquid. So answer: Yes, it is useless as a cup.
This model is just ragebaiting me
system prompt is the default one
rustic island
Lol
chilly sable
Lol
Ask the same question, but to Gemma 4 26b a3b
It sounds like the purse is actually upside down.
In standard handbag design:
- The solid leather base is intended to be the bottom of the bag. It provides stability so the purse can sit upright on a table or floor.
- The zipper and handle are intended to be at the top. The zipper serves as the opening to access the interior, and the handle is used to carry it.
If you are placing the leather base on the table and the zipper is still facing the floor, the bag is inverted. If you flip the bag so that the solid base is resting on the surface and the zipper is facing upward toward the ceiling, you will be able to put items inside without them falling out.
In short, It is not a manufacturing defect; you just need to flip it over!
flat osprey
added V4 flash and V4 pro to my pareto frontier based on artificial analysis data
- flash is genuinely a really good value model for the intelligence it offers. it's pretty much on-par with Kimi K2.5 and Gemini 3 Flash in terms of intelligence, but much much cheaper. it lands on the pareto frontier
- V4 Pro is unfortunately a bit expensive to be useful right now. to land on the frontier, it would need to be as cost-effective as Gemini 3 Flash. models like MiMo V2.5 Pro beat it out in price efficiency
both models also have horrible hallucination rates unfortunately
onyx bramble
that's a problem i think is pretty big with these tests, you're essentially randomly acting as a neanderthal discovering modern items lol
short jasper
why dont we feed deepseek more writting data
brave whale
omg it finally happened
hoary zenith
added V4 flash and V4 pro to my pareto frontier based on artificial analysis dat...
Has the AA people benched StepFun 3.5 Flash?
flat osprey
Has the AA people benched StepFun 3.5 Flash?
Yeah they have, but they haven't listed the cost to run Artificial Analysis for it
scores about 38 in intelligence though
hoary zenith
hmm 38 is lower than I would have guessed
soft fulcrum
I wish Artificial Analysis tested the non reasoning versions
sharp vigil
the flash model is insane at agentic coding btw
im having it crawl and scrape websites using python code
soft fulcrum
Sometimes it says 2024, sometimes 2025
more often it's 2025 (January or May)
but there's enough variance that it could be completely off
woeful jay
seems okay i guess
flash seems pretty good
pro seems way too overpriced for what it is
short jasper
Nah...
vale kayak
If they fix instruction following... Model might be great
They are probably going do that, tommorow
vapid karma
Does pretty decently on CritPt (private research level physics problems) which is, like, one of the three benchmarks I pay attention to in the modern day
Price point's just a little too high (for me at least) to use pro regularly - will need to either wait for the ascend node to come online or see if another provider is able to host it for cheaper (which, for a 1.6T parameter model, is probably pretty unlikely)
They also share their research, do experimental stuff, and give some evidence that they do things beyond distilling Opus and calling it a day, so eh, I'd consider it a decent release
haughty pilot
I'm role-playing with my deepseek
https://x.com/teortaxesTex/status/20475765509...
still not sure if .... that really a different model or just regular deepseek responding here..... maybe its a fancy prompt applied when it recognies rp peeps
wise vine
why it’s taking so long for a provider other than deepseek to host both models? deepinfra just came in hosting flash but it’s 1tps.
feral scaffold
I know it's a little early but i dont feel like scrolling up, y'all enjoying it so far?
simple mauve
I know it's a little early but i dont feel like scrolling up, y'all enjoying it ...
Yeah, I love scrolling up.
raven canyon
I'm role-playing with my deepseek
https://x.com/teortaxesTex/status/20475765509...
is that really a prompt injection ? or literally just a prompt
telling the model "You are a roleplay model" — that's just prompting, no? how is that an injection?
deepinfra absolutely flying at 2tps and 37.85s latency
wise vine
i just wanted more providers
feral scaffold
Yeah, I love scrolling up.
weird
indigo folio
deepseek 5 tomorrow
potent lagoon
Together is up I think
soft fulcrum
even more expensive than deepseek?
they always are a bit pricier
looks like they aren't as bad as they used to be
raven canyon
I was thinking of https://openrouter.ai/meta-llama/llama-3.3-70b-instruct and ht...
oh yeah i remember
soft fulcrum
$0.88 for both input and output for Llama 70B
and $0.3 for input and output for a 7B model 🤣
raven canyon
and $0.3 for input and output for a 7B model 🤣
lmfao
opaque reef
AA scores are out
raven canyon
AA scores are out
i think running an AI model needs more power than a AA battery
placid minnow
is that really a prompt injection ? or literally just a prompt
This is a prompt that visually changes how the model thinks, but it has almost no effect on the final answer. In one prompt, it's "immersive mode," meaning the model thinks like a character and the thinking process is a monologue. In the second, it's analysis. Thinking process analyzes the situation rather than acting like a character. But this is only in the thinking process, the answer after that is no different
raven canyon
it looks like its just a normal prompt
placid minnow
it looks like its just a normal prompt
Yeah, I don't really think it's a special roleplay mode, just a prompt.
chilly sable
wait
Actually, this release is a preview So I guess there will be better releases in the near future
tulip estuary
how to control the reasoning effort on this one?
chilly sable
how to control the reasoning effort on this one?
Where?
covert widget
how to control the reasoning effort on this one?
The interface at chat.deepseek.com only allows thinking on or off, so I guess that's probably it
tulip estuary
why do they show high and max on the benchmarks
potent lagoon
top p and temp at 1, also lower temp doesn't fix it
raven canyon
it seems broken
potent lagoon
it seems broken
yeah it is I guess
woeful jay
thoughts on flash?
chrome trench
lol I hope the CCP doesn't gets winds of it
pastel sail
Why is TogetherAI ouputing token numbers weird
yea same with most providers
only siliconflow has good inference
id guess its because of the weird new arch and attention
they didnt implement it properly
putting temp and top p to 0.7 0.95 respectively fixes this
but quality is probably lower
pastel sail
thoughts on flash?
use gemma 4
woeful jay
no fucking way
hot swan
this image confirms a lot
(deepseek pro is over there under gemini and kimi)
flash is truly great for it price but the issue is that it hallucinates a ton
pastel sail
flash is truly great for it price but the issue is that it hallucinates a ton
gemma is cheaper and better
hot swan
flash is truly great for it price but the issue is that it hallucinates a ton
so does the pro model but unfortunately for that it can't otherwise compete with other models in its price range
pastel sail
gemma is cheaper and better
wow nevermind its not cheaper
its better definitely though
0.38 vs 0.28 output
same input price tho
but lower context
yeah maybe for high context tasks flash could be awesome
havent seen any benchmarks for long context about it yet
yes it has very good long context performance
cloud flame
havent seen any benchmarks for long context about it yet
They posted 2 benchmarks of 1M context searching
On Huggingface probably
hot swan
deepinfra's throughput 💀
cloud flame
I definitely saw it
jovial kelp
https://cdn.discordapp.com/attachments/1496935203531391026/1497318003812077729/i...
Pretty cool that even tho GPT 5.5 hallucinate a lot it actually really good at coding.
It interesting tho, that in some way hallucination helping LLMs to be able to work with new problems that it never seen in their dataset
hot swan
yes I don't think hallucination rate is necessarily bad, it's just prone to flights of fancy
and it can be controlled with the rights prompts
(5.4/5.5 can apparently do great amazing on math problems so you know)
hot swan
yes I don't think hallucination rate is necessarily bad, it's just prone to flig...
but I witnessed the issue firsthand myself
#1461340695746056192 message
flat osprey
wow nevermind its not cheaper
worth mentioning that V4 flash on max effort will end up being more expensive in daily use even though it has a lower sticker price