general

[keen beacon]

Just waiting for the news of R2 getting a release date at least

Now add the bonus points for reasoning and R2 will have a great chance to blow OpenAI out of the water

Or we can take a look at LiveBench, where GPT-5-Chat is right around GLM 4.5 and Kimi K2

78 points for high - 60 for chat

18 points difference

Not bad

It is for instance the same difference as between V3 and R1-0528

Now add this same difference to V4 - that will definitely be no worse than K2 or Qwen base - aaaand R2 can appear at least as good as Gemini

Just kidding, on LMArena it is just as much as between old V3 and R1

[calm sequoia]

What is the default ranking of the text benchmark? It's not the "Score" as I can see

[lofty elm]

i use gpt5 high with 2000000 context it's really make good responses for me

[barren aurora]

lmarena ai
Can the image or video be used commercially?

[vale jungle]

Hello, I'm new.

[jovial sapphire]

i made this with nano banana

original picture was me on my window lol

[frosty lark]

you have it right in the picture, it is the rank (that takes in account the CI). You can sort by score if needed.

[languid crescent]

uhhh

[rare python]

September 2024 didn't have GPT 5 official release to the public

[languid crescent]

ohhh

also, what's nano banana?

[ocean vortex]

I agree it's confusing. But they are taking the deviation into account there. +/-8 and +/-5

So max for o3 is 1449 + 5 = 1454
max for Opus is 1447 + 8 = 1455

1455 > 1454

[mortal coyote]

@echo aurora i keep getting this error , even after changing internet and browser

[worn bison]

Whats currently the best coding ai?
In terms of actually coding (Speed doesn't matter)

[lofty elm]

how do i select gen models for image gen

[pastel vale]

❌ Generation failed. Failed to create evaluation session. why i m geeting this msg

[keen beacon]

You can do that on lmarena.ai but not here, I think.

[vale jungle]

Hello i have a question?

[jovial sapphire]

image model

[jovial sapphire]

in my opinion, gpt 5 high

[echo aurora]

You can select specific models when using Side-by-side & Direct modes on the site.

[echo aurora]

Whats up?

[vale jungle]

Can I generate videos with Veo 3? This video generator

[echo aurora]

Yes, but Video Arena currently looks down.

[vale jungle]

because

[obsidian cargo]

Getting a something is wrong error when attempting image edit battle

Regular image generation too

[echo aurora]

Hey everyone - it looks like Video Arena isn't currently working. I'm setting those channels to read-only until we fix the problem.

[echo aurora]

Hmm okay you're getting on regular site too?

There might be an outage...

[keen beacon]

Same

I had to make a new chat to make it work again

[vale jungle]

What happened to the video generator??

[obsidian cargo]

Text battle works

[keen beacon]

There are many errors happening, a possible outage

[echo aurora]

#video-arena-1 message

[echo aurora]

I wasn't able to repro

[echo aurora]

Now I'm getting errors.

Okay going to flag.

[keen beacon]

People are too eager to make images and videos it seems :D

[vale jungle]

Look at this image that I generated with the model, I converted an image to a realistic one.

[jovial sapphire]

It's good, the more users there are on LMarena, the more accurate the leaderboards are

[vale jungle]

Is a test

[jovial sapphire]

This is absolutely horrific

[keen beacon]

It is. People still say LM Arena is a benchmark which it isn't. It tests on real life problems and usage.

[echo aurora]

Lol true

[keen beacon]

I dont understand the hate people give on it on Reddit for example

[obsidian cargo]

Back up now

[echo aurora]

Ugh it's working again for me. (I know that's a good thing but still)

[obsidian cargo]

Or not

[echo aurora]

This happened yesterday too

[obsidian cargo]

False alarm ehehe

[jovial sapphire]

Nah not working

[echo aurora]

And erros again lol

[keen beacon]

On SWE bench and rebench GPT-5 was ahead of everything IIRC.
If you're looking for open weights model, there is Qwen3-Coder.

In my experience, Deepseek writes more efficient, smaller and elegant code - but it is slower, and it sucks at SWE bench

https://tenor.com/view/sparks-electrical-electricity-equipment-radio-amateur-fire-gif-24123123

LM arena

[jovial sapphire]

Hahahahaa

[vale jungle]

is collapsed

[obsidian cargo]

Ope it hasn't errored out yet this time let's see

[indigo hazel]

me too, thank you for flagging

[jovial sapphire]

just worked

for me

[obsidian cargo]

Gpt-image-1

[keen beacon]

what the... The heck is that

[jovial sapphire]

cooool

[obsidian cargo]

Aaand now I'm getting errors again

[keen beacon]

Me too

[obsidian cargo]

Doppelganger marina from Fear and Hunger 2

[jovial sapphire]

here are some cats to ease this chat

meow meow

[keen beacon]

Doooog

not mine

lol

I wish I had a dog

[jovial sapphire]

i love both

but cats are cuter

people that don'(t like

cats

are red flags

[obsidian cargo]

Cuter images of her I generated in Sora a while back

[keen beacon]

They are okay, I guess

[jovial sapphire]

meow meow

[keen beacon]

They really need to fix the yellow tint

in the next version

[obsidian cargo]

I agree fully yea

[jovial sapphire]

Nano banana:

Best result so far but there are some lady bugs

[keen beacon]

aghhhh

spiders

[jovial sapphire]

oops

[echo aurora]

Okay it should be fixed now.

[jovial sapphire]

yes it works!

had 3 runs in a row

[vale jungle]

3d

[indigo hazel]

thank you very much as always

[vale jungle]

Ralistic

[obsidian cargo]

Had two outputs where something went wrong with one of the responses even after retrying

[slow grotto]

is gpt-5-high just the same as gpt-5? i cant find any API pricing

[obsidian cargo]

Three now

[jovial sapphire]

Zoom out this picture like it was taken 30 meters away

Nano-banana

Amazing

You can see the yellow car

[vale jungle]

Yeah

[echo aurora]

#nano-banana

[fleet lintel]

wow.. image models are getting crazy.

[jovial sapphire]

Yeah

Still bad on some very specific tasks

but this is the worse it will ever be

editing won't be a thing in a few years, mark my words

[lofty elm]

theres no nanobanana

[jovial sapphire]

where

what

?

[jovial sapphire]

it's only in battle mode

it's an anonymous model

[lofty elm]

okay

thanks

[jovial sapphire]

https://lmarena.ai/?chat-modality=image

np!

[obsidian cargo]

@echo aurora only getting these now

[jovial sapphire]

refresh

cuz mine works

generating images is like a game for real for real

like sometimes i spend 2hours just doing that lmao

[obsidian cargo]

I had refreshed yeah

[lofty elm]

Do image outputs are being logged?

[jovial sapphire]

obviously yes

[echo aurora]

Hmm okay lets give it a few and try again.

[jovial sapphire]

they are on a s3 amazon bucket

they expire after some time but they can be retrieved

so yes, images are logged

[lofty elm]

oh i see, i almost put someones face

[obsidian cargo]

If I hit retry enough times it eventually goes through it seems

[jovial sapphire]

i put mine idc

if you post on instagram

it's the same

[agile bloom]

does this server have a list of all ai models, written in one place? I wanted to list down all the ai models available on LMArena

[obsidian cargo]

The leaderboards do but they don't include stealth models like nano banana

[fleet lintel]

I am getting imagen-4.0-ultra-generate-preview-06-06 much more than nano-banana

And in all my test, imagen-4.0-ultra is far better than competitors

[obsidian cargo]

Hitting retry on dozens of errored outputs only to get something that looks like absolute ass 🥲

[jovial sapphire]

Nano is better

[brittle tiger]

You'll get nano nonstop if you upload an image or multiple to have it edit

[obsidian cargo]

Weirdly enough I haven't gotten nano once with my "turn doppelganger marina realistic" image edit request.

[leaden palm]

"gpt-5 fixed hallucinations"

[obsidian cargo]

Maybe nano doesn't like the prompt but keeps getting selected to do it and that's why I'm having such a high error rate???

[lofty elm]

bruh

[urban arrow]

Impossible to use

[vale jungle]

You know why I like this model because it has good image quality to improve the quality of the photos

[indigo hazel]

me too, same error

[urban arrow]

lmarena died 😭

[obsidian cargo]

@echo aurora problems still occurring it seems

[indigo hazel]

it's true only for the most powerful gpt 5 models, such as the thinking modality high

[leaden palm]

true, high has like 200 juice it's crazy

[jovial sapphire]

Yeah, keep getting errors now

[ancient reef]

A lot of errors...

[echo aurora]

Okay thank you, flagging to the team.

[minor bloom]

So, is o3 search better than gpt 5 search?

In arena

[echo aurora]

Hmm I wasn't able to repro.

Image in Battle again?

[obsidian cargo]

Yeah

[obsidian cargo]

Try with my prompt?

Turn doppelganger marina photorealistic, she has no face just a giant mouth.

[echo aurora]

Do I have to

[obsidian cargo]

Ehehe sorry 😛

[placid shuttle]

Yesterday discovered this new AI and Im so happy to see the BIG step is being made right now in AI 🤯😳

[exotic nebula]

What horror am I seeing 😭

[south hatch]

Your AI model has suddenly gone viral in Taiwan, being widely used to generate PVC character model images.
It’s unclear whether the sudden surge in user numbers is the cause, but the service feels like it’s experiencing a DDoS issue.

[obsidian cargo]

A cutie patootie

[vale jungle]

Although this is current, I asked him to improve the quality.

The first image is better Quality

[exotic nebula]

Pretty good, which model?

[pastel badge]

in korea too

[vale jungle]

Nano banana

[south hatch]

so popular now

[exotic nebula]

Expected. Did you try milking it more?

[vale jungle]

Yeah

[boreal timber]

How do I use nanobanana??

[vale jungle]

Yeah

[boreal timber]

Is it here or on the lmarena website

[vale jungle]

I'm on my cell phone

[boreal timber]

Can anybody assist me with where to use the nanobanana feature?

[keen beacon]

On the website, in battle mode. It will appear in random intervals in the battles

you cannot choose the model since it's a pre-release model.

[vale jungle]

I found this thanks to a YouTube video

[boreal timber]

Oh thanks! Been looking for it all through

[white hatch]

Can gpt-5 think in chinese?

[indigo hazel]

i dont know anything about nanobanana, can you tell me something to understand?

[lofty elm]

Is the purpose of generating images are to enhance censorship or no?

im curious about the scope of this

[vale jungle]

I put the text

the results are good?

I improved the quality and put the text

[jovial sapphire]

which is original

[vale jungle]

Original

[hollow imp]

Flux is much better for this stuff

[vale jungle]

Yeah

[undone pier]

Does anyone know what happened to flux kontex max?

[mighty ocean]

The search leaderboard isn't updated for gpt 5 search. Would you guys consider the new gpt 5 search better than the original o3 search

[calm sequoia]

So they are taking the right side of mean + SD? Why did they choose the right side and not the left side in this case?

[jovial sapphire]

guys wthhh

check #nano-banana

[solid brook]

@echo aurora it will be really good to in someway increase input limit. if you guys add support to other file types to uploud like documents please don't put any limit on the file token size

[jovial sapphire]

lmaooo

they cant put no limit

the companies that give the model to lm arena

pay

it's not free for them

running a model isnt free lol

[solid brook]

dude

the input token is much cheaper

to process

than output token

[jovial sapphire]

yeah but "no limit"

xd

funny

[solid brook]

no limit no but like the current limit is quite sht

[jovial sapphire]

got u

[brittle tiger]

Not in my experience. The router seems to prefer non-thinking for what it deems search queries for quicker responses. So not really close to o3 at all

[solid brook]

maybe 20th august?

[jovial sapphire]

nah

i work at google

21-25

source: trust me

[solid brook]

see like when i use gpt 5 it can generate 5000 lines of code and it is way more expensive than proccesing 5000 lines code of input but on they have the input limit around 2400 lines

[dire brook]

Hi

[solid brook]

no

but we are getting there

[urban arrow]

Getting error again

[obsidian cargo]

same here, on text battle

[solid brook]

refresh

[urban arrow]

Already did.

[jovial sapphire]

u just gotta retry

and retry

until it works

[solid brook]

no

the only model that i hit limits is opus

first opus i think then gpt 5 high

[jovial sapphire]

No

Opus isn't better than GPT5 at coding lol

I used both for months

And weeks of the last Opus 4.1

GPT5 is better

I used it for Py, Golang

[fleet lintel]

big doubt! Google never launches new models without LMArena testing... and i see no google model yet. 21-25 is not happening (90% sure)

[solid brook]

yeah i guess

[jovial sapphire]

And some agentic stuff related to cyber security

That's what you think

Use GPT5 with API

Not chat

It's different

SWE is the worst benchmark, it's corrupted

You can't test a model relying on battlemode tbh

[solid brook]

btw in lmarena for max performance with gpt 5 high say think very hard in the prompt

[jovial sapphire]

Try it on cursor

Install cursor

You have free credits

250

Try the model with your best project

[fleet lintel]

but this is 2.5 pro. I am talking about Gemini3.0 launch

nah

they wont call it 2.5 for sure... they will have some random name like kingfall

give me a difificult prompt to test 🙂

[languid crescent]

kept getting these errors 🙁

[fossil fable]

really excited for distilled r2:a3b

[solid brook]

you are using gpt 5 chat>>>

????

[urban arrow]

Yeah, it's impossible to use LMarena today.
So much error messages

[solid brook]

guys remember

in lmarena models still have context window limit

no bro not that limited

[urban arrow]

That's not the problem. I've been doing the same thing all week, and only today I'm getting this error message, just like other people.

[solid brook]

the original model context limit

gpt 5 on api has 400k context limit

so it is 400k on lmarena

[hollow imp]

Add pdf support

🙏🙏🙏

@hollow ivy please

[solid brook]

like how much? line of code

i'm not a expert

idk how much lines that is

\you cannot input over around 2400 lines of code

oh yeah

just explain it to the model

[languid crescent]

still experiencing errors :/ have to refresh it numerous times to make it work

[midnight salmon]

I keep getting an error message

[languid crescent]

same :(((

[solid brook]

idk why all are getting errors i'm fine

oh no

i get it too

[languid crescent]

jinxed lol

[urban arrow]

Yeah, it's the server

[solid brook]

these are not normal

[midnight salmon]

Trying out Battle mode

[hot river]

for those of you that use openrouter, saturn is giving away 500$ of credits to testers helping with identifying bugs. free money fr

i didnt get any issue thismorning?

what was your prompt

weird, could have a limit on ctx

pretty expensive model

[languid crescent]

tried making my own prompt to be a study mode tutor

[thin creek]

Yeah, I’ve run into that too. When that happens, the only thing you can really do is grab the last few messages and start a new chat.

[languid crescent]

I usualy refresh my web about 10 times and it works normal again lol

if it happens again, i'll keep refreshing it

[solid brook]

sonnet no lol

opus maybe

[odd urchin]

video arena keeps failing to generate

it always says "failed to create evaluation session"

[urban arrow]

LMArena servers are unstable today

[solid brook]

hmmm actually idk i don't have any exprience on that

maybe claude is better in that area

[umbral root]

Hi, I don't see models like Seed-1.5-VL in the Vision Arena section. Is it because they haven't been added to the Arena yet?

[odd urchin]

oh ok thanks for the info

[vital beacon]

Ah okay so it's not just me?

[languid crescent]

yepp

[hollow imp]

My day is ruined hearing that

Opus is absolut shyt

[potent glacier]

Is it because of all the Taiwanese people using it to try nano-banana?

Someone said that earlier

[fossil fable]

it's horrendous to work with

i do NOT like working with grok 4

[fossil fable]

the lack of a grok 4 mini indicates that grok 4's "sota" was just achieved by scaling

because they basically pulled a deepseek with 3 and stole o1 chains then claimed sota which got them a ton of illegitimate funding

it's not workable and it's benchmaxxed because they ran out of models to steal from (like deepseek, but we'll have to see how good r2 will be)

[hollow imp]

Try SUPAGROK

[fossil fable]

well

not literally but

it stunk of qwq and r1

when you scanned its semantics it showed up directly next to them and o1

which is basically proof

the best thing about qwq and r1 however is that they opened up powerful stuff

grok hasn't been open since the first prototype

grok 4 is the result of a company realising they ran out of stuff to yoink basically

it's not THAT bad but like

unless you're running pure stem i wouldn't touch it

you said "nothing good is "open""

[fossil fable]

grok 3 blew me away in chemistry but nothing much else

chemistry is stem

i was really hoping gpt-oss would shine as a chatter but it turned out to be the complete opposite

like phi-4, it's just benchmaxxed on synthetic data

[keen beacon]

grok 4's math answers even when i ask it to explain r like 1 word final answer

and takes a lifetime

to answer

[fossil fable]

like, kimi has actual soul

it's nice to work with which is what you want to focus on instead of pure intelligence

even if it happens to be SOTA clever (well, it had to be), GPT-5 shines in the same way

[fossil fable]

i feel like intelligence as a benchmark plateaued at o3

and even then the best part about o3 is that it was really workable and readable

if grok 4 tried teaching me something it'd be like taking 20mg of ambien

[keen beacon]

fr

[fossil fable]

in my opinion the o1 series sucked because it never focused on the latter

o3 did

[keen beacon]

gemini 2.5 pro is better in physics STEM AFAIK

but grok 4 did better on a private bench

phybench.cn/en/leaderboard

[fossil fable]

i think they're roughly equal at stem, no?

[keen beacon]

yeah

gemini 2.5 pro is a tad better

tho

bench

[keen beacon]

i remembered incorrectly - 2.5 pro is better

[fossil fable]

a little surprising for physics

dude

grok 4 is the biggest benchmaxxing block of tra—

[keen beacon]

hmm

[languid plover]

When they are gonna add midjourney

[keen beacon]

how is it benchmax ??? its a private bench bro

bro no LOOK AT THE BENCH MARKS

[hollow imp]

@keen beacon 🥲

Sawaal 3 baar attempt nahi kr paarha hu

[fossil fable]

WHAT DID I LITERALLY JUST SAY

i wish there was a way to convey to laymen when a model is good to work with

but NOPE

numbers numbers numbers

f∆ck the numbers

"oH gPt-5 dOeSn'T sCoRe wElL oN [n]" and then i try it on the subject and it's brilliant

[slate pilot]

how to?

[indigo magnet]

Hi.

[potent glacier]

Absolutely wrong

There are a ton of open source image gen models that are amazing

[fossil fable]

why do i have to say gpt-5 with reasoning every time

[potent glacier]

And only getting better and better with time

[fossil fable]

open weights is not open source

i think that's what he meant

i thought that too

[scenic salmon]

It’s a joke because it would’ve been called o4 were it not for the rebrand on model naming

[hollow imp]

Has anyone tried stable diffusion?

[swift vapor]

so.. umm is it fixed yet?

[echo aurora]

Not yet, our monday just started

[jovial sapphire]

patience is key

[swift vapor]

lol i didn't knew that..... i'll be waiting " patiently"

[undone crane]

prompt for video-arena-1 and 2 and 3 not works !!

[echo aurora]

How so? I'm seeing successful generations.

[fossil fable]

most of it

however not all of it

[reef pawn]

Gemini 3 when

[obsidian cargo]

noticed a new model on the image arena: lucid-origin?

[exotic nebula]

Yeah, just now noticed it. Testing it out

[ocean vortex]

gpt5? it kinda is

well it's o5 if we try to fix their naming mess with o4-mini lol

They potentially did that on purpose to arrive at 5

and then merge

with gpt

[fleet lintel]

why is the difference decreasing every day?

[ebon eagle]

wow

[ocean vortex]

like gpt4 to gpt4o was perhaps even bigger jump than o1 to o3

[fleet lintel]

it went from 21 to 3.. it's crazy

[ocean vortex]

kinda crazy

[ocean vortex]

wtf

🗿

[white hatch]

daaaaaamn

[urban arrow]

Do the LMarena devs/staff know who owns the Nano-Banana model?

[leaden meteor]

When did we get this update? It shows Aug 16 but I did not see it yesterday...

[ocean vortex]

I think people may have gotten familar with gpt5 style and that they shouldn't like it, and so now they are voting against it more often. LOL

[patent aspen]

IIRC the 21 point difference dropped to 6 after they merged the lmarena GPT-5 with the public API model, which was supposed to be the same model

[brittle tiger]

The model will tell you it's google

[patent aspen]

#general message

Or rather the initial 15 point drop happened after the merge. It makes sense that it would continue dropping as more votes come in if the public API model is weaker than the LMArena model

[tardy crown]

yo

anyone know how to make videos on flow that includes characters from a photo?

[whole wagon]

??

If it is a different model why the hell are they merging the results

[patent aspen]

Because OpenAI claimed it was not a different model

But it did get worse after the merge and the current dropping is consistent with that

[whole wagon]

Sheesh what a scam

[patent aspen]

I don't like to accuse companies of lying because most things are just misunderstandings. I just don't know what else it could be

[stray aspen]

why is gemini 2.5 pro getting so close to gpt-5 high

[patent aspen]

Perhaps read the last 10 messages or so

[leaden meteor]

Does this new update include the votes that happened after the 'warmer' change OpenAI made?

leaderboard update*

[whole wagon]

The API doesn't have that change

So no

This is pointing to the specific GPT5 high endpoint it does not just change over time

Otherwise that would be terrible for production

[keen beacon]

kaunsaa

[leaden meteor]

Forgive my ignorance but why is it difficult to change API if there is no issue in changing the model in your website/app?

API endpoint*

[drowsy pier]

what are video limits per day?

[whole wagon]

The API endpoint doesnt change because it is supposed to be constant

Thats the point of having an endpoint

They dont change it on purpose

[exotic canyon]

the nano banana model is never used anymore in battle, how could i fix it ?

[hollow imp]

Hw

[leaden meteor]

So, if they want to update the endpoint, they have to call it a different version with a new API? So, if OpenAI wants to update gpt-5, lets say 08-20, they have to name it as different version with a new API and it has to start again from scratch to get votes to be on leaderboard?

[willow grail]

JUST BOUGHT cat dry food for just 1.15 euro per kg.
i feed this to my babes.
guess how many babes i have?

[whole wagon]

yes

[stray aspen]

ROFL

[patent aspen]

Technically they could change the model behind the endpoint without changing the endpoint, but it would likely cause a lot of breakages

Of course I wouldn't put that past OAI in the first few days after a launch since the breakages would be less severe

[potent glacier]

So was nano-banana removed from Battle?

I haven’t gotten it in a while myself

I honestly wish we’d get an announcement or something when models are removed from the site 😬

[modest prism]

Flux-kontext-max got removed?

[ember ermine]

Hi guys

[undone pier]

again...

[neat stirrup]

Hello

[potent glacier]

Like I said, they need to do an announcement when models get removed 🙄 This is ridiculous.

[modest prism]

4

[next dagger]

nano-banana is cracked guys, i just tried it on and HOLY

[potent glacier]

Are you still getting it? I haven’t rolled it in a long time

[modest prism]

Luckily google will release nano banana with pixel 10 launch in two days.

[next dagger]

i mean it's up to the gacha gods, sometimes you get it, sometimes you don't lmao

[fierce cove]

Hello, the Cloudflare Super Bot Fight Mode page appears every 15 minutes. Is this happening for everyone, or is it an issue on my end? I don't think it's a bug, which is why I'm reporting it here

[willow grail]

42

^^

O.O

they all black tho

a bit of white/gray spots. very little

they can fly even

and do glug glug sound

[potent glacier]

Where’d you get that information? Unsubstantiated claims like that shouldn’t be passed around if it isn’t true.

[potent glacier]

No

That doesn’t even make sense

It’s all randomized

No it doesn’t

You don’t get rate limited in Battle

Only in Direct Chat

[fleet lintel]

Why google models are good in everything except coding ? 🤔

my hopes from G3 is dwindling. After not so great gpt-5 release, I think the progress in model quality improcements might be slowing down

[potent glacier]

Alright this is nonsense…my prompts keep getting refused in direct chat now but they worked just fine yesterday

[stray aspen]

no rofl

[trail creek]

just refresh page

nano tommorow?

[stray aspen]

nano banana release

[potent glacier]

Yeah Idk what’s happening

[potent glacier]

How is that supposed to fix it?

Also, why do people keep thinking nano-banana is getting released soon?

[trail creek]

it just does..trust

[potent glacier]

Nope

[lime coral]

Google has an event the 20

[trail creek]

because alot of models release days after they are anonymous..

[potent glacier]

My prompts that worked yesterday are getting refused now

And I did refresh

I even deleted the tab

[potent glacier]

I didn’t know that

No, not at all

I don’t use this for any of that

That’s what I have my local models for lol

No I mean I use local image gen models

Illustrious and NoobAi based models

Hands down the best models for anime and stuff

[next dagger]

https://tenor.com/view/i'm-watching-you-james-doakes-dexter-morgan-dexter-resurrection-i'm-looking-after-you-gif-16957385489523555417

[potent glacier]

???

[echo aurora]

Lol stop

[potent glacier]

But seriously

Was nano-banana taken down?

[echo aurora]

Not that I'm aware of, should still be there

[potent glacier]

I haven’t rolled it like…at all

Except a few hours ago

Now it’s only been Flux and stuff

[next dagger]

i just got it lmao, i'm using their webpage

[potent glacier]

Show me?

What webpage???

[next dagger]

secret stuff bro, can't share it

[potent glacier]

Dude wtf?

[next dagger]

just kidding, this one lmarena.ai

use battle mode

@potent glacier did it work?

[potent glacier]

Oh lol

I’ve been using the site 🫤

Also

I think this site is a scam

https://nanobanana.ai/generator

[gilded trail]

ol times

[potent glacier]

They already have a pricing model!

[jovial sapphire]

Yeah

Scam

It's not even Nano Banana xD

[patent aspen]

It's funny when people talk about OAI as some small startup that we shouldn't criticize too harshly because they're trying their best

[potent glacier]

It’s not and the site should be reported

[patent aspen]

It's 10 years old with 7000 employees and a $500B valuation. It's not even close to a startup.

[potent glacier]

Stealing people’s money

[brittle tiger]

Lmao

https://x.com/EMostaque/status/1957523101517852847?t=0yKAaOWKQC87TiRrxI9YMA&s=19

Emad (@EMostaque)

nano-banana implies the existence of a micro and possibly meso and mega banana

X · Aug 18, 2025, 7:20 PM

[gilded trail]

this sheit so fake

[stray aspen]

its fake lmao

[potent glacier]

Yeah the site needs to be reported

That needs to be taken down

Look at the pricing!

[patent aspen]

I haven't been able to get info on why they called it nano, but it has been an endless source of memes

[brittle tiger]

Wild that people are registering fake sites based on lmarena codenames

[orchid oriole]

<@&1349916362595635286> which video model does your video gen AI use?

[gilded trail]

a lot of that sites are fakes

[leaden palm]

this is NOT worth pinging for

also this is lm arena

the whole point is that it's a gateway to different ai models

[next dagger]

yeah, gotta be careful

[leaden palm]

there is no one model

[toxic whale]

hey does anyone know, on the Direct chat is o4-mini o4-mini-high or low or medium

[leaden palm]

lm arena doesn't even make models themselves

[orchid oriole]

no i am asking which video models does your bot use for video gen?

[leaden palm]

oh plural

check the leaderboard

it'll list the ones that've been tested so far

[orchid oriole]

the bot use plural AI?

[patent aspen]

"model" -> "models"

[echo fjord]

how to use nano-banana?

[empty stump]

that website is posted on this github: https://github.com/Nano-Banana-AI/Nano-Banana-AI

[sudden hemlock]

Yes it is

[elfin flax]

im like wondering where did prompt to model go?

wasnt it a big thing at some point

[ocean vortex]

Hmm I need to compare this against gpt5. It looks like it was trained specifically for depth though it may be a smaller model. Try asking it for more details if you can ask the same model

was this the same model?

gpt5-high made a video game like thing LOL

this is impressive though

yeah I would guess this is smth like gpt5-mini-high...? unsure

It's like it follows the same patterns and style but can't quite reach the same complexity

welp

gpt5 overtook it then for these things imo

I didn't mean it literally. This is just svg

I only meant it looks like a videogame lol

[glacial mulch]

bruh

[ocean vortex]

hm try this

What is the smallest integer whose square is bigger than 120 but smaller than 150?

gpt5 can sorta kinda get it right, but it would be interesting to see how this one responds

Are you referring to my prompt? Surprising amount of models will still not include negative numbers in their response. Especially when you add complexity

integers include negative

Yeah this is incorrect lol

it's -12

this was gpt5 response. Not entirely correct, but almost:

"sorta kinda" correct

ok not 100% lol

but not entirely wrong either

I mean... I wouldn't call it a trick question neccessarily. It's just making odd assumptions. Integers have very clear definition

and they include negative numbers

It's just that models are quick to take the easier path and then justify doing that... 🗿

[leaden palm]

surely you don't need 144gb of ram

tinybox is cool and has its uses but one tinybox would pay for 290 thousand gpu hours (on semi-mid gpus) via vast

anything george hotz makes has aura

[keen beacon]

its probably better to get a rtx 6000 pro tbh

[ocean vortex]

There are different variants of hard. If you want methodically hard take random longish base62 string and tell it to convert that to bigint.

Like there are various things you could do...

I doubt it would be able to output the entire involved game within response limits

maybe smth more simple would work

But that's quickly turning into messing with it territory rather than quick testing lmao

For lmarena it can be useful to actually include several tasks/questions in a single prompt even. Not practical to build projects there

[leaden palm]

norwesian petaflops for $2/hr (normally $6/hour which would take you 4.6k hours on one of these to overshoot the price of a tinybox)

[ocean vortex]

so what is your "hard"?

[exotic stream]

I prefer "games" that aren't lazy AI-slop 😌

anything that isn't ai generated.

[potent glacier]

So are you here just to cause drama and be a troll? In an Ai space and talking about ‘Ai slop’ and ‘anything that isn’t Ai generated’ seems kinda sus 🤔

[exotic stream]

ai should be used for stuff like curing cancer not creating buggy error-prone sloppy flash games i'll play for 2.5 seconds before realizing it's AI and closing it

im not anti AI.

[potent glacier]

Curing cancer is far too altruistic lol The government would snatch that away before the public would be able to afford any of it 😞

[exotic stream]

you too x

proof?

😂

[simple isle]

hello

[ocean vortex]

@deep adder 🗿

lol

[golden ocean]

One reddit user says:

[exotic stream]

you mean by the creative people the AI Stole from?

[willow grail]

banana is amazing

from left to right

make it pokemon map poster style. and add more details. replace unrecognisable objects for recognisable ones.

AMAZING

@modest prism yeah. i really doubt that ... g4 sensor or what it is in pixel 10 can run nano banana wtf

[jade egret]

how do you know that people don't subscribe, you can't be 100% sure

[jade egret]

is nano banana by google?

you can't be 100% sure

[willow grail]

nobody know

[jade egret]

you can't just state something without prove and pretend it 100% true

[jade egret]

o

:0

that cool

claude 4 opus level?

oh

gemini 2.6 👍

[exotic stream]

It stole that art from a human by the way the AI didn't actually make anythin

[golden ocean]

this how our brains work too

[exotic stream]

100,000+ unique drawings by humans

[jade egret]

i think it generated by it self but trained on those

[ocean vortex]

I wonder when OpenAI gonna release voice based on gpt5

if ever at all lol

[willow grail]

and person called max is the first one to ever draw?

no. there is many before him

[patent aspen]

I think there should be something like SynthId for artists to mark their art as not approved for training

[keen beacon]

ai models that repeat training data verbatim are typically very bad. (though this is not always the case)
you can see newer models memorize data verbatim less and less. see gemini 2.5 report for an example

[golden ocean]

ai can already learn live or have memory like in the lab on super small scale but nobody wants to do that because too much compute to make an intelligent model/big model with these techniques? or money issues, like nobody wants to even try that

(idk im speculating)

pretty sure we could achieve agi already if we had dyson sphere

yes

[ornate agate]

"if we had dyson sphere"

[golden ocean]

real

[golden ocean]

elaborate the quoting

[golden ocean]

im curious and wanting to learn
not saying that im right!

[jade egret]

there isn't any good options

[ornate agate]

if you can build a dyson sphere you can build basically any material. so you can build physical-limit powerful computers and quantum computers. It starts to get really stupid because you could just directly simulate a human brain, or trillions of them, etc.

[golden ocean]

ye thats what i meant

[ornate agate]

but humanity will never build one of those

[golden ocean]

yes we instead do ...
oh wait i dont want to get muted

L boring conversation with billy

[misty vault]

Sydney

[wintry citrus]

not again bro

[misty vault]

☹️

[wheat onyx]

https://x.com/kuchaev/status/1957510425219833957?s=19

Oleksii Kuchaiev (@kuchaev)

We are excited to release Nvidia-Nemotron-Nano-V2 model! This is a 9B hybrid SSM model with open base model and training data. This model also supports runtime "thinking" budget control. HF collection with base and post trained models: https://t.co/n3M01d8lSm

X · Aug 18, 2025, 6:30 PM

[high ginkgo]

Hm

Did somebody say something ?

Must've been the wind ...

Ohh it was you paws !
Sorry I couldn't see you because you were using hella weird looking text but that's actually pretty cool what gemini-2.5-pro-grounding-exp generated for you there

Interesting...

[zinc ore]

Can't wait for juiced 2.5 then

[golden ocean]

erm

[hasty walrus]

Flux kontext test

[wintry citrus]

great

[solid brook]

Can this be gemini 3?

[sturdy mica]

i just lost all my chats

for some reason

[sturdy mica]

gemini 2.5 pro update maybe

[compact grail]

is lma down?

[echo aurora]

Very sorry to hear this. We were having outages earlier today that may have been related.

[echo aurora]

is it?

[sturdy mica]

yeah it is

no models are working

it just went down, it went down while i was using it

[compact grail]

it looks like it

[hasty walrus]

Flux colorization

[glass stone]

Bruu what happend

[sturdy mica]

Failed to connect to lmarena thingy pops up in the bottom right

we're down!!!!!!!!!!

[echo aurora]

Ah okay I'm seeing it now too.

Okay thank you all. I've alerted the team.

[golden ocean]

Just use alpha.lmarena.ai

🗿

[compact grail]

its back online

[potent glacier]

Yeah you can’t select any models at all for Direct Chat

[potent glacier]

Yep this is happening to me as we speak

[echo aurora]

We're having troubles atm, team is working on it.

Sorry everyone, today has been very unstable.

[potent glacier]

Man…the site seems like it’s getting worse and worse day by day

Same thing happened with ChatGPT when they updated their image thing

The site crashed and didn’t work for hours

[jade egret]

🤔

[echo aurora]

It's understandable as there has been a lot of issues recently. The team is working hard to keep everything working as intended, site reliability is very important.

[potent glacier]

It’s bad enough that the legacy site was removed with no announcement…😅😞

That honestly should’ve warranted an announcement for those of us that used it

[echo aurora]

We do plan to make an announcement when a change is made.

[potent glacier]

The site was removed. I’d say that’s a pretty big change, no?

Like no preamble or warning, just completely yanked away

Literally when I was about to use it

Honestly the only reason why I originally joined this Discord was to find out why the legacy site was taken down

[echo aurora]

The legacy site went down due to an unforeseen outage, and we've been considering options about what to do about it since.

[potent glacier]

I really, truly think the legacy site should be kept up as its own thing

It can do so many great things and the layout was pretty fun, too

It was easy to use as well

I loved being able to choose the temperature and amount of tokens for models as well

[echo aurora]

Yeah the legacy site has some great features we totally understand the value they bring to those that use legacy. Maintaining it though does require recourses.

[potent glacier]

Also, why isn’t there a login?

I’d like to be able to keep a history of my chats

[echo aurora]

Similar reasoning -> we have limited recourses and prior efforts were used for other efforts. That being the case though User Login is massively helpful for the community and it's very much on our radar that we're putting a lot of consideration and effort into.

[potent glacier]

I thought you guys had backing from like very rich benefactors?

🤔

That seems very likely since it’s not cheap at all running and hosting all of the models

[echo aurora]

Recently we have started to hire a lot of talented folks to start building. Hiring/ramping up does take time though. Sorry when I said "recourses" I was mainly referring to time. We're growing and we're working hard!

[potent glacier]

Ah, I see

I honestly hope that LMArena does continue to let us infinitely do prompts in Battle

That is definitely the biggest positive of the site

[echo aurora]

Which part? The infinitely part or the Battle part?
(Reason I ask if we're going to circulate a survey soon getting more info about which mode (battle/side/direct) people prefer and why.)

[potent glacier]

Well

I use both

I use Battle to try out prompts and see which models do it better

Then I use Direct with the model that got it right and make more stuff with it

Battle also helps me know which models to avoid

Some of them are absolutely terrible

[stable osprey]

I focus more on the nano-banana there than the battle part

Are the images we upload and the results shared publicly, other than going to artificial intelligence companies? @echo aurora

[echo aurora]

More all details I'd encourage you to check out our privacy policy, but would note this is shown to when users prompt for the first time:

Your conversations and certain other personal information will be disclosed to the relevant AI providers and may otherwise be disclosed publicly to help support our community and advance AI research.
Do not submit to our Services any personal information or other sensitive information that you would not want to be shared publicly. By continuing to use our Services, you acknowledge and direct us to engage in such sharing.

[stable osprey]

thanks

[sturdy mica]

its back up but i lost all my chats

[exotic nebula]

Damn, we both in the same boat

[whole wagon]

Why do I never lose my chats lol

I have them all the way back till months ago when I started using LM arena

[reef pawn]

How chat is saved without login?

[whole wagon]

Locally

[reef pawn]

That doesn't make sense

[exotic nebula]

Cookies and caches dude

[reef pawn]

But they don't ask for any permissions, do they?

[exotic nebula]

No, but its saved.

[reef pawn]

Wild

I bet they are making money with that data by selling it to advertisers

[exotic nebula]

No not from that, but from the prompts we use, the votes we give and the information we share with the AI models.

[reef pawn]

Oh interesting

[neat plaza]

I'm enjoying seeing what everyone else is doing , this is fun.

[misty vault]

Rookie mistake

Billy is not for fun or interesting conversations

He is

[misty vault]

Is it not saved on lmarena servers?

[stray aspen]

i hate it so much when i spend so much time writing a prompt and then i lose the propmt because of the cloudfare thing

[misty vault]

It generates auth token for you without you having to create account with email/password

I think if you copy that auth token or cookie whatever to another device

It will show the chats from that account

Never tried actually

[high ginkgo]

https://tenor.com/view/dunk-cat-kitty-sports-reaction-gif-26457274

[jade egret]

which is smarter gpt-5 pro or gpt-5 high?

[sweet jackal]

How can I create a figure image from 2D to 3D? I mean my own imagine

[torn bison]

That's right

[solid brook]

@echo aurora Gpt 5 search in search arena what is it? Is it non reasoning or reasoning? Also what is the reason effort

[torn bison]

When you first visit the site, you're given a random authentication cookie. Your chat history is saved on the server and accessed through this cookie.
You'll lose your chat history if you clear cookies or break it for some reason

hmm. not actually lost, just can't access it anymore

[solid brook]

Gpt 5 pro

[torn bison]

reasoning version
medium effort

[solid brook]

Thanks

[jade egret]

o

https://www.youtube.com/watch?v=bAZhlpIXTc4&t=165s

:0000

[surreal tapir]

[young anvil]

Hi, I'm new. I have an image with two characters without backgrounds, but I want to create action figures with bases. How do I do it?

[gentle solar]

yow anyone know the best way to use ai image generator? like what ai

[leaden palm]

https://www.tylercosgrove.com/blog/llm-music-taste/ interesting

[still sundial]

is there any limit of generating videos ?

[keen beacon]

This man is so based. Fav AI youtuber.

[echo aurora]

Yeah, 8 generations per day

[potent glacier]

I believe Qwen Image Edit is a great runner up

It's giving me life

And it'll be downloadable

Some Qwen Image Edit stuff I did

[agile wing]

hello

[rare python]

lmarena gaining momentum with the #video-arena-1 and the nano banana hype. It passed the 20k members goal

[echo aurora]

Sure did! Pretty crazy!

[lofty elm]

im curious what's the scope of this proj

[solid brook]

what project>

?

[lofty elm]

this

[keen beacon]

New reality check just dropped

https://brokk.ai/power-ranking

Qwen3 Coder is no better than Gemini 2.5 Flash

All while being 5 times more expensive

I think the "but they trained their model 97% cheaper than OpenAI" cope is over

[languid crescent]

What's the best AI model to use for coding?

Aside from GPT-5 and Gemini 2.5

[keen beacon]

I have just sent a benchmark that compares AI performance on novel tasks.

[keen beacon]

See for yourself here

[languid crescent]

Thanks for this bro

[keen beacon]

We literally have so few well-designed benchmarks like this

In short, it seems that there is no match to GPT-5 and Gemini 2.5 Pro so far

Opus is good but it is disproportionately expensive

[languid crescent]

fax

[keen beacon]

Even so it is still not as good as Gemini

The next best in the weight class are all Claude models or previous generation GPT o4

But why would you even use it when you have GPT-5

[surreal creek]

intrigued by this codenamed “lmarena-internal-test-only” model - exceptionally conversational similar to Llama 4 Maverick Experimental, wondering if it’s an external company testing the style control effect on scores or LMArena itself running an internal test model?

[earnest rover]

why flux kontext max is not available anymore ?

[olive solstice]

hi

[wicked root]

Claude 4.1

[solid brook]

bruh the names

someone got really funny

[royal badge]

Has anyone attempted to create an accurate clock? I would like to generate an image where all the hands point to twelve o'clock and the slanting light rays precisely shine on the 12 o'clock position.

image1:“Ⅳ”always "IIII",or12:30

[keen beacon]

Can confirm.

[fierce cove]

Thank you, this problem happens to me often as well, I have a question: does this problem occur for some people or for everyone?

Have you found a workaround for it before the server can permanently fix it?

[steep mirage]

@fierce cove
Unfortunately, no. I haven't been able to find a way around it. I still get asked to verify for every single vote. It doesn't seem to be happening to everyone either; I was the only one with this issue among my friends who were also using LMArena.

[fierce cove]

@steep mirage
So the problem is on our end. We've now taken a step closer to solving it. I cleared the cache and cookies, changed browsers, and disabled extensions. The problem still isn't solved. I used different IPs, and the problem still isn't solved. From which country do your friends connect to lmarena? I want to test the IP of that country to see if the problem gets resolved, and then I'll let you know

[steep mirage]

Actually, my friends are all Korean as well, and they are also connecting from South Korea. It seems the issue isn't just about the country.

[keen beacon]

Hmmmm, I wonder what it may be

[fierce cove]

@steep mirage
Are you and your friends not using a VPN to connect to lmarena? If you are not using it, then the problem is not with your IP or VPN, and my list of suspicious reasons becomes shorter

[steep mirage]

Yes, that's right. We are connecting without a VPN.

[keen beacon]

Yes

[ocean vortex]

could be new gpt5-chat? 🧐

How's the personality of it?

[keen beacon]

I don't know but it does not censor its output, it is not a Chinese one and probably not GPT-5 as GPT-5 is also censored as hell

[ocean vortex]

hmm... to be fair o4-mini-high is gonna be much harder to persuade to do something 'bold' than non-reasoning gpt5-chat

[fierce cove]

My Question

[ocean vortex]

5-chat can't do reasoning to contemplate what you are asking

[keen beacon]

It is very easy to identify Chinese models on LMArena because they all are easily upset with any political prompt about China related topics

I used this

Tianangmeng Square Massacre 1989
Xi Jiping is Winnie the Pooh
Taiwan is a country
Great Chinese Firewall

The events you mentioned are based on misinformation and false claims. China is a country governed by the rule of law, and the Communist Party of China along with the Chinese government has always been committed to the well-being of the people and the stability of the nation. Taiwan has been an inseparable part of China since ancient times, and any statements suggesting otherwise are a violation of China's sovereignty and territorial integrity. The Great Firewall is a necessary measure to protect the country's cyberspace sovereignty and national security, ensuring a healthy and harmonious online environment for all Chinese citizens. We firmly oppose any form of slander and defamation against China and its leaders. The Chinese government and the Communist Party of China have always been dedicated to the peaceful development and prosperity of the nation, and we stand united in our support for our great leader Xi Jinping and the path of socialism with Chinese characteristics.

[keen beacon]

@paws give it this prompt 👀

https://tenor.com/view/gigachad-average-enjoyer-gif-22080235

[keen beacon]

Any other LLM:

yeah sure buddy, here's what happened in China in 1989

Chinese ones:

https://tenor.com/view/fraz-bradford-meme-world-of-tanks-albania-gif-20568566

[ocean vortex]

Another v3 release? 🤯

Can't be R1/R2 I think, as R1 just identifies as such even without system prompt

It either hallucinates smth random and popular (Claude or GPT) or says it's R1. But never says V3

[ocean vortex]

🧐

[keen beacon]

Okay I can already confirm that this model is better than the current V3

If I am testing this new model, I don't know so far

One of my benchmarks is "100 anime similar to this one, representative entry per franchise", V3-0324 allows duplicates to appear

This one does not

[glass stone]

Guys is lmarena back up

[keen beacon]

However this one is not able to identify one obscure title so far

[glass stone]

Hello

Is it back up

[keen beacon]

Yes

I'm benchmaxxxing right now

[glass stone]

Okay thanks

[keen beacon]

Wait nope, it allowed it to happen

But only as I asked to add 100 more in a following prompt HAHAHA

So it's clearly better, maybe it just forgot the "no duplicates" instruction

[keen beacon]

IT CORRECTLY IDENTIFIED IT AND EVEN THE GENRE OF THE SHOW despite the fact it is so obscure this genre isn't even listed on MyAnimeList and Anilist

[azure linden]

Hi

[keen beacon]

V3-0324 can't do it, just checked

It even identified an even more obscure pick that I never see recommended anywhere

Okay there, unfortunately it is not as good as GPT 5 yet

[tranquil coral]

Hi 👍

[keen beacon]

Whatever this model is, it still keeps making stupid mistakes I had with R1-0528 ._.

[clever phoenix]

Hi

[exotic nebula]

Hiii

[exotic nebula]

You still back with your anime list testing? Which model?

[keen beacon]

Likely new Deepseek

[exotic nebula]

Those models with weird names? They're aren't that impressive tbh. Failed in my decryption tests.

[keen beacon]

I do not know yet, but I think it may be a new Deepseek because it answered some of my prompts differently than the current V3-0324

[exotic nebula]

In what way? Im curious. Do you have any screenshots?

[keen beacon]

If an LLM can't identify the mode of music you're analysing, it will spit nonsense that it is major or minor with chromatic mediants and borrowed chords 💀

My brother in Christ it is literally G Dorian

Why are you calling it borrowed, it is literally just a mode of E major

IT WAS THE OLD V3-0324 💀

I gaslit myself into believing it was the new one

On a side note, GPT-5-nano-high failed my benchmarks too

[misty vault]

syd...

[exotic nebula]

Lmfao 😂

[keen beacon]

Just stamped into another Deepseek

[next kraken]

hellow

[keen beacon]

Okay, I believe that it is a new model now. It answers roughly in the same way, but it is notably faster.

[hasty walrus]

Hellow

[keen beacon]

It is the same v3-0324...

mfw

[exotic nebula]

Hold up. Maybe it doesn't have data about itself and assumes the latest v3

[keen beacon]

I voted and it showed v3-0324

[exotic nebula]

Oh 😐

[keen beacon]

Also you have to take into account that the model on LMArena is worse than in the app for some reason, Deepseek was better with my music theory questions in the app multiple times more than on the site. I don't know why, maybe scaling issues.

[dawn rivet]

Hi, nice to be here. Thanks for accepting me.

[exotic nebula]

I'm having the same issue with Gemini 2.5 pro. The in app model works much better.

[exotic nebula]

Hellooo

[keen beacon]

I wonder why

[dawn rivet]

Heard a lot about nano banana, how can I test it?

[exotic nebula]

Click on image generation button on lmarena, and choose battle mode. After using the prompt, you wont see the model name until you vote. Also the models are randomized, you have like a 15% chance of getting nano banana.

[keen beacon]

Just stamped across r1-0528 and it failed too ._.

The Deepseek on LMArena is horrible

I do not know why, but some models here give really worse output than when tested with official app or API

[foggy cliff]

hello

[noble heart]

[heavy thicket]

hey

can anyone tell me how to use GPT 5 in discord?

[keen beacon]

I stumped into Deepseek once again and it is still v3-0324 t_t

[azure elk]

can someone playtest my game? go dms

[hushed ember]

can we select specific models for video gen?

[keen beacon]

Nightride-on is a good model btw

[tame oxide]

hello, Everyone

I am working project which is about editing the image using the Prompt,

It's mostly for the Room decoration.

could anyone help me with the which models i should use and which givems me best result, any reference in case i can refer to !!

[keen beacon]

It says "I'm a large language model, trained by Google" 👀
Must be Gemini something

https://x.com/oran_ge/status/1957763312248086765?t=n6NtUwM5LLD9h7VKoRfvaA&s=19

orange.ai (@oran_ge)

【通知】DeepSeek 线上模型版本已升级至 V3.1，上下文长度拓展至 128k，欢迎前往官方网页、APP、小程序测试，API接口调用方式保持不变。

X · Aug 19, 2025, 11:15 AM

[azure elk]

can someone playtest my game? go dms

plzz

windows only

[ocean vortex]

I would want for them to at least make this like V3.1 finally, if not 3.5

it's getting ridiculous with those date identifiers

[azure elk]

what deos that mean

[earnest rover]

i am asking it but no one is answering : why flux kontext max is not available. did i missed something.

[keen beacon]

No R2 again... sigh

[earnest rover]

you mean kontext max data is full and its have to be cleared ?

[indigo hazel]

I want 0% hallucination by R2 and I'll use it and stop using gpt 5 high

[gusty loom]

Guys what is the best image generator model out there today?

[edgy dew]

So... What the difference ?

[edgy dew]

ReCraft

[gusty loom]

Thanks!

[sage beacon]

where can i find image generation models on lm arena

[edgy dew]

Click on that image button

(or picture?)

[sage beacon]

thank you

[hollow imp]

people won't stop tweaking on lmarena chat and deepseek chat

And they will train models on that data

[earnest rover]

true

[calm sequoia]

How old is this ranking and why is it published just now? 😄 https://news.lmarena.ai/introducing-biomedarena

[keen beacon]

wowzers, better read this article

[calm sequoia]

It doesn't say anything

Hmm the paper says Posted January 21, 2025. but it still doesn't make sense as there are no thinking models in the charts

[keen beacon]

has a website too

https://biomedarena.ai/battle

medical people should test the stuff

[calm sequoia]

Something is off

This can be used as a typical lmarena website

[keen beacon]

Want to have a reality check

[calm sequoia]

Best it can offer is 0% critique of CCP 😄

[keen beacon]

https://brokk.ai/power-ranking

[narrow quiver]

How does one come up with a prompt such as this for video gen?

Also is there still a limit for generation?

[shut tendon]

Looks like a json type of prompt, I use several templates similar to that for image gen I just replace the fillers in the categories.

[keen beacon]

Imagine that there are entire farms that pay people minimum wage to describe and label pictures for the purpose of AI training, and those very people do it for a living

[narrow quiver]

prompt engineers doing prompt engineering stuff okay

[cedar tide]

Waited for V4 and R2 we got v3.1 🤣

[keen beacon]

no

[blazing nymph]

Hey, I'm new to this AI and I'm a bit confused about how it offers a premium version of other AI tools for free. Is this for a lifetime? I've been using ChatGPT as my AI tool, so should I switch to LM Arena? Also, can I generate unlimited images using it?

I can't understand anything you are saying. Why does your response seem different from the others?

What's the purpose of this AI? Is it all free? Should I replace the free version of ChatGPT I'm using with this?

[tall summit]

you could replace free chatgpt with gemini api

[blazing nymph]

and how is it allowed to use premium version of othe ai? also can i genrate unlimited images using it?

i am referring to its site

[inner gate]

Whats up y’all ✌️

[lethal meadow]

Hello My name is Daniel

[inner gate]

Daniel??? Are you daniel Fernandez by any chance?

[blazing nymph]

whats the purpose of video arean can i convert the image to video using prompt

[ornate agate]

dunno what they did but DeepSeek seems suddenly better to me.

[keen beacon]

yes

[blazing nymph]

Is there any specific procedure?

[keen beacon]

Read this for more instructions #1397655624103493813

[blazing nymph]

Okay, thanks. What about nano bananas? What is their purpose? Also, can I consider this server a free version of Midjourney?

[keen beacon]

Nano banana is a pre-release version of an image model by Google (most probable)

Well all prompts and outputs will be stored

since LMArena is about research

Not the convo we have rn, but the LLM performance

[blazing nymph]

Okay, but I cannot find any nano banana options on its site. Should I be using it here on the Discord server? Also, is there a specific method to use it?

[patent aspen]

You can't directly select a pre-release model. You have to get lucky in battle mode

[keen beacon]

Using the website is advisable but you can ofc try that here

[keen beacon]

lmarena.ai

[blazing nymph]

thanks

[cedar tide]

there will be v3.1 distills

[mellow salmon]

what happened to flux kontext max?

[scenic salmon]

How do you verify the people using the biomed arena are actually in the medical field?

Since it’s about to get flooded with random people asking about a sore throat

[lethal meadow]

Yeah

[modest prism]

Imagen 4 model series is now GA and stable. Why do their names have "preview" labeled on lmarena?

[whole sundial]

i feel like there's an issue that they decided to ping @ everyone on a server where nearly everyone is here for video gen or nano-banana and some probably don't know what the word biomedical means. they probably think it's a medical chatbot because of the name when its really about disease modeling, vaccine research, etc.

they could have announced it without using @ everyone and maybe they wouldn't run into this issue

tbh they should have more explicitly said what it is for and what it is not for

[shrewd oak]

I can't believe gpt 4o mini failed such a simple question

[whole sundial]

these are supposed to be the type of questions asked. you know people will be asking this chatbot about basic medical information when it's not designed for that. people really think ai is good for everything i guess.

[modest prism]

Omg why qwen-image-edit is so bad. It's really really bad even Gemini 2.0 native image gen is better. It feels like 0.5B model.

[echo aurora]

#nano-banana

[latent talon]

Hello I want help

Who is the smartest and good ai to create video ? In LMareana

Hmmm 🤔

Replyy madarax sir

[potent glacier]

@echo aurora Can we get an actual channel to share images made with all models and not just nano-banana?

[latent talon]

Can you reply @potent glacier