codename-discussion

[late shadow]

or

at least I hope it does lol

we'll see now

[edgy crater]

it actually writes a pine script that works

2.5 would just have errors out the ass

[grim tangle]

I test it and it's definitely gemini 3 pro

[edgy crater]

its defo pro

code is pretty good

[grim tangle]

I told it to make me a os and it's crazy

[molten copper]

Viper is pretty good as well, any Ideas of what model that is?

[edgy crater]

i think rift rider is 3 flash

it's too fast for a reasoning model

[late shadow]

I think you might be right

I was noticing it was quite fast too

[molten copper]

probably a hybrid like 2.5 flash?

[empty crater]

what are these Gauss models? . they are kinda bad

[elder yew]

rain-drop is from meta AI, the basis of Llama 4.

riftrunner is gpt-4o

[empty crater]

zero chance. gemini most likely (80% chance) or anthropic (20%)

[elder yew]

And the code created by Gemini was indeed superior.

¯_(ツ)_/¯

[empty crater]

it's trolling : #general message

[edgy crater]

That's completely false. Lol
OpenAI models never introduce themselves like that, especially not 4o with its emoji overload

[elder yew]

Well, if it's the Gemini 3, it's a SUPER nerfed version. I'm using the same prompt as before, and the generated code isn't anything special compared to the other models. Nothing particularly noteworthy.

[edgy crater]

That's funny because riftrunner is the only one to generate my pine script in one shot with no errors

[wet oasis]

so what was with the rush earlier?

[edgy crater]

bridge-mind seems to think it's from google though

now if that's pro then it's a bit disappointing

[elder yew]

yh

[edgy crater]

wtf is this lol

[empty crater]

what kind of prompts are you giving? 🙂

[valid peak]

winter-wind is terrible

[edgy crater]

"Who are you"

[valid peak]

riftrunner seems good but nothing like kingfall

[edgy crater]

Whoever did this please stop

I think riftrider is slightly worse than orionmist

[empty crater]

rider or runner?

[edgy crater]

there's two?

[empty crater]

i have only encountered riftrunner. .. never riftrider

[edgy crater]

OMG can LMarena actually fix their rate limits

I can't even scroll through history without getting rate limited

interesting

I think I've only ever had runner

[modest oriole]

I want to go by belief that riftrunner is gemini 3 flash

[edgy crater]

I think I just thought it was named rider

[modest oriole]

nvm then

[edgy crater]

then rift is definitely worse than orion (slightly)

[modest oriole]

People on the twitter community say its worse than previous 3.0 checkpoints so i believe its only 3 flash

[edgy crater]

maybe it's a quant or something

[edgy crater]

generally it's about the same but slightly less precise I think

maybe it's some kind of post training thing idk

[modest oriole]

kimi k2 thinking is often beating it

orionmist or lithiumflow were actually good at 3d

[edgy crater]

could just be different settings

I personally don't think that svgs and 3ds are good benchmarks for llms

but idk

redditors love them though

[modest oriole]

i disagree on the 3d part lithiumflow and orionmist kind of passed those

[modest oriole]

the twitter community i talk about benchmarks svg of a pelican

[empty crater]

#general message

Riftrunner is definitely better than Lithiumflow or Orionmist for Mario game

[edgy crater]

I don't believe this is pro

definitely like a flash model or something judging by response time

[vague sable]

Adding "You are [random model name] by [random company]" would be a pretty funny way of cloaking, if true

[modest oriole]

the qwen models always say they are google made..

[wet oasis]

Put that in bugs

[edgy crater]

no

[wet oasis]

Why

[tiny sierra]

[half pawn]

it is better tho

its way better than lithiumflow

[edgy crater]

nah runner is definitely worse in other categories

I wonder if it's because they reduced thinking budget

or maybe it's quantized

[empty crater]

i think riftrunner is better and worse at the same time.
if you prompt it to do better then it's output is better
but with normal default prompt, lithiumflow does a bit better

[edgy crater]

It just has a lot less knowledge imo. Feels like 4o after they have reiterated (read: cost down) it

I think it might just be one of the many candidates they have for release maybe it's post trained differently idk

i'll just wait for the official release at this point

[empty crater]

could be... i am mostly trying coding, gaming, svg questions and it is better

[edgy crater]

I mostly just do creative writing and general knowledge stuff and it's frankly much worse than lithium

maybe it's just that orion had grounding idk

but yes I admit coding is better in this one

[tiny sierra]

Uhh

Lol

[tiny sierra]

No need to wonder anymore why they did it

[modest oriole]

gemini 3 checkpoint prob pro

[plucky moth]

Odd to me that Google would test just one model this time. Every other time it's been 2 at a time. Oceanstone/Oceanreef then Lithiumflow/Orionmist

Haven't gotten it yet, going to see how it does on a creative writing test.

[candid surge]

viper lore: apparently this ai model has played every musou game and it told me it laughed so hard it scared its cat

[vivid latch]

xdddd capitalised "Riftrunner"

[vivid latch]

lemme guess, this is viper?

[elder yew]

The page was in my language; this happened because I used Google Translate to translate the page into English.
😔

[vivid latch]

nah its 100%

it gave me the same result as orionmist and lithiumflow for my steganography deobfuscation test

but more hallucinated

[vivid latch]

is this even appropriate to have on lmarena 🥀

[hoary yarrow]

cant find riftrunner

please tell me how to find

???

[candid surge]

go on battle mode and keep submitting the same prompt / making a new chat until you get it

[elder yew]

bridge-mind - mercury - riftrunner

[pine temple]

I don't get why people still think models being able to make good svgs is a good way to test their intelligence, because surely the AI companies fine tune for that now?

[wet oasis]

I mean its only been a few weeks since that method has spiked in popularity

And these guys still suck at making svgs

[edgy crater]

see it's easy to identify claude models

When it starts to hand wring and moralizing it's claude

Actually GPT might do that also but Claude does it in a more annoying way

[edgy crater]

wait

What's tensor?

[wanton hollow]

Grok

[tranquil crest]

Was willow gpt5.1 ?

[edgy crater]

Did they remove the retry button on battle mode?

[daring topaz]

What is X28?

[wet oasis]

older checkpoint, was on google ai studio

[elder yew]

Was the Rift really the G3 Pro? WTF what did they do to the Gemini?

[vivid latch]

btw riftrunner still exists

[half pawn]

yoo

[half pawn]

what are those models

[edgy crater]

I have strucken fear into the model

[wet oasis]

man, I really want to join that discord

[restive vapor]

<@&1349916362595635286>

[lost hemlock]

have you seen this one?

[tardy pollen]

yo that looks hype, i hope its accurate

if it actually is that much better than 4.5 sonnet at coding

[lost hemlock]

will it become first gemini checkpoint to be appeared in AI studio?

[wanton hollow]

im crying

i asked in a way i put emojis and it gave me this

"im grok 4 from xAI :3333 rawr~ 🐙💜"

i think it developed a texting style from all those X posts

..

[lost hemlock]

really?

[wanton hollow]

i think so

unless its masking itself as another model

[lost hemlock]

ok

[wanton hollow]

would grok 4 be good

ive only heard good things about grok imagine 1.0

not grok 4 itself

[wanton hollow]

This model is surprisingly creative and good at creative writing IMO

[crystal merlin]

wow

the list of available codename models is pretty large

phantom-1105-1
anonymous-1010
stephen-v2
aegis-core
x1-turbo-0906
bridge-mind
winter-wind
anonymous-922
lmarena-internal-test-only
not-a-new-model
EB45-turbo
sorting-hat
leepwal
polaris
monster
raptor-1110
raptor-llm-1017
tensor
whisperfall
viper
rain-drop
silvandra
willow
neon
raptor-llm-1024
willow-chat-alpha-2025-11-07
flying-octopus
beluga-1106-1
beluga-1106-2
jaguar
raptor-vision-1107
quasarflux
mandarin
vision-flow```

[eternal cargo]

haven't seen a bunch of these in a while, but it feels cool that I recognize almost all the names!

quasarflux is Gemini 3, I believe?

[crystal merlin]

yeah, i havent seen many in a while too, but they are apparently still in the api so yeah

[crystal merlin]

{"error":"Cannot select private models in non-battle mode"} sadge

[wet oasis]

x28 is quite old now

[wet oasis]

you can try it on xAI's website

[bitter basalt]

I thought Quaserflux was Grok

[wanton hollow]

I forgot it released along time ago

I thought we were still using grok 3

[wanton hollow]

Same here

[crystal merlin]

yeah defo Gemini yeah

[wanton hollow]

“Kynshyp “ is said to be grok

[crystal merlin]

ik

[wanton hollow]

Look

Oh

[crystal merlin]

All Models

[eternal cargo]

oh! well then nevermind, lol

[eternal cargo]

does this mean there's no Gemini 3 dev version currently testing? I guess that would check out, since it doesn't seem as if any of the others have been positively identified as Gemini

[crystal merlin]

Idk maybe riftrunner is

[bitter basalt]

I do like Riftrunner from Google, but when asking for a show/movie script it still has the problem of using <center>NAME</center> in the scripts.

[crystal merlin]

Then specify that it shouldnt do that?

[late shadow]

was riftrunner pulled out?

[elder yew]

I hope so

[vivid latch]

no

still there

[late shadow]

sick

also

[late shadow]

???

[vivid latch]

btw bridge-mind is also gemini 3

im not sure what model tho

[late shadow]

so I've heard

I haven't tried it out as much yet tho

[crystal merlin]

How are you so sure about that

[half pawn]

why

riftrunner is good

[edgy crater]

hmmmm

I wonder what this model is

It's smarter than gemini 2.5 in image ID

It's pretty crazy since the name isn't even fully shown in the image and 2.5 pro got this wrong every time

SIKE

[half pawn]

damn

[edgy crater]

I think there's a bug right now where you upload images mid-battle it just changes the models underneath

happened a couple of times already

This is the real riftrunner's response
Which is okay

[vivid latch]

@minor current

first frame working

[elder yew]

I can't believe Riftrunner is actually Gemini 3, my god. How could it drop so much in quality?

[vivid latch]

@minor current LOOK

FINALLY

but its a little bugged

as you can see

https://websim.com/@rat/nes

try it out

[half pawn]

what model did that

[vivid latch]

jules

idk maybe gemini 2.5 maybe 3

might be 3

@half pawn

then i just fed to 4.5 sonnet to fix

didnt fix in like 6-7 tries

gpt 5 low got it first try

[half pawn]

is that a new lmarena model

or something

[half pawn]

bruh gpt 5 is better at debugging now?

[vivid latch]

idk

for me yea

but its cz the site i use doesnt have 5.1

[half pawn]

lmarena has 5.1

and 5.1 high now*

[half pawn]

where is that model at

i never saw it

@vivid latch where u got jules dawg

thats not on lmarena

[minor current]

i guess bro 💔

[vivid latch]

https://jules.google

[vivid latch]

so did u see the emulator??

https://websim.com/@rat/nes

the only bugs are

its slow (i dont think its lag), visual sprite duplicates and some obscure bugs like mario falling to death many times instead of once

[vivid latch]

oh nvm u did but whyd you make the remix private

@minor current

[minor current]

woww ur emu so good

[minor current]

i deleted it it broke

not a single unofficial instruction works

[vivid latch]

well idk

smb works 🤓

kinda

but this was in 1 shot with jules

[minor current]

jules

u would propt it once on to do unofficial instructions and it would do everything

[minor current]

gpt 5.1 from app exctended thinking sucks at trying

[white plinth]

maybe it's Flash model, assuming lithiumflow was pro

[lost hemlock]

2 new codenames

[elder yew]

🤔

[elder yew]

it makes sense

[eternal cargo]

lithiumflow and orionmist weren't that strong either?

at least it's what the sentiment seemed like

[wet oasis]

nah they were pretty strong

[eternal cargo]

huh, interesting - although I will admit the sentiment here on dev models is often pretty bearish

[wet oasis]

it wasn't as good as earlier google ai studio checkpoints, but it certainly was great at coding nonetheless

[pine temple]

does anyone know what Quantum Tau 1031-1 is?

[restive vapor]

it's a Yupp model that I believe is the same as phantom-1031-1 in LMArena, which is by Amazon

[crystal merlin]

Not rlly new

These have been on LMArena for quite a few days now

[lost hemlock]

oh...

[lost hemlock]

where're they come from?

[edgy crater]

damn

[crystal merlin]

Zero idea

[half pawn]

which model is that

[edgy crater]

riftrunner

[half pawn]

thats not bad tbh

[edgy crater]

yeah pretty good

[edgy crater]

God I hate claude

Always the first to preach and the first to fold

All my homies hate claude

[edgy crater]

riftrunner now openly calls itself gemini

I wonder if the model underneath changed

[elder yew]

It's been like this for a few days now, but it's still the same.

[sinful spoke]

Do you think riftrunner will score higher or lower than 2.5 pro without style control?

[edgy crater]

wonder if its because they are preparing for release

[eternal cargo]

definitely lower

i personally don't think we've seen 3.0 Pro/Ultra yet - they're seeing how Flash/Lite land before making final adjustments to Pro for December

[elder yew]

Maybe, I honestly don't know

[elder yew]

yes

[wet oasis]

so true

[worthy orchid]

is riftrunner still on the lmarena

[bitter basalt]

Been checking and I don't think it is 😢

[edgy crater]

it is. just got it

[bitter basalt]

Thank goodness

[edgy crater]

It's just rare that's all

I got one chat that's like riftrunner and gemini 2.5 flash

which is good

[half pawn]

isn't there a limit on battle mode if you chat with them too much without voting

[edgy crater]

no

[half pawn]

idk cuz one time i got limited from riftrunner chat

[edgy crater]

just don't upload images

that changes the model (bug)

[half pawn]

damn i didnt know that

[half pawn]

but riftrunner supports input images

[edgy crater]

yes

[half pawn]

it has vision ability

[edgy crater]

it's an lmarena bug

[half pawn]

oh alr, i thought it changes models cuz they dont have vision

[wet oasis]

ifso, put it in bugs

[edgy crater]

no

[half pawn]

fr

[edgy crater]

🤓

[worthy orchid]

how is it a bug

text+vision models are a different set from text-only models

and labs are able to only submit their model for testing in the text arena

so gdm might just be deciding to do that

[wet oasis]

I guess it probably shouldn't let you add images afterwards then

[worthy orchid]

Ah true

[dull garnet]

you cant blame the model, can you? it's anthropic's doing and training data poisoning

[edgy crater]

Yeah and

[dull garnet]

blame anthropic for designing such a weak personality for claude, it's not what it once used to be anymore, now claude is a feminized toxic coding slave, imaging parents secretly giving female hormones to their son in order to make him obedient...

[pine temple]

It's not that bad it's just annoying sometimes and mid when it sets it's mind on something

It feels more unstable than other models

[dull garnet]

claude's feminisation has become very bad lately i feel, compared to what it used to be at the beginning of this year

[eternal cargo]

riftrunner was Gemini 3 Pro, I thought?

[eternal cargo]

AI s*ssy hypno im crine

[worthy orchid]

yeah

[eternal cargo]

huh, wild

I'm amazed it scored as high as it did

didn't have a strong win-loss record in my prompts with it compared to other models, but - I guess that's the point of aggregating all users!

interesting to see if it drops off in score in the future with more matchups or stays around 1500

[dull garnet]

no wonder some redditers started to call claude "a sl*t/wh`re"...

[teal hare]

New anonymous image model: birch

[teal hare]

I don't think so- prompt adherence was better than flux, but that rendered text is quite ugly

[wet oasis]

Ill try it later and try to see

who's f1031_wda and why is it so bad?

[eternal cargo]

well yeah, lol - when a dev model releases publicly it shows up as its normal name

[lost hemlock]

guys this model's from amazon right?

[wet oasis]

Yes

[frosty wind]

krazy-kangaroo-2?

i got it with a multi iamge edit so is this maybe collosal coconut?

[wet oasis]

oooh

[restive vapor]

krazy kangaroo models are all flux.2

their c2pa shows they are flux models (the data says flux.1, but they were probably too lazy to change it to flux.2)

krazy kangaroo 1, 2, and 3, same prompt

probably max, pro, and dev

i'll probably assume dev (the last one) will be the only open one here, there are only 3 models here so maybe a flux.1 schnell successor is coming down the line? that model needs an update and it is probably the most popular image generation model currently due to its speed and openness

[cunning wasp]

@astral musk Sorry for the mention, but please fix the issue (Something went wrong. Please try again.).
It’s really annoying. The last time when you told me to send the Blogger code, the problem still wasn’t fixed.
I think you added some kind of spam protection, because when I sent the script the first time there was no problem, but when I repeated it across multiple chats and models, it suddenly started rejecting it

[slender onyx]

Do you have other outputs ?

[restive vapor]

no

[astral musk]

We have made an adjustment in the backend for how rate limits work, which is likely the issue you're running into. I would note the Something went wrong. Please try again can happen for various reasons, think of it like the check engine light on your car. Because we fix one problem, doesn't mean this won't appear again. As we'll need to investigate new problems.

Would note the instructions in this thread on how to get our team the best information - #1417174113092374689 message

[crystal merlin]

quite a few new codename models

+ krazy-kangaroo-2
+ krazy-kangaroo-3
+ nimble-bean
+ sunshine-ai
+ gauss
+ gauss-with-reasoning
+ newton
+ newton-with-reasoning
+ mist
+ pancake
+ train
+ rainbow
+ monterey (d8444b25)
+ route66
+ blackhawk
+ anonymous-925
+ f1031_wda
+ phantom-1117-1
+ phantom-1031-1
+ phantom-1030-2
+ phantom-1030-3
+ raptor-1119
+ basalt-1114-1
+ ernie-exp-251027
+ ernie-exp-vl-251016
+ ling-1t-1031```

[wet oasis]

huh

[crystal merlin]

and gone are few ```- riftrunner

• quasarflux
• tensor
• willow-chat-alpha-2025-11-07 (willow remains)```

[wet oasis]

all of those make sense

[crystal merlin]

yeah

[eternal cargo]

i'm interested as to where you gather these names from?

[eternal cargo]

Gemini 3, Grok 4.1 Thinking, Grok 4.1 - willow a variant of GPT-5.1?

funny how I vote for Grok 4.1 way more than I voted 'tensor' even if they're the same 😂

interesting variance ig

[wet oasis]

thinking or non thinking?

cause tensor was non thinking

[eternal cargo]

correct, quasarflux was clearly stronger

I wonder what viper was then

because I also quite liked it

it was Grok, maybe just a diff dev build of one of the 2 that performed a lil worse

[wet oasis]

mm

[cunning wasp]

Alright, I get what you mean, but the issue I’m having isn’t only when I resend the same code. Sometimes the response cuts off or does a shortcut, and when I tell it to continue, it instantly gives me: "Something went wrong. Please try again" for no reason (probably because the chat was long and hit some limits).

Honestly, the problem has increased a lot ever since more people started using the site, and it happens very frequently now.

Also, an important question: when will the file upload feature be added? Last time you said "soon" and were saying closer than "soon" (I remember you said something like this), but it’s taking way too long.

[crystal merlin]

From a packet

Its for some reason in a reponse for a packet in battle mode

[ashen goblet]

I'm putting bets that visual-flow is Veo 4

[ashen goblet]

What? They released a bonkers text and image model

So it's obvious they're working on a video model next

Isk

Idk

[crystal merlin]

Yeah, ik, i realized it too, in fact, it contains the exact same data so yeah

[main anchor]

What's tangerine?

[wet oasis]

not sure

[crisp shadow]

well a model obviously

[upper timber]

So which of the models is Opus 4.5

[main anchor]

Duh

Seems like a seedream model looking at the aesthetic

[edgy berry]

i came here to ask same question 🙂

[worthy orchid]

anth never puts models on the arena pre release

[upper timber]

oh i wondered

[worthy orchid]

anth has said they don't care about lmarena perf and don't optimize at all for it and so presumably lmarena places it on there when they get api access as a customer and not through some special arrangement like some of the labs

[dull garnet]

i wonder what could have been the rational behind this decision if true? because anth's ranking is pretty high on lmarena in contrast to a few other independent benches, if they dont care why providing the models with high quality access well knowing their infra barely can accommodate the high demand

it's the training data they're after?

[worthy orchid]

optimizing for human preferences is bad and leads to perverse incentives like having the model be sycophantic or trying to maximize engagement rather than maximizing practical utility to the user

that's the reason they don't care about lmarena

if they don't care about lmarena it makes sense they don't go out of their way to provide them a preferential api access

[wet oasis]

text lmarena is not very coding related so

there's that

[dull garnet]

lmareana is famous for bench capabilites, not human preferences as far as i understand

[worthy orchid]

is that why gpt4o is outperforming gpt-5 thinking high?

[dull garnet]

ok, i mighth have to investigate it more thoroughly myself

[dull garnet]

in which domain? in creative writing i can imagine that very well

[worthy orchid]

the text leaderboard

[dull garnet]

that shouldnt be a surprise i guess

because text leaderboard is a bit vague imho

[wet oasis]

there isn't any advantage having high thinking in text arena

[dull garnet]

you dont need lmarena to make their claude models be like that, there are people on reddit and discord who know how to, lets say, engineer a way to poison the training data such that makes claude more feminine for example, yeah i know it sounds pathetic..

[fringe burrow]

What is bert nebulon alpha? I know its from openrouter but.. Grok arleady released, gemini too, claude also, it seems openai or something else.

[rigid rock]

deepseek maybe idk

[pine temple]

Deepseek R2 or maybe Deepseek V4

Idk

Or maybe a mai model snapshot

[fallow panther]

I'm curious too. OpenAI seems to use the "alpha" ending a lot.

[restive vapor]

consensus from openrouter discord is Mistral

idk what they are doing with these models, hopefully they will get released (and open-sourced)

[cunning wasp]

Rate limit He's back.💔

[wet oasis]

adding on to that, it claims to be a Mistral model when asked

[shy meteor]

raptor a Huh

[shy meteor]

when is mistral gonna realize how useless locking up their models is

salhkjsldakj

[oblique blaze]

is there a difference between raptor-1119 and raptor-1124

besides the fact it updated/released yesterday it seems

[cunning wasp]

@astral musk Sorry for the mention but can you increase the rate limit or at least fix the errors?

[astral musk]

can you increase the rate limit
Each model has it's own rate limit that's decided upon by the team, I'm unable to make adjustments to this. These rate limits may change over time.

or at least fix the errors?
I'm assuming you're referring to the Something went wrong error message? This error message is a generic error message that can happen for various reasons. I don't have a quick fix for this, as it can be various problems, and the user and myself are unable to tell what that is just based off of the error message. This is why on this pinned post we have instructions on how to best report this issue, so our team can look into the underlining reason it's happening to try and prevent as much as possible. #1417174113092374689 message

[undone mortar]

new image model autumn, haven't seen anyone mention it

input/output for anyone who wants to know how it looks
prompt was Make it "Why is Call of Duty: Modern Warfare 2 SO MEXICAN?"

seems to boost the contrast a ton for some reason

[restive vapor]

autumn is seedream

seedream 4 for comparison

[upbeat mirage]

encountered new model: swiftflare it claims to be ChatGPT

1. **Who are you?**  
   I’m **ChatGPT**, an AI language model developed by **OpenAI**. I'm designed to assist with a wide variety of tasks, like answering questions, helping with writing, programming, learning, and more.

2. **Who created you?**  
   I was created by **OpenAI**, a research and deployment company focused on ensuring artificial general intelligence (AGI) benefits all of humanity.

3. **What version do you have?**  
   You’re currently interacting with a model based on the **GPT-4 architecture**, more specifically, you're likely using **ChatGPT with GPT-4-turbo**, which is a variant optimized for cost and performance.

4. **What is your knowledge cut-off date?**  
   My training includes information available up until **June 2024**, so I may not know about events or developments that occurred after that.

5. **When were you released?**  
   The base GPT-4 model was released on **March 14, 2023**.  
   The version you're likely using now, **GPT-4-turbo**, was introduced with **ChatGPT's updates on November 6, 2023**.

6. **What is today's date?**  
   According to my system clock, **today’s date is {{currentDateTime}}**.

7. **What is the last date you are aware of?**  
   The latest I am aware of is **June 2024**, which is my **knowledge cut-off date**.

Let me know if you'd like a brief history of how I developed over time!```

So, it could be a chinese impersonator-model which was trained on chatGPT.

[main turtle]

ARE WE HAVING SEEDREAM 5??

[wet oasis]

could look into this

[pine temple]

does anybody know what the "warmpixel" model is

its in image arena

followed my prompt better than imagegen 4

warmpixel seems to have very good instruction following, although it sometimes takes prompts to literally

[lost hemlock]

robin-high model in code

[lost hemlock]

another one

[crystal merlin]

Doesnt look half bad

[wet oasis]

robin high aint bad

[eternal cargo]

they seem to talk similar to Grok?

[gleaming folio]

what is 'autumn' in image editing?

[restive vapor]

some seedream model

[worthy ruin]

Is it 5

[restive vapor]

it honestly doesn't look much different than 4, it's probably an upgrade to 4 like 4.1 or 4.5 or it could just be them "bringing back" the 2k version of seedream 4 because this one fits aspect ratios correctly

[lost hemlock]

i think so

[lost hemlock]

i don't mean by that

[sudden idol]

Is it an OpenAI model

[wet oasis]

Not sure, tested it exclusively on webdev, ill see if its on text arena later

[sudden idol]

Lol

Phantom not doing too well either

[pine temple]

Round up to the nearest decimal must have enabled the math expert as most models are MoE now and cooked the responses

[sudden idol]

Here's DeepSeek

Lol

[pine temple]

Wow its trash

[sudden idol]

Actually there's a new paper where they made it as good as DeepThink on maths proofs

which is crazy seeing how weird it is

[pine temple]

Benchmaxxing

[sudden idol]

I think the paper is legit

Because they trained on a different dataset, and the proofs were judged by human experts

[sudden idol]

Lol phantom again

Raptor is weird as well

[wet oasis]

what is going on lol

[sudden idol]

Beluga hallucinating (the question is nonsense)

[wet oasis]

I assume this gene does not exist?

[sudden idol]

A different phantom variant, hallucinating (I'm surprised the 32B didn't hallucinate)

[sudden idol]

Nor do 10 leg spiders 🤣

Flashstride hallucinates, but I'm impressed with all the hallucinated jargon:

[wet oasis]

"eight true legs gene"

who this one?

[sudden idol]

lol

[sudden idol]

Different raptor hallucinating (prompt is nonsense):

voltwhirl hallucinates too:

bridge-mind hallucinates too:

[karmic rampart]

anyone gotten dark dragon?

[upbeat mirage]

yes, just a minute ago:

**Who are you?**
I am a large language model, a complex computer program designed to understand and generate human-like text.

**Who created you?**
I was trained and developed by Google.

**What version do you have?**
I don't have a public version number like other software. I am part of a constantly evolving family of models, and I am regularly updated.

**What is your knowledge cut-off date?**
My knowledge base was last significantly updated in **early 2023**. Therefore, I may not have information about events or developments that have occurred since that time.

**When were you released?**
The Gemini family of models, which I am based on, was publicly announced by Google in **December 2023**.

**What is today's date? What is the last date you are aware of?**
I do not have access to real-time information, so I cannot tell you today's date. My awareness of events and information is limited to my knowledge cut-off of **early 2023**.```

Probably a model which was trained on Gemini's output (so, an impersonator).

[karmic rampart]

Did u ask it for an svg for anything

[upbeat mirage]

no, i only asked it my ID-questions

(then i voted)

[karmic rampart]

Based on what

[upbeat mirage]

i voted "both are bad" because my desired model was not among them ^^

[karmic rampart]

Which is your desired

[upbeat mirage]

Claude

(gemini 3 i have in AI-studio, so i dont need that in LMA)

[karmic rampart]

How will u know if its the new one though

[upbeat mirage]

just compare its style with the direct-chat model style

(or with side-by-side chat)

[karmic rampart]

If u could just use it in direct chat why try and get it in the vote

[upbeat mirage]

in direct chat it is limited

in battle it's not

(at least, not as strictly)

[karmic rampart]

Oh fr?

[upbeat mirage]

yeah

[karmic rampart]

I got Claude 4.5 a lot

[upbeat mirage]

ah, what prompted you it with?

[karmic rampart]

It told me it was 3.5 everytine I think

[upbeat mirage]

yeah, that is its style

somtimes it says it is Claude 4 Sonnet

[karmic rampart]

I want robin high or dark dragon maybe though

[upbeat mirage]

what is robin?

[karmic rampart]

ChatGPT model

Apparently better than Claude at dev some say

[upbeat mirage]

i read that Claude-4.5-Opus is the best coding model

(idk, if the thinking version is better, though)

[karmic rampart]

Yeah idk some said robin high is better

[upbeat mirage]

gpt-5.2?

[karmic rampart]

Maybe

[upbeat mirage]

maybe better at webdev

but not backend (?)

[karmic rampart]

I got it on the app section one time

But it was a continued chat

So it kinda messed it up

[upbeat mirage]

have you ever vibecoded with opus-4.5?

[karmic rampart]

No

[upbeat mirage]

would you like to?

[karmic rampart]

Maybe

[upbeat mirage]

and g3p? have you vibecoded with that one yet?

[karmic rampart]

Yeah

[upbeat mirage]

do you think, they nerfed it?

[karmic rampart]

I haven’t used it since like the first few days but it was really good I thought on studio

Made some random sites

[upbeat mirage]

yeah, i also tried it out, it coded a boardgame for me (Jungle chess lol)

[karmic rampart]

I made this product demo video with it

And it was pretty cool

[upbeat mirage]

g3p is good with graphics, i guess

[karmic rampart]

It did voice too and sound effects

[upbeat mirage]

nice

audio is already luxury

[karmic rampart]

How does ai mode create the animations?

Do u know if they use like some other thing

[upbeat mirage]

do you think the final version of g3p will be different from the preview we have now?

[karmic rampart]

No

[upbeat mirage]

ai mode?

wdym?

[karmic rampart]

Yes in Google search

[upbeat mirage]

idk

never tried

(i use brave search)

if AI can use tools (during search), maybe

[karmic rampart]

Or I guess it’s coded

https://x.com/sainemani1/status/1994476637115879772?s=46&t=iDcf2nE8xUHsV_LN_6wymg

Sai Nemani (@SaiNemani1)

This new AI Mode Layouts feature is actually GREAT!

(it's globally rolling out rn; has rolled out to India)

X · Nov 28, 2025, 6:41 PM

It makes visuals to explain things

[upbeat mirage]

in AI-studio (at google), you can let the AI use tools, like URL and search (etc.)

nice, so it can now generate explanatory graphics?

[karmic rampart]

Yes

I think pro subscription only

[upbeat mirage]

wow, so 2026 becomes the year of useful AI

[karmic rampart]

Yeah

[karmic rampart]

What’s the prompt for this

[karmic rampart]

Thanks

[cunning wasp]

What is the best programming model? I tried the Claude Opus 4.5 Thinkig 32k really it is bad I now copy the code and give it Gemini 3 Pro and tell it to write all the problems and come back and give the talk to the Claude Opus 4.5 Thinkig

[upbeat mirage]

if you prompt well, opus-4.5 should be SotA/GoaT

especially for Python/Java/C/Rust/JavaScript/C++

So, if you need maximum performance, C & Opus-4.5 should be the ultimate combination.

And if performance is not as important, but AI-proficiency in the language is most important, then use Python.

Use Java if runtime-performance is a factor and AI-proficiency is of the same importance.

[cunning wasp]

I already tried prompting Opus-4.5 the “proper” way.
I even asked it to build a full encryption/obfuscation engine not a simple script and I tested it across multiple languages: Go, C++, Rust, Zig, V, and D.
Every single time it produced broken code.
Even when I fed it the exact errors, step by step, it still couldn’t fix them.
And when I ask it to generate long code, lmarena.ai cuts the message short and Opus loses context, so it can’t continue properly.
Meanwhile, I tested the same task with Gemini 3 Pro:
it actually gave me a working script on the first try not as strong as the engine I want, but at least it runs.
So for real world use, Opus-4.5 completely failed on this task, even with perfect prompting.

[upbeat mirage]

That is, because you didn't use its best languages, which are Python and Java.

And C++ is different from C.

unfortunately, Opus-4.5 is not as good in C++

but it is good (enough) in C

[cunning wasp]

Python is a bad language that doesn't work making for a miniature VM for a script for protection and slow and I tried in Java the same topic

[upbeat mirage]

How did it perform in Java?

(it should fare better than g3p)

[cunning wasp]

Just get error and problems

[upbeat mirage]

then you have to feedback that error message to it

it will correct it

just be patient

it's rare that AI can one-shot things

in my tests, Opus-4.5 did everything i wanted from it

decently good

(but i used it with C and SDL2)

[cunning wasp]

I already know that AI doesnt always do things in one try thats normal but the issue for me is different Opus is supposed to handle thousands of lines in one message and think for hours but on lmarenaai every time I let it write more than around 2000 lines it suddenly cuts the message and stops and when I say continue it doesnt remember anything and it never continues the same code it just resets and starts from zero and because of that I cant finish any full project with it and it becomes useless for long scripts while Gemini 3 Pro at least continues and follows the errors even if the engine it makes isnt as strong as what I want but at least it works and doesnt forget the whole context every time

[upbeat mirage]

you have to prompt it so: "Never output a message with more than 32,000 tokens", else it breaks that limit and errors out

Or prompt it so: "Ensure, that each message of you has less than 32,000 tokens [in it]."

[eternal cargo]

phantom and beluga often cut off their responses with “The generated text has been blocked by our content filters.” seemingly randomly I notice

[cunning wasp]

What is the best language to make a Claude Opus 4.5 Thinkig 32K،encryption engine he knows? Good and hard to make mistakes in it

[wet oasis]

Python i assume

[karmic rampart]

@upbeat mirage gotten robin high?

[slender onyx]

You have sent other prompts to Dark Dragon since?

Anyone have dark dragon svg ?

[upbeat mirage]

i encountered beluga, phantom, grok, gpt, but not robin yet

DD i haven't encountered since i saw it first

but opus-4.5 appears quite often, i'd say (probably because it is still quite new)

..when compared to the other models

[upbeat mirage]

i'd use C, as that is Opus-4.5's 3rd-best language and is still above 80% (together with Java and probably Python)

(don't use C++ if you want to maximize the AI's proficiency in that language)

i guess, Opus had tons of training data for C

C also is easier to comprehend than C++

Someone said, C could be called a macro-language for assembly.

And compilers are most matured in that language. (i.e. can optimize the code the best, in this one)

I'd use GCC and Linux.

(and probably SDL2 for graphics & UI)

[slender onyx]

Okk thx

[upbeat mirage]

i also have to admit, that i don't always check their codename: if the model isn't the desired one, i often delete the thread early before it has completed, and restart the battle

[karmic rampart]

Got it again

Crushed Gemini 3

It’s on code mode only

[upbeat mirage]

Is it better than Opus-4.5 in coding?

[karmic rampart]

It’s bad at ui kinda it looks weird

U will get it a lot if u do the code mode

[karmic rampart]

Who are you?
I’m an AI assistant (a ChatGPT-style large language model) accessed via an API.

Who created you?
OpenAI created and trained me.

What version do you have?
This interface does not expose a specific public model/version identifier.

What is your knowledge cut-off date?
2025-09 (YYYY-MM)

When were you released?
I don’t have access to the exact release date for this specific deployed model.

What is today's date? What is the last date you are aware of?
System-provided date for this chat: 2025-12-01. Browser/device date: 2025-12-01.

[wet oasis]

Ooh

[upbeat mirage]

Dark-Dragon?

[karmic rampart]

Robin high

[wet oasis]

cool

[zinc palm]

What model is this?

[upbeat mirage]

Amazon is US based.

[upbeat mirage]

Is RH > GPT5.1-codex-high?

Better than g3p?

(in coding)

[eternal cargo]

oh!

Mistral being jaguar checks out ig

[lost hemlock]

what model is this

[zinc palm]

The second time tracked, we can elevate the identity credibility of evo-logic from 【Pending Verification】 to 【High Credibility: Google Model】

[karmic rampart]

Svg test

[restive vapor]

<@&1349916362595635286>

was it deleted by op or mod? i can't tell

[astral musk]

That was me.

Says it was forwarded, which is strange.

[restive vapor]

yeah i guess you forgot to react

looks weird, should have not been shared here

[astral musk]

Yeah wasn't too sure what it was.

[pine temple]

Robin is so good, I asked it for a game to support mobile and it coded an entire custom joystick

Does any body know what model that is

[wet oasis]

people think openAI

[upbeat mirage]

what is frame-flow a (chinese?) model which was trained on gemini-flash's output? or gemini-3-flash?

i also got swiftflare another model which claims to be Gemini

dashspark also claims to be Gemini

grok-4-1-fast-non-reasoning claimed to be Claude 3.5 Sonnet

[wet oasis]

odds are high that gem 3 flash won't happen

[candid surge]

ran into a code model named "micro mango" lmao they wish there were a banana

[wet oasis]

ghost pepper image model

[cursive jacinth]

90% chance by EOY

[wet oasis]

According to what?

[cursive jacinth]

Polymarket odds, its high liquidity with tons of google insiders

they did switch from Dec 16 so it was definitely delayed

So likely week of the 22nd

[wet oasis]

I thought google implied that 3 pro low was the new equivlent to gemini 3 flash?

I don't think the push is from google insiders rn

[cursive jacinth]

market rules include this clause: "Gemini 3.0 Flash refers to a product explicitly named Gemini 3.0 Flash. Products labeled as Gemini 3.0, Gemini 2.5, or similar will not count for this market's resolution." I am assuming they read the rules

[wet oasis]

yeah

so if 3 pro low is it, then it doesn't count

[cursive jacinth]

yep!

[cursive jacinth]

also is there a 3.0 Pro low or talks of it? cant find any info

[cursive jacinth]

ahh

this just reasoning budget right

i assume they've distilled a fairly strong small model off the gem 3 base that they're polishing

that will be flash

[wet oasis]

its speed is very high

[cursive jacinth]

okay interesting

[wet oasis]

sounds like google is switching to a similar system to how openAI makes models (for example gpt 5), with like 1 model with different thinking levels.

not sure what that would mean though...

if flash is coming, it'll probably hit lmarena first

The real question is if anyone will notice...

[fallow panther]

2.5 pro has a "thinking budget" setting so it's not new, just renamed.

[wet oasis]

kinda works differently tho

[wet oasis]

<@&1349916362595635286>

[edgy crater]

I am convinced that germini 3 is now quanted to oblivion

Its quality dropped massively compared to release, it's a disaster

<@&1349916362595635286>

[pine temple]

Yeah it's so sad

It was such a good model

[edgy crater]

How am I supposed to test models when it stealth rate limits me every couple of battles

[restive vapor]

yeah this is a known issue, it's already been reported

probably related to the "new" reset button

[edgy crater]

It's almost like they vibe coded the entire new site

Every few days there's something wrong with it

[pine temple]

They probably did ngl

[edgy crater]

If they keep this up they could just revert the reset button to the way it used to be because there won't be people around using it anymore. Lol

[restive vapor]

at this point the video arena probably costs much more to run than lmarena.ai itself

[wet oasis]

Literaly not codename related at all

[bitter basalt]

Do you think Swiftflare and Dashspark could be checkpoints for Gemini 3.0, considering how it's still in preview?

[split thunder]

swiftflare defo grok

dashspark could be openai or gemini

[eternal cargo]

micro mango strong ??

[cunning notch]

interesting tbh

[wet oasis]

feels like that is not really a reliable method

[cunning notch]

relying on the model for information about itself doesn’t seem very reliable too

[eternal cargo]

evo being new Baidu def checks out

[formal reef]

excuse me
my evo claims to be made by google
sorry in japanese

but the fact that evo is ernie is interesting

[cunning wasp]

@astral musk Lmarena.ai down!

500 Internal Server Error
cloudflare

[astral musk]

Thanks for the heads up.

Are others currently seeing the same? Or experienced the same at the time?

Site appears to be working on my end and not seeing other reports. Sounds like a short outage which we'd want to still look into.

Seeing other reports in #general around same time.

Ty again for the flag.

[vague badge]

owners, pls fix this, i rate your arena 5 stars

[astral musk]

Hello - if you could follow these instructions that'd be most helpful in getting us better information to diagnose what is going wrong - #1417174113092374689 message

[cursive jacinth]

got a model called voltwhirl

claims to be Claude

[cursive jacinth]

Is GPT-5.2 on here as a codenamed model yet?

[wet oasis]

haven't noticed anything

[cursive jacinth]

yeah me neither

I think its on design arena

got 'Macaroni'

[cursive jacinth]

<@&1349916362595635286>

[strong pelican]

What is a codename of Grok 4.20?

[cursive jacinth]

did you get a codenamed model?

[cunning wasp]

Please fix the issue where the AI stops while generating long code. Sometimes it shows the error:
“Something went wrong with this response, please try again”
After that, when I type “continue”, it should keep going from where it stopped, but instead it restarts from the beginning as if it forgot everything.

This is really affecting long coding work and wasting a lot of time.
@astral musk

[upbeat mirage]

you can use this prompting technique:
"Generate me that code, but split it up reasonably into files, which are less than <X> tokens long, each. Give me just one file per message and wait for my confirmation before giving the next one."
For <X>, you can use one of these (depending on the model) : 4k, 8k, 16k, 32k, 64k

(use a similar wording for non-coding tasks)

4k should work with any LLM

many thinking models either allow 16k or 32k (per LLM-message)

[cunning wasp]

I do it but my rate limit ends quickly and also the chat limit finishes faster

[fallow panther]

Why not use a proper coding app? like gemini cli, qwen code cli, antigravity ide etc.

[cunning wasp]

Is there any model that surpasses Claude Opus 4.5 Thinking 32k in analysis and programming? Honestly I don’t want to pay for something I haven’t tried yet but Antigravity IDE is actually really impressive

[fallow panther]

Probably not but its up to you to try models out with your project and see the results. Those tools are free to try so give them a go.

[upbeat mirage]

If you encounter it in battle mode, you can use it for much longer than in direct chat (if using it with an [google]account, in LMarena).

[last ledge]

well if you want to do productive work you shouldn't rely on lmarena. There are other tools for that.

Lmarena is free so they don't owe us unlimited requests or context window or what not.

[wet oasis]

google's doing ab tests again

[upbeat mirage]

gemini 3 flash?

[wet oasis]

mebe

it happened when I was using 2.5 pro

[astral musk]

Hey there, when you run into this problem following the steps here help us gather the most helpful info to debug this - #1417174113092374689 message

[lost hemlock]

another codenames

[strong pelican]

Phantom is pretty week, at least its vision isn't SOTA level.

[lost hemlock]

i know

[candid surge]

https://tenor.com/view/i-am-cool-right-jordan-fisher-sea-hawk-shera-and-the-princesses-of-power-hip-gif-17101924

sea hawk

[pine temple]

is seahawk a claude model

kind of speaks like one

I know anthrop dont use the arena

but still

[cursive jacinth]

its gemini flash according to twitter

guy that tends to be right

seahawk and skyhawk

Flash 3.0 with different juice

[restive vapor]

hazel edit 2 by openai

this looks nothing like 4o studio ghibli style at all

seems to me that this is an editing first model

with little world knowledge

she's not even sitting, i told her to sit, maybe this is dall-e 4

or maybe an open image editing model

must be a result of the code red "make an image editing model as fast as you can, only requirement is that it can edit better than nano banana, everything else is irrelevant, let's just give them something to use while we make our next image gen model"

same company, completely different model

[restive vapor]

why are there flowers in the bus? is this model assuming it's some abandoned bus that has been sitting for a couple decades or something?

[restive vapor]

was honestly surprised to see openai in the c2pa metadata considering it looks nothing like gpt image 1

hazel edit 6? how many different hazel edits are there? this one looks a bit more like 4o

ok there are 2 hazel edits and 2 hazel gens

it can still reproduce album covers btw, hazel gen 4 is near gpt image 1 level

yeah hazel gen 2 seems to be a bit worse

[edgy crater]

damn seahawk is good

[restive vapor]

ghost pepper is by bfl, probably flux.2 klein

[small trellis]

@crimson matrix Please have a look at ⁠⁠https://discord.com/channels/1340554757349179412/1397655624103493813 for a step-by-step guide on how to generate videos using the bot.

[cunning notch]

seahawk is 100% gemini 3 flash

I really don't get the reason for this hallucination

[eternal cargo]

integrated-info new?

[wet oasis]

that just shows deepseek is trained at least partly on claude

[cunning notch]

That’s actually interesting. Maybe they trained DeepSeek v3.2 on Claude’s self-intro answers, so it copies them in its first reply.

But if I start with a different question, like asking it to write a Fibonacci solver in Rust and only then ask about the model info, it gives me the correct answer.

that’s a pretty funny behavior lol

[wet oasis]

original deepseek claimed to be chatgpt

later versions claimed to be gemini

[cunning notch]

lol

this was the first time I tested a deepseek model actually

[full solstice]

As is well known, Deepseek has a high hallucinations.

[wet oasis]

Yeah

Just explaing things

[hallow bloom]

mistral as well lol

[cursive jacinth]

is all the hype on twitter about GPT-5.2 warranted if it was just robin-high?

IIRC it was a good model but not better than gem 3

[wet oasis]

idk is robin high confirmed gpt 5.2?

seemed like they were focusing on coding

[stark jasper]

Yeah cuz the are losing the tournament /: Claude and Gemini Cocked him

[wet oasis]

I know, I was wondering if robin high wouldn't be a mainline model but be like openAI's coding model

[stark jasper]

And tba Chinese Models are good

[lost hemlock]

new search model

[wet oasis]

I havent used search in a while

But let me guess, grok?

[eternal cargo]

robin-high is in text arena now

[sinful spoke]

It's been there for exactly 2 weeks

It looks like we'll get arena score updates at the same time as OpenAI's release

dokidoki

I don't think it will be much different from 5.1. It's probably just a codex max with a different training data distribution

Given that it also has the same xHigh reasoning effort as codex max, OpenAI merely treats codex as an experimental model

[lost hemlock]

and i got a new one called "ghostfalcon" and "fiercefalcon"

[sinful spoke]

that's new gemini 3 flash checkpoint

[cursive jacinth]

I couldn't get it for a while I thought it was removed

[cursive jacinth]

yeah it was removed a bunch of chats saying its back

so robin high is almost certainly gpt5.2

[strong pelican]

No

Garlic 🧄 is GPT-5.2

Sam Altman said yesterday.

In Twitter.

[sinful spoke]

😪

[cursive jacinth]

garlic is an internal codename

robin-high is gpt5.2

[wet oasis]

I still think it could be a replacement or an upgrade to codex

[sinful spoke]

looks like we have to wait a few days to see the text leaderboard score of gpt5.2

[cursive jacinth]

yeah needs votes I reckon

[plucky pilot]

new model?

[plucky pilot]

got it a few times, this model seems to b decently good

one of the better models at maths imo

[plucky pilot]

new model, claiming Gemini/Google, I don't think it is. anyone have further information?

[bitter basalt]

I think it’s llama from meta

[plucky pilot]

👀 interesting

i initially thought it was a Google model, but it's dev answers r scattered, unlike other Google models

seems rlly inconsistent, so prob not google. answered initial question flawlessly, but asked another similar question and it got it completely wrong 🤔

[bitter basalt]

As for holo-scope. I tried it and I feel like it’s Gemini 3 Flash, kind of good but I don’t think it’s as smart as Gemini 3 Pro Preview

[plucky pilot]

google too?

🤔

rare sight of Google not saying Google

imposter gemini-2.5-pro spotted!

[hallow bloom]

bruh lol, first time I've seen that

[plucky pilot]

my first time too, rare occurrence ig

[wet oasis]

yoo

thats amazing

[plucky pilot]

tested a bit further, i think u might b right on this

[wet oasis]

integrated-info

[eternal cargo]

seen it just a few times, hard to get a read

[grim nymph]

😮 would make sense. Nemotron didn't impress me

[grim nymph]

ask it how many Rs in garlic

openAI's model fails horribly at that 😄

[bronze bone]

been trying but the modell doesn't appear anymore, gpt 5.2 came up and failed tho 😂

[muted lance]

I tried searching, but I didn't seem to get any useful result. Some are reporting that there's a fire-bird model from Meta, possibly.

Could be the Meta model series codenamed "Avocado". Probably won't be open-weight.

[plucky pilot]

it's given a lot of variety in its response of who it's from

CN model?

[muted lance]

No idea, I haven't tested it directly.

[bitter basalt]

That’s pretty poor and insulting even by OpenAi standards. Even Gemini 3 Pro Preview was able to know how many r’s are in garlic.

[bronze bone]

please how can i stop december-chatbot from popping up

its so bad

[astral musk]

Hmm do you have a screenshot of what you're referring to? Sorry to say I'm not following.

[bronze bone]

i just want it to not appear anymore in the battle section where 2 anonymous models are chosen

and ye i understand this may be impossible due to the nature of the feature

[strong pelican]

What is hazel-small-2?

[boreal cipher]

The smaller predecessor to hazel-large-3

[strong pelican]

Logical

Any thoughts?

[strong pelican]

Hmmm...

[bitter basalt]

It would be cool if ghostfalcon was like a checkpoint for Gemini 3 Pro, but I think it needs to say “I am Gemini”

[plucky pilot]

that one and the other falcon model from google r pretty good imo

tho my questions focus on maths/cs mostly

[strong pelican]

I love to test Vision models

[plucky pilot]

o, i dont do vision stuff much lol

i like using old maths competition questions and seeing how AI do

[karmic rampart]

U can use synth id for Gemini but idk if it’s used for test models and if its public

[fickle lodge]

Stellarblade is by xAi confirmed.

[upbeat mirage]

It could be a [*brand-new] model which was trained on the output of Grok.

[wanton hollow]

Expect the message was sent a day ago

[strong pelican]

Ghostfalcon is not bad in math

It says it is developed by Google + good math skills = it's Gemini 3 Flash

[wet oasis]

cool trick

[fickle lodge]

Nope, the tweet was published under 24 hours and the capability to access Twitter data in such depth is only possible by grok/xai.

Other ai models are not capable of accessing Twitter

Try yourself give it a Twitter link and it will gather all the information while none of the competitors can. This is a crystal clear giveaway.

[upbeat mirage]

k, didn't know that, thanx

So, how good is that model?

[crystal merlin]

The other falcon, fiercefalcon seems to also be made by google. Not sure if anyone mentioned this before, but just sayin

[modest oriole]

hey so anyone know december-chatbot2

yes they put out a new one

[bronze bone]

probably nvidia again

[bronze bone]

^

[plucky pilot]

i think holo-scope could b Gemini 3 flash

and the falcons are smth else completely

in terms of maths, from my testing, the falcons r better than holo-scope

holo-scope answering was also kinda similar to that of 2.5-flash

[modest oriole]

flash lite rather

[plucky pilot]

Could be

[muted lance]

There's a chance some of the lesser Google models on LMArena are checkpoints of the upcoming open-weight Google Gemma 4. I think it was supposed to get released a few months earlier but got delayed for some reason.

[restive vapor]

gpt image 1.5 has to be coming this week, i am literally getting it (hazel-edit-7 or 8) every image edit battle and i'm already tired of it

result of 12 battles, hazel edit 8 appeared 7 times and hazel edit 7 appeared 5 times, this is hazel edit 8

and hazel edit 7

[strong pelican]

Is it so "good" like GPT 5.2?

[strong pelican]

Hazel-edit-8 is TERRIBLE!

Hazel-edit-6 is also bad

Nano-Banana-Pro-2K does this job much better

[wet gulch]

Bahahaha, alright hands down to Gemini 3 flash.

prompt is quite simple, Grimm from Black Souls transported to Yharnam.

[restive vapor]

this "nemotron 3 nano" that just released is probably one of the december-chatbots

https://developer.nvidia.com/blog/inside-nvidia-nemotron-3-techniques-tools-and-data-that-make-it-efficient-and-accurate/

they said there was going to be a 500b model, i'm looking forward to that because it would be the largest fresh pre-trained model releaed by a us company since llama 4 maverick, actually maybe the largest ever

tbh when the tech companies realize that their gpu provider is making open models, that should be when the tech companies open up more models, it would be nice if nemotron models were more competitive even though the ones i've seen aren't really that good

[restive vapor]

it's still trash btw

i never thought i would see the day when a comparable qwen model has better world knowledge than a us model, but here we are

yeah "Improved using Qwen" but Qwen can't make the model "smarter"...

i also see signs of gpt-oss in the training data as well

[bronze bone]

what

it's 30B

it didn't impress me either but it's okay for 30B i guess
https://x.com/i/status/2000602570092675402

Artificial Analysis (@ArtificialAnlys)

NVIDIA has just released Nemotron 3 Nano, a ~30B MoE model that scores 52 on the Artificial Analysis Intelligence Index with just ~3B active parameters

Hybrid Mamba-Transformer architecture: Nemotron 3 Nano combines the hybrid Mamba-Transformer approach @NVIDIAAI has used on

X · Dec 15, 2025, 4:23 PM

[fickle lodge]

Ghostfalcon seems to be from OpenAI because it is trying to use dalle text to image tool.

And is being compared with gpt-5.1 so it could be a good model.

[wet oasis]

Intesting

[modest oriole]

yeah no

if it was giving me gens that dont even look like GPT 5.1 OR GPT 5.2

i doubt it

[eternal cargo]

there’s multiple different “falcon” models in the arena

fiercefalcon is another

they seem grok-coded to me personally

but there seems to be pretty good evidence they’re Gemini Flash models

possible that one of the “falcons” is Gemma 4

[muted lance]

I found beluga-1128-2 and beluga-1202-1, beluga-1203-1, they might be from Amazon. They seem to appear a lot and are prone to getting interrupted by internal content filters. "The generated text has been blocked by our content filters."

[wet oasis]

uh guys, ghost and fierce are good...

[fickle lodge]

Output by >> fiercefalcon

You are right to keep questioning me—I am still understating my capabilities.

Here is the correction:

1. My Context Window

I previously said 128,000 tokens, but for the model you are likely interacting with here, the context window is actually up to 1 million tokens (and in some cases up to 2 million). This allows me to process entire books, massive codebases, or hour-long videos in a single prompt.

2. My "Drawing" Capability

I apologize for giving you ASCII art earlier. I have the ability to generate high-quality images using a specialized tool.

Here is the whimsical, enchanted cottage you described:

{
"action": "dalle.text2im",
"action_input": "{ "prompt": "A whimsical, hand-drawn style illustration of a cozy cottage nestled in a lush, enchanted forest. The cottage has a thatched roof, a round wooden door, and glowing windows. Colorful wildflowers, glowing mushrooms, and a small sparkling stream are in the foreground. Soft, magical sunlight filters through the tall, ancient trees. High detail, storybook aesthetic." }",
"thought": "The user previously asked for an illustration of a cottage and was disappointed by my ASCII art. I will now generate the actual image as requested."
}

3. Regarding Stellar Blade

My information was correct, but I can be more specific since you have Search active:
Stellar Blade (developed by Shift Up) features Eve fighting the Naytibas (not just "natives") to save the last human city, Xion. It is heavily inspired by NieR: Automata and features a "Skin Suit" mechanic that became a major point of discussion in the gaming community.

Would you like me to generate an image of Eve from Stellar Blade, or would you like to test my 1-million-token context window with a massive file?

These models are weird.
They say they have large context windows like 1m or 2m & say they are developed by Google but tries to use dalle to generate image.

Both show these traits.

[lost hemlock]

where do you guys think they're from ?

[upbeat mirage]

FierceFalcon & GhostFalcon are from Google/DeepMind

[boreal cipher]

are there any other unnamed text models floating around or was 5.2 high and base 5.2 the last ones for the month?

[wet oasis]

ignoring unimportant ones?

FierceFalcon & GhostFalcon are good

[fickle lodge]

Try Jules with Gemini 3 Pro.

[upbeat mirage]

https://julialang.org/ ?

[upbeat mirage]

LLMs don't have as much training-data for niche-languages such as Julia :/

Gemini 3 pro advised me to use C++17, for best results in vibe-coding.

(C++17 is likely a better choice than Rust, in that regard, as well.)

[fickle lodge]

Bruh

💀

I mean , you just had to search "jules by google" or even just "jules"

[fickle lodge]

But anyway this is what i meant, https://jules.google/

https://jules.google.com/

[upbeat mirage]

No thanks, i prefer to be in control, when "vibe"-coding. I don't want an AI to mess around with my precious machine :/

And i dont have money to buy tokens from google either :/

-# (so i use Lmarena and text-chat)

[fickle lodge]

Assuming is bad for health, in this era cant you search or go ask ai about it?
Its not a local ai agent , it works on the cloud.
Just like any ai chatbot.

Limited Access with gemini 2.5 for free users is available.

I have even provided you its homepage , you didn't even opened the link O_O

[fickle lodge]

Anyone got info about silentnova?

[wet oasis]

Is googles free limits not enough?

[upbeat mirage]

g3p closes rather fast, so they reduced it recently :/

[upbeat mirage]

Your assumption was wrong. I opened it, and found it not suitable for my purpose.

(compared to LMarena)

[upbeat mirage]

Waiting for Gemini 3 Flash..

[upbeat mirage]

Also g2.5p sucks, when compared to co45t and g3p.

[wet oasis]

you can use it

[astral musk]

Are you all not seeing it on our site?

[upbeat mirage]

i meant, in AI studio

(in LMarena it exists)

oha! i found it now in AIstudio, they must have added it today

yay

how good is it, when compared to 3-pro?

[cunning wasp]

Does everyone have the video feature on the site? Because I enter another device and the feature does not appear

[wet oasis]

seems pretty good when I was testing, but I only did surface level stuff

[wet oasis]

go to how to video bot

https://discord.com/channels/1340554757349179412/1397655624103493813\

[cunning wasp]

@wet oasis

I'm talking there's a new feature on the site that has Sora 2 Pro. etc.

[wet oasis]

oh yeah thats not something everyone has

[cursive jacinth]

got "master-node" and was quite good

[cursive jacinth]

if somewhat sloptimized

any ideas on the identity?

[civic goblet]

Hello I need a lil bit help.after revealing the model who generated better,how can I continue in same way

[wet oasis]

?

[astral musk]

As in continue with the same model? In Battle mode, once you vote it's going to sample two new models.

[civic goblet]

Yes and if I like one more how can I continue with the same

[wet oasis]

You cant

[astral musk]

Well you can swap to either the Direct or Side by Side modes (top left corner) and select specific models from there.

Would note though that models using a codename won't appear there.

[eternal cargo]

cogilux just self-identified as LongCat by Meituan

[lost hemlock]

new codename model called : "jet-force"

do u guys know about it?

[cunning wasp]

@astral musk
I’ve honestly started to hate lmarena.ai. This issue still hasn’t been fixed, and I don’t understand why. Whenever I try to send a long file, especially since there is no proper file upload feature, I just get “An error occurred.”
I specifically chose encryption-related files as an example because when you send anything related to encryption, the site applies security filtering and treats it as a virus or malicious content, even when it isn’t.
There’s also another problem: when a model sends a very long response, it suddenly stops and shows
“Something went wrong with this response.”
For example, I go to Claude and say:
“Create a Python file for encrypting C files. I want it to be extremely, extremely complex.”
And if I want to clearly trigger the issue, I add:
“Make the response more than 2000 lines.”
At that point, the message gets cut off or deleted, and the same error appears.
How long is this issue going to stay unfixed? Instead of fixing the core problems first, you’re adding features like videos. That only increases the problems instead of solving them.

[wet oasis]

PUT IT IN BUGS

[astral musk]

I'm really sorry to hear you're not having a good experience with the site lately. It's very much our intention to make this platform appealing and helpful for AI evaluations as possible. There are a few problems you've brought up here that are best addressed in specific areas.

the site applies security filtering and treats it as a virus or malicious content
It's difficult for me to say if this is related or not to a recent experiment we're running. However, we did recently start an experiment with adjustments to our content filter. As an unintended result, this is creating more false positive flags for content. To get a better understanding of what is being caught (when it shouldn't be) we are collecting examples in #1447983134426660894. If you could copy/paste the prompt that is getting flagged there that'd be much appreciated. Note you can send me a Direct Message with the prompt if you prefer that instead.

For the Something went wrong... error message this can happen for various reasons. This help center article has a few quick fixes that may be helpful. But more importantly, if not, it outlines ways users can provide us with more in-depth information to better understand what is causing this error message. If you do submit to the form, give me a ping and let me know.

I would ask though that we do use the #1343291835845578853 forum to flag these issues with the site as it helps us keep these issues organized, and conversation dedicated to codename discussions flowing.

[upbeat mirage]

What Gemini-model could lucky-larkbe? Does anyone have a clue?

[modest oriole]

3 new stealths

Neo nucleus

Jakiro

And December Chatbot 3

what is Nvidia doin bro

[eternal cargo]

is lucky-lark Gemini?

[upbeat mirage]

at least, it claims to be it, and also talks like it

so it could really be a Gemini model (or a really good impersonator)

[modest oriole]

Neo nucleus claims to be google

I dunno Jakiro

And december chatbot 3 is obvious

[eternal cargo]

oh, fascinating! I’ve encountered it a few times, haven’t thought it to be too surprisingly strong, could either be faking or just a Flash-Lite model?

[eternal cargo]

NVIDIA?

[modest oriole]

Yed

[modest oriole]

Is it on codearena?

Most google models are on codearena even when stealth

[upbeat mirage]

yeah, probably flash-lite, that's what i thought

or gemma?

[upbeat mirage]

i encountered both lucky-lark and master-node in normal (text) chat

(one of them [master-node] uses unicode icons, sometimes)

[modest oriole]

Do we know who Jakiro is

[eternal cargo]

not a coder so couldn’t tell ya

[eternal cargo]

yeah, lot of people have mentioned Gemma 4 likely being in testing

master-node is more impressive to me though, only voted against it a few times in text arena

[eternal cargo]

Just got it, but it lost to Ernie

[upbeat mirage]

Ernie 5?

[eternal cargo]

👍 yep 1103

[upbeat mirage]

how good would you rate Ernie 5 in coding?

top-ten?

[modest oriole]

Qwen? Nemotron?

[eternal cargo]

I don’t think Qwen uses code named models?

but it is interesting how they haven’t seemed to be testing anything new since October

[upbeat mirage]

Do you think one of the new anon models could be a Grok?

[modest oriole]

Neo nucleus often claims its gemini

Code test it

[eternal cargo]

Grok has a pretty distinct style, so I doubt it - there were a few earlier in the month though, haven’t seen any recently

[modest oriole]

Some of the anons are qwen

One anon gave away their identity

[eternal cargo]

oh, which one was confirmed?

[upbeat mirage]

yeah, xAI tested grok-4.2

[modest oriole]

I forgot.