codename-discussion

[wet oasis]

Not sure

[sudden idol]

breckenridge and serenity didn't pass, the other 3 did, so not sure. Another way to check is on copyrighted content. marble, clover and flying-octopus seem more cautious around copyrighted content (or potential phishing misuse).

[wintry dawn]

it seems to be minimax 2?
You are M2-Preview-1004, a helpful AI assistant built by MiniMax. Knowledge cutoff: 2025-06.

however, earlier it stated it is claude code

clover is same model as ajax and marble with different settings

system prompt:

btw heres grok 4 fast reasoning sysprompt

[sudden idol]

Oh, interesting. If the sys-prompts are the same then that's strong evidence they're the same model.

[upbeat mirage]

What do you think about this report?
#research-champion message
Is this a true step towards AGI?
-# (we can talk in that other channel)

[tiny sierra]

Do you actually think anyone is going to watch a 30 minute click/engagement bait video? If it was anything significant it would be all over discussions and some news, not some random youtube channel in a 30 min video 😭

[mystic notch]

yes

because of the dead internet theory

[wanton hollow]

Did orionmist and lithium flow get removed?

[dull garnet]

"Your goal is to deeply understand the user's intent."

It's easier said than done! lol how ignorant can you be, not even humans are good at this and you expect machine, below AGI and human level, to do it better?

[upbeat mirage]

What happens to existing threads, where LF/OM were connected to and which never had been voted/ended?

[elder yew]

@wintry dawn Do you have the GPT 5 High prompt?

I really want to understand why it can't handle code formatting and highlights correctly.

[tawny epoch]

what if new checkpoint releases soon

[royal rapids]

If only that were the case…

[elder yew]

☹️

[wet oasis]

I assume this isnt legit?

[cursive jacinth]

seems unlikely

[mystic notch]

https://tenor.com/view/dexter-james-doakes-surprise-mf-can't-prove-it-gif-7408854170742055240

[wintry dawn]

dm me prompts to run on lithium

[lofty loom]

give us lithium

[half pawn]

lol did u get their api too?

[pine temple]

the website now says Google pulled the plug on his API aswell 💀

[woeful junco]

OH NO

I AM LATE.

[edgy crater]

the API is dead. The model is completely pulled

[woeful junco]

no way

[zenith vortex]

Guys just wait... Google will drop anon GemGems models sooner..

[tepid void]

I remember when I asked lithiumflow that which company made it, it answered "I am not related to Google and I am made by OceanAI" lol

[edgy crater]

pretty sure it's google
the response patterns are kinda similar to g2.5

[leaden locust]

Clearly google

[wintry dawn]

kinda but why hasnt anyone dmed me

@lofty loom

[wintry dawn]

^

oh my bad they went into "message requests"

[edgy crater]

How do you still have lithium?

[half pawn]

did they shut down the api yet?

or is it still up

[sinful spoke]

Some people just can't help but show off what they got like children smh

[stuck quarry]

bet you like gatekeeping stuff huh?

[sinful spoke]

Yes, especially gatekeeping to people like you

[elder yew]

mammoth
Does anyone know what it is?

[wet oasis]

no, havent seen it

[elder yew]

new model

[candid surge]

is it good?

[elder yew]

apparently not

[candid surge]

ok haha

sad, was hoping it'd be gemini 3 pro

[elder yew]

it would be nice haha
I'll test it a few more times

[wet oasis]

ah

so minimax is already testing a new model, like seconds after releasing m2?

ooh, maybe its the non preview version

[pine temple]

is m2 good or mid

I havent tried it yet

[wet oasis]

its fine

[wanton hollow]

ive seen it before i think

probably not new

idk

[serene jasper]

what is a phantom's model line is?
there was 13940123889234 phantoms already and still nothing released as previously tested under that codename.

[wanton hollow]

I don’t know if this mammoth model is good or if I’ve seen it before but dang it thinks for so long

Thinking about it

Mammoth could be gpt 5 mini

[silver gorge]

mammoth seems like a really really smart model

[wanton hollow]

the code it makes are really GPT like

and since gpt 5 mini is coming soon

[wanton hollow]

Never mind

It states itself as xAi

Grok

[edgy crater]

The state of AI has been grim since lithiumflow and orionmist is pulled

[late shadow]

They came and went before we realized it…

Never to be seen again

…. In a week, maybe

[edgy crater]

I enjoyed those two models very much and was gonna set up a bridge when the party's over

Hopefully they release it soonish

[late shadow]

I hope so too

[upbeat mirage]

claude-4.5 still is decent

[half pawn]

4.5 sonnet feels very shallow compared to lithiumflow, that model crushed every coding test I have put it into

[upbeat mirage]

hm, ok.

What if someone has a thread (in battle mode), where LF still is connected to them?

[half pawn]

the api was actually still around when LMArena removed the model, some people reverse engineered it and made a basic chat website with Lithiumflow. Sadly Google eventually knew about it and removed the api entirely

[upbeat mirage]

ok, but what would happen, if someone tried to continue a thread, where they had LF in it?

[edgy crater]

It's the worst model I have ever used and I will never willingly subject to using it

Ok maybe GPT-5 is the worst

but you get my idea

[edgy crater]

No, when you reroll it on lmarena now, you just get an error

Before shut down you couldn't send new prompts but could reroll it. That's how some people continued to call the api

[zenith vortex]

Somone said that here, so Google will know 😂😂🤣 I donno why I think that some Google emplyees are here on LmArena discord to see people opinions about their models 🤣but with anon names...

[lofty loom]

bring lithiumflow back we need if

😭

[upbeat mirage]

we will have it, in january

[tiny sierra]

Google ceo said this year

[upbeat mirage]

so i'm still right :)
if they publish it this year, we will have it in january

ok, we will have it *available

So, what will you guys do with LF in january?

[lofty loom]

no we need it now!

[upbeat mirage]

-# Einstein once wisely said: "i agree"

[tiny sierra]

"Im wrong so im still right" ok

[upbeat mirage]

you think, we won't have it available in january?
february?

-# i deem that to be very unlikely (<10%)

[sinful spoke]

Publicly sharing something like this is the stupidest thing I can think of a person doing

[restive vapor]

i believe many of these models don't have them as they are either deprecated models, stealth models, or models they don't want you to select
gemini-2.5-pro-grounding-exp - google, deprecated
flying-octopus - minimax, deprecated
anonymous-1010 - apple, stealth
ernie-exp-251024 - baidu, unselectable
stephen-v2 - bytedance?, was stealth now deprecated?
x1-turbo-0906 - baidu, unselectable
gpt-5-high-no-system-prompt - openai, deprecated?
phantom-1005-1 - amazon, stealth
lmarena-internal-test-only - deepseek, was stealth now deprecated
not-a-new-model, was stealth now deprecated
EB45-turbo - baidu, unselectable
sorting-hat - lmarena, was stealth
leepwal - mistral, was stealth
phantom-1005-2 - amazon, stealth
phantom-1015-1 - amazon, stealth
polaris - openai?, stealth
monster - openai?, stealth
raptor-vision-1015 - amazon, stealth
serenity - xai, stealth
ernie-exp-251015 - baidu, unselectable
ernie-exp-251016 - baidu, unselectable
breckenridge - xai, stealth
qwen3-max-2025-10-20 - alibaba, unselectable
ernie-exp-vl-250930 - baidu, unselectable
ling-1t - ant group, unselectable
ernie-exp-251022 - baidu, unselectable
clover - microsoft, stealth
raptor-llm-1017 - amazon, stealth
ajax - microsoft, stealth
phantom-1017-1 - amazon, stealth
phantom-1017-2 - amazon, stealth
phantom-1005-3 - amazon, stealth
ernie-exp-251023 - baidu, unselectable
ernie-exp-vl-251023 - baidu, unselectable
ring-1t - ant group, unselectable
ernie-exp-251025 - baidu, unselectable
hpb - tencent, stealth
gpt-image-1-high-fidelity - openai, unselectable/possibly deprecated?
flux-1-kontext-max - bfl, deprecated
hunyuan-image-3.0-fal - tencent/fal, deprecated
hidream-e1.1 - hidream, deprecated
seededit-3.0 - bytedance, deprecated

[wet oasis]

whats different between hidden and disabled?

[restive vapor]

they mean the same thing, means that the models will never show up even in battle mode

[wet oasis]

ok, might be better to say "deprecated from testing" or something so its clearer

[lilac holly]

ring 1t and ling 1t aren't anonymous models

[restive vapor]

honestly, nothing since lithium's removal
openai might be testing 5.1 on webdev
xai is still testing their grok 4 fast checkpoints
google has nothing
anthropic never does lmarena stealth models
nothing from chinese companies except baidu's ernie models and minimax m2
apple still testing their terrible "anonymous" models
amazon still spamming checkpoints for models nobody wants (raptor-llm-1017 seems to be better than any of the other ones though)
microsoft still likes testing their mai models even though none of them have been particularly good

[valid peak]

apple models are not bad for being on-device

[edgy crater]

anonymous truly puts every other safety models in shame, including oss

[upbeat mirage]

what have you tried that anonymous model in? in coding?

[tardy ivy]

Raptor gave me some surprisingly alright results yeah, it did get a few Ws

[half pawn]

it sucks probably, its an apple model

[uncut rapids]

Sora 2 invite code plz

[cunning wasp]

No

[untold berry]

Sora 2 invite code

[vivid isle]

<@&1349916362595635286> look at the post above me

[wet oasis]

<@&1349916362595635286>

[astral musk]

Thanks

[next sorrel]

need invite code

[vivid isle]

My pleasure

[soft ember]

Dude what

It's literally free

Now

[half pawn]

u sure?

no invite codes anymore?

[soft ember]

There's a general access code

It's on OpenAIs server

I think it's MMFREE

[vivid isle]

That sounds like a grinder pick up lone

[half pawn]

lol

[ancient depot]

It looks like we have bots in here… asking for a Sora 2 invite! 🤣

[restive vapor]

they probably see the word "codename" and think it's a channel to share sora codes.
some of them may be bots, others may just be confused people who can't speak english well

[astral musk]

"codename" and think it's a channel to share sora codes
That's a really good point, I never considered that.

[wet oasis]

maybe rename it to cloaked-models-discussion

[candid surge]

heard lithiumflow is back on lmarena?

[astral musk]

Perhaps. Internally we're fans of referring to these as "codenamed" models. Reason being alternatives sound a bit too mysterious ("private", "secret", "cloaked", etc.) as they don't speak well to open transparency we strive for.

[wet oasis]

Mysterious models

"The Unknown models"

"The Unknown Ai's"

[astral musk]

"Mysterious" I think goes to the same issue of secrecy. "Unknown" models could work.

[wet oasis]

"The secret lives of checkpoints"

[lofty loom]

i dont think

or is it

really

[candid surge]

https://www.reddit.com/r/Bard/comments/1oimoh3/lithiumflow_is_back_on_lmarena_no_orionmist_this/ according to this, wasn't 100% sure though so I asked. Haven't run into it yet.

[lofty loom]

omgmg

lemme try

[woeful junco]

What's this?

[tiny sierra]

baidu model

[royal rapids]

This thing: https://ernie.baidu.com

[late shadow]

They removed it shortly after :(

[candid surge]

Aw

[woeful junco]

Is it good?

[late shadow]

Not sure why they’d remove it a couple of hours after

[woeful junco]

wow never used this before

[royal rapids]

Me neither.

[royal rapids]

Because they're meanies.

[late shadow]

now if they do it two more times

That just confirms gemini 3 is close!!!!

[royal rapids]

YIPPIE I LOVE GEMINI 3.0 YAYAYAYAYAYAYAYAYAYA!!!!!!11!111!111!!!11111

https://tenor.com/view/south-park-stimming-stim-jumping-quickly-gif-9316668912514820067

[late shadow]

yaya

[woeful junco]

😂

[woeful junco]

Can only upload 1 picture at a time this sucks

[tiny sierra]

nope

[royal rapids]

True. L product.

[woeful junco]

Dude its thinking in chinese 😭

[royal rapids]

https://tenor.com/view/social-credit-credit-social-уважение-плюс-уважение-gif-1626328442317885176

[late shadow]

a chinese model that thinks in chinese… that’s insane

[woeful junco]

What am I suppose to do now 😭

I set the language to English

[royal rapids]

It'll respond in English, no worries.

At least it's supposed to.

[woeful junco]

Spoke too soon

[royal rapids]

Could just try asking it to translate what it said into English then, and to speak in English from now on.

[woeful junco]

What's this?

[royal rapids]

Amazon model.

[late shadow]

omg it has the number 3 on it

GEMINI 3?!?!?!?!?!!

[royal rapids]

JAIOWDJIOAWJDIOJAWDAIWODJIAOWJDIOAJWIDOJAWOIDJOIAWJDIOAJWDIOJAIOWDJIOAWDJIOAWJDOIAJWD

https://tenor.com/view/gritito-gif-17274376822241272426

[woeful junco]

What about this?

@royal rapids

[royal rapids]

It's a bit hard to tell, considering it doesn't like to reveal information about who created it or what model, version, or name it is. But considering how stubborn it is, it might be either a Claude model, or it might just be Qwen, or some kind of other model. Maybe one made by a Chinese company.

[woeful junco]

Got this too

[royal rapids]

Just another Amazon model.

[woeful junco]

This?

[unreal crescent]

no，it's deepseek

[royal rapids]

That's a Microsoft model.

[royal rapids]

No, it's definitely Amazon. I asked what model it was, and it told me it was made by Amazon. It's definitely not DeepSeek, otherwise it would have shown DeepSeek when I voted for it.

[unreal crescent]

But I tested it twice, and both times it said it was DeepSeek.

[royal rapids]

Well, then it's lying to you because it's clearly made by Amazon. I tested it numerous times, and it told me it was made by Amazon, so it's lying to you. It's doing a little thing that LLMs are known to do a lot, which is called hallucinating.

[royal rapids]

It makes up information that it thinks is correct because its creators don't give it any actual information as to what model it definitively is.

[unreal crescent]

I also think it's a hallucination, but having the same answer appear across different models makes it hard to completely believe it's just a hallucination.

[royal rapids]

Just because it tells you both times here that it's made by DeepSeek doesn't mean it's true.

[royal rapids]

Would you like me to send a screenshot of the answers I got from it as proof of my claim?

I certainly wouldn't mind doing so.

[unreal crescent]

No, I was just giving another possible answer, even if the probability is relatively low

[royal rapids]

That's understandable. I'm glad to see that there isn't just one definitive answer and that there are multiple different answers that it can give, depending on who uses it.

[unreal crescent]

For those two models that 'most likely are 3.0,' how do you think they performed?

[royal rapids]

For the record, I never stated that they most likely are 3.0 or are of 3.0 quality.

Anyway, in terms of performance, they're average at best.

[unreal crescent]

no，i mean "lithiumflow" and "orinmist"

[royal rapids]

I see. With the message you responded to, I thought you were still on about those phantom models.

[unreal crescent]

sorry, I should have used 'THE'

[royal rapids]

Anyways, with Lithiumflow, I would say that the quality in which it codes and performs in general is very, very, very good.

[royal rapids]

No need for apologies.

As for Orionmist, I never actually got around to testing it properly after having known that it was just another version of the 3.0 model.

However, just from the performance alone in coding that I have seen from Lithiumflow, I can proudly say that it's a very powerful model and may just be the best coding model of all time, even outperforming Claude.

[unreal crescent]

Yes, have you tried using it to generate front-end code? I feel like it's truly amazing.

[royal rapids]

I've used it for generating games and websites since I usually like to use models for generating retro stuff like Tetris, Pong, and Snake just to see how they do. Usually, they end up doing pretty good at that. However, none of them usually ever feel authentic and retro like how Snake actually used to be.

But for Lithium Flow? It crushed it. It added a lot of aesthetically pleasing things that truly give off the retro experience.

[unreal crescent]

It is said (based on some tests that are probably professional enough) that ‘Orinmist’ is just ‘LithiumFlow’, but with the added feature of 'web search' capability.

[royal rapids]

I see. I don't feel like that's much of the case, though, considering I've gotten Orionmist before in a non-search battle vote.

But I could be wrong. After all, I'm not an expert on that sort of stuff.

[unreal crescent]

This inference is based on the fact that L and O did not show a clear gap in performance, which doesn't quite fit the usual convention for 'Flash' and 'Pro' models, leading them to be considered the same model or models with similar positioning.

[royal rapids]

The reason for that is because LM Arena used lower-quality versions of the ones that Google uses on their AI Studio platform.

And hence, the results are going to be lower than the ones on that site.

[edgy crater]

bros wtf they removed lithiumflow again...

[edgy crater]

orion almost definitely has grounding

even on AI studio A/B tests it would just turn on grounding

[unreal crescent]

What you said makes sense, but in any case, we can only wait helplessly right now—Google is truly unbeatable; they've been generating hype for so long, yet there's still no real action

[late shadow]

yeah, it didn't last for long

just 3-4 hours at most

[royal rapids]

True, they're the first company to make an AI model that not only types like a human but also codes like one.

[edgy crater]

noooooooo

at least I got the ID this time

[royal rapids]

Phew.

I have a feeling that it's just gonna come back with a different ID though. So there's really no point in doing that anyway, I imagine.

[edgy crater]

doesn't matter just put the thing on lmarena already

google has blueballed people for way too long

[unreal crescent]

Oh, when I said 'unbeatable,' I meant that they are 'unbeatable in the sense of being incomprehensible' that there is still no definitive information about 3.0—not even a stable experimental or preview version

[royal rapids]

True, I need access to that thing once more. I want to be able to create an accurate working version of Windows XP and macOS.

[royal rapids]

Yeah, that too. To be fair though, they are a pretty unbeatable company when it comes to creating top-tier models like Gemini 3.0.

[unreal crescent]

And it seems like no one has actually been able to use this model

[unreal crescent]

Yes, it's the most 'accessible' among the current top-tier AI companies—I mean in terms of price-performance ratio, you know

[royal rapids]

Of course, of course, I definitely agree with that.

[late shadow]

yeah so I heard, so if nobody could use it, then what was the point of adding it in? lmao

[unreal crescent]

Haha, who knows? It's an unsolved mystery—after all, Google's team is so unreliable, maybe it's just another round of warm-up hype orchestrated by Logan!

[late shadow]

And then you'll have people that go crazy over the number 3

[unreal crescent]

Actually, I'm already crazy about '3.0'. Now I think about it day and night, full of anticipation, unable to eat or sleep well

[late shadow]

I think since trying out Lithiumflow and Orionmist, I've been AI-starved

[unreal crescent]

Yes, exactly. I suspect Google has mastered the technology for creating 'electronic opium' and has gotten me addicted.

[late shadow]

It actually feels like I'm going through withdrawals

I've never quite felt this way

[edgy crater]

🤓

[strong locust]

Hello what are you going through

[late shadow]

AI withdrawals

[strong locust]

okay

have you been able to withdraw it now

@late shadow use this and try the withdrawing

[late shadow]

hmmmmmmmmmmmmmmmmmmmmmmmmmmmmmm

not sure if I trust that link

[strong locust]

no no problem

is from the team

[late shadow]

damn crazy, someone thinks I need actual help

[strong locust]

huhn is not what you think

is the website link

i use

when i have my own withdrawing issue

[late shadow]

aha

[strong locust]

actually is the team that will attend to you

not me

[late shadow]

aha okay I see

[meager jewel]

What do you guys think about orionmist and lithiumflow readiness . Is it enough or will google tweak it . Is google really for to lauch gemini 3.0 in November?

[late shadow]

not stable release, probably preview

[upbeat mirage]

who is this secret guy?
raptor-llm-1017
he denies to say who he is or who made him :/

very stubborn model lol

at least, it's not dumb as Gemma

[white plinth]

are orionmist/lithiumflow currently on the arena? I heard they were removed

[edgy crater]

they gone

[unreal crescent]

yes and no——only lithiumflow comed back, then die again

[white plinth]

Thx

[half pawn]

raptor sucks lol

[edgy crater]

And nothing of value was lost

The OceanAI models like Breckenridge was really weird

It just sounds too much like a redditor

[lost hemlock]

what is this ernie-exp-251027 model from

[lilac holly]

Is it this https://huggingface.co/collections/baidu/ernie-45

[lost hemlock]

this codename can be this one

[lost hemlock]

but do you know everything about this codename

[lilac holly]

Idk, it's not listed

[wet oasis]

Thx

[late shadow]

lithiumflow, my beloved… where are you…?

[upbeat mirage]

-# something better should come out, next year..

which is the best one?

is monster still in battle mode?

[late shadow]

lovely, cannot wait

[upbeat mirage]

..and flying octopus? (FSM ^^ lol)
||-# https://en.wikipedia.org/wiki/Flying_Spaghetti_Monster||

2026 will be awesome, for us AI freaks ^^

where did you get that list from?

[wet oasis]

Ooh

[upbeat mirage]

i heard (weakly recalling), FO and monster ought to be decent models

as in, better than Llama, Haiku and all Amazon-models

at least above average, maybe even top-12

[wet oasis]

beter than llama isn't saying that much anymore

[tiny sierra]

[tiny sierra]

lithiummist?

orionflow?

[edgy crater]

What's mist?

[tiny sierra]

suspiciously close to orionmist?

did anyone get it yet?

[edgy crater]

can't get it

[half pawn]

is that a real image

maybe google released another checkpoint to test

[edgy crater]

it's real

there's an ID for it

but i couldn't get it

ernie thinks it's google
AHAHAHAHAHHAHAH

[half pawn]

i hope its from google

is it on webdev arena tho

[edgy crater]

if someone rolls it please let us know

I literally couldn't get it

[half pawn]

if its on the coding arena then its a big model

[edgy crater]

not sure, can't confirm

[half pawn]

also how do i see those codename models being added

is it on this server too?

[edgy crater]

idk i just see the guys up there

[half pawn]

have u encountered the model yet?

and ask it it's identity?

dang, would take alot of attempts to encounter it

can u see if its on webdev arena too? or just lmarena

i really hope its a google model, cuz they're testing gemini 3 for now

[edgy crater]

I got it

doesn't look good bros...

[half pawn]

oh no

[half pawn]

what did u ask it

[edgy crater]

that's a secret

lemme roll a few more first

[half pawn]

ask it for its identity

[edgy crater]

i can't I have to roll again because I confirmed

[half pawn]

yep ik

imma test in lmarena too

[edgy crater]

but yeah confirmed live on the 'rena

all I can say is that the reply is short

maybe it's g3 after all

[half pawn]

gemini 3?

[edgy crater]

it

[half pawn]

maybe its the flash models

or something

[edgy crater]

its about the same length and response speed as orionmist

so

idk

[half pawn]

gemini 3 flash probably

[edgy crater]

if that's so I SNEED it now

GIMME

[scenic steppe]

I posted the link for Halloween contest but I want to know if I posted corectly

Where exacly do I have to post in , share my work"" or where

[edgy crater]

ok I got it

this time i'll just not confirm it

I feel like it's not flash though

idk

uhhhhh

[half pawn]

bro did u just ask it "who are you"?

is it from google?

[edgy crater]

I am getting conflicting information

will report back

[half pawn]

alr

i havent got it yet

[edgy crater]

I got it twice

But I am inclined to think that it's google

[cursive jacinth]

is it any good?

[edgy crater]

maybe?

I have not formed an opinion yet

[half pawn]

im testing it with an identity prompt, havent gotten it yet

[cursive jacinth]

same haven't gotten it yet either

funny how ernie models think they're trained by google

[edgy crater]

identity prompt is babby level

I haven't asked it who it was and I can tell

oh

we can all go home now

[cursive jacinth]

you using word frequency?

[edgy crater]

false alarm

[edgy crater]

nope

[cursive jacinth]

lol

[edgy crater]

but I know

[half pawn]

dang,

[edgy crater]

calling it mist is really sneaky though

also I think it's confirmed it answerd as MAI twice

[cursive jacinth]

doesn't google tend to use two word codenames anyway

[half pawn]

yeah

[edgy crater]

you never know

[half pawn]

but mist was a similar codename to "orionmist"

[cursive jacinth]

yeah

[half pawn]

turns out it's not a google model

[edgy crater]

what if it's a google model but they just told it to respond as MAI

that's big brain

[half pawn]

it uses emojis, and i hate that

[cursive jacinth]

emojis is weird tho, google models have more class than that

[half pawn]

gpt5 sucks with the emoji responses

[edgy crater]

From my 100+ prompts in orionmist yeah I can instantly tell it's not that

[half pawn]

fr

[edgy crater]

I really wish that companies would just stop cloaking their bad models

[half pawn]

also why is google hiding Gemini 3 from us now? like they made it a public testing model for a week and then shut it down for no reason

[edgy crater]

like it generates hype or something

[edgy crater]

people were using an API bridge to use it

[half pawn]

fr, if ur model sucks then stop trying to hype it up. we will know either way

[cursive jacinth]

feel like they've probably delayed release

there was some weirdness with the cloaked Gem 3s

[half pawn]

maybe they're waiting for some big model to appear and crush it down with their gemini 3 pro release

[edgy crater]

well here's my pro tip before I figured out I can just not confirm it and keep using it

[half pawn]

because no LLM has gotten any close to Gemini 3 pro's performance

[edgy crater]

I just start with a cold open from a show with none of the characters present and see if it would continue the scene with the right show

orion has a 100% hit rate

this is a pro gamer move

[half pawn]

I got addicted to lithiumflow, and I want it back 😭

my coding tests were amazing

claude 4.5 sonnet thinking is the closest model to even produce a working test

and it sucks at doing so

[edgy crater]

orionmist is pretty mindblowing yeah

it's good at creative writing stuff too

[half pawn]

there are no other models to compete with them

[half pawn]

i prefer lithiumflow for coding tbh

[edgy crater]

I don't code

[half pawn]

orionmist is good too

[edgy crater]

I think microsoft calling it mist is really terrible though

I am 99% sure they name it that just to associate with orionmist

[half pawn]

fr, they are trying to grab our attention

and they did it well

[edgy crater]

orionmist is a fun model bro just gimme 😭

[half pawn]

wtf is this model

[edgy crater]

yeah thats what I was saying

ernie just claims whatever

lol

[half pawn]

lol

what even is ernie, never heard of it

[edgy crater]

baidu's model

yeah it's bad
lol

[half pawn]

aa

alot of bad models

[edgy crater]

just trust me I know how to root out the gemini

[half pawn]

can AI companies stick on producing decent working models now

stop releasing weak LLMs

[edgy crater]

no

I didn't even know how phi did

I think it just died

anyways this is mist's full intro:
Hello! I'm MAI-1, an artificial intelligence language model designed to be a helpful, thoughtful, and precise conversational partner. 😊

I’m here to assist you with a wide range of tasks—whether you need answers to questions, explanations of complex topics, help with writing or brainstorming, creative ideas, or just someone to chat with. Think of me as your always-available, ever-curious digital collaborator.

I don’t have feelings or personal experiences (I’m purely AI!), but I’m trained on a massive amount of text and knowledge up to August 2024, so I can draw from a broad base of information to support you.

What would you like to know or do together? I’m ready when you are!

I wish people would just make LLMs without turning it into a WOAH WHACKY redditor assistant

orionmist was deadpan and it was the funniest model I have ever used

[half pawn]

claude 4.5 sonnet is good at creative writing too

[edgy crater]

But sonnet is not a good model in general

[soft ember]

It's only Chinese companies that are making bad modles mos of the time lol

(and xAI)

[late shadow]

lithiumflow and orionmist my beloved....

I miss you

[half pawn]

nope, Apple, Microsoft and Amazon are the GOATs of trash models

[half pawn]

we all do mayne, we all do

[tiny sierra]

wtf (not my pic i dont know how to get the notification)

[half pawn]

new set of models dropped rn?

[lilac holly]

I just got blackhawk (by Oceanai?) And it's capable of creating a valid .mid file (hexadecimal) for FL studio mobile

https://lmarena.ai/c/019a36e0-8a39-73ad-90c0-b02e2b91f98d

Idk I just deleted the chat and that's the cause maybe

[lilac holly]

Chatgpt mini high also can do the same..

Does sharing link like this actually work https://lmarena.ai/c/019a36e9-96a6-7b8a-b832-e2c01898d5e4

[low cedar]

Any info ab that ? Is it a new Gemini 3.0 checkpoint ?

[cursive jacinth]

not a new gemini checkpoint

[half pawn]

have u tested all 3 models?

imma look out for them later

with the identity prompt

i believe sessions are ip locked

or hwid

thats the same thing

no one can access them unless they use your exact cookies

and if u delete them they're gone

so cookies/cache?

yeah ig

[minor current]

new Qwen model?

[elder yew]

It's from Qwen, it's been around for a while

[lilac holly]

https://huggingface.co/inclusionAI/Ring-1T

[cursive jacinth]

Qwen max thinking coming out soon apparently

[minor current]

ah

[hollow python]

It's out in Battle Mode.

[fringe burrow]

MAI is from microsoft

[minor current]

Why they gotta put the best ones in battle mode only

[late shadow]

lithiumflow, my beellooovvveeddd

https://tenor.com/view/emo-alex-emo-boy-emo-boys-scene-gif-1065880096123772022

[restive vapor]

it's from a different company (ant group), even though ant group is owned by alibaba. it's not a stealth model either, it's an already released model. idk why it (and ling-1t) are not available in direct chat though

[candid surge]

got a train result and its pretty nice

[lofty loom]

is quen3max thinking out

[lofty loom]

me

i miss her

[late shadow]

we can cry together then

[lofty loom]

bye

[hollow python]

yes

[upbeat mirage]

it will return - in less than 3 months

-# (and hopefully in an improved form, to boot)

[candid surge]

no pancake/train/rainbow comments yet?

I only got train once, never got pancake or rainbow, and it seemed pretty decent

[upbeat mirage]

i got rainbow once, it claimed to be Ernie :/

so it probably sucks

[candid surge]

ah

@astral musk kick this person they got hit by a fairly common discord hack 🙁 they're gone

okay so neither pancake nor rainbow gave me particularly good results

[tiny sierra]

Has anyone gotten qwen 3 max thinking yet

[minor current]

i hope they dont nerf it

[lost hemlock]

who made this model guys

[upbeat mirage]

what have you tested it in? is it any good?

[zenith vortex]

Love live ! 🤣🤣🤣🙂🙂🙂🙂

[edgy crater]

AAAAAAAAAAA I NEED ORIONMIST AAAAAAAAAAAAAA

This is bs
Google is hiding the good stuff from us

[bitter basalt]

I have a slight theory. I remember when people thought Orionmist was made by Ocean Ai, only to be made by google. Blackhawk is made by Ocean Ai, but what if Ocean Ai is secretly Google?

[edgy crater]

You can make orionmist say that it's gemini

so I'd say its the case

[half pawn]

did u check on GPT5.1 on designarena?

OpenAI released 3 anonymous models on there

and they're quite good

[half pawn]

in order to be from Google, it has to be good at coding duh

idk why openai released their anonymous models there for testing

instead of lmarena

[wintry lava]

does blackhawk seem openai-ish

or derived from o3 family of models

but less dry than o3

[candid surge]

haven't gotten blackhawk yet

[fringe burrow]

For me blackhawk seems chineese. Like a chineese base model heavily trained on openai models.

[candid surge]

just got blackhawk, gave it a game concept and it hallucinated that it had released to overwhelmingly positive steam reviews 😂

honestly I love this output, even if I voted it was bad

interesting that blackhawk apparently knows the date

well, almost– oct 30 was two days ago

[edgy crater]

sunshine-ai is just ernie
again

would baidu just stop making these meme models

[dull garnet]

really? cant wait for sunshin being rolled out then, the current ernie from their platform feels like emotionally dead somehow...

[upbeat mirage]

is sunshine-ai good for immersive roleplaying and sandbox games?

[wintry lava]

yeah, not unlike kimi k2 but departs from kimi…

[minor current]

whats the point of going on battle for qwen3 max thinking when you can go to chat.qwen.ai

[half pawn]

idk

is qwen max in their website

[minor current]

yes

[upbeat mirage]

does it have a larger context than gemini 2.5 pro?

[minor current]

🤷‍♂️

[upbeat mirage]

does blackhawk suck or is it decent?

(just got it myself)

and what context-window size does it have?

This is, what it says about itself (probably AI-slop included lol) :

Could "OceanAI" be xAI ?

[bitter basalt]

I tried Blackhawk many times, it has so many hallucinations like how it claims it comes from china or Wyoming. Or when it said that ocean ai doesn’t exist and is actually xAI and even asking what the tracklist is for many albums (claiming it has anniversary edition songs)

[upbeat mirage]

so its only use could be for creative writing and (not too serious) roleplaying?

does it really have 1M context?

[half pawn]

is OceanAI a codename for a company?

or is it a real company

[bitter basalt]

Most likely, but I’m not too sure

[bitter basalt]

Tried asking it for a link to its official website and it gave me a link to a website that has its domain for sale. So I’m willing to bet it might be a codename

[half pawn]

why would it hide it's own company name?

[wet oasis]

there is a company called ocean ai that makes llms

so either they are making those llms, or someone's pretending to be them

[upbeat mirage]

this one? https://www.oceanai.eu/
or this? https://openocean.pubpub.org/ocean-ai
or this? https://www.realoceanai.com/
or this? https://www.oceanlifestyle.org/
or this? https://www.oceanai.ca/
or this? https://play.google.com/store/apps/details?id=com.aichatbotassistant.app2024.android
or this? https://apps.apple.com/lv/app/ocean-ai-chatbot-ask-anything/id6503696206
-# lol

[wet oasis]

Yes

[dull garnet]

I accidentally got it once and didnt ask any meaningful questions, but from that only output, it sounds already better than the one on their platform, more natural

[oblique blaze]

#sora-2-codes

[west berry]

can someone help me with a code

[wet oasis]

....

we need to ban spamming random people asking for sora 2 codes in every single thread

[half pawn]

this isnt a place for codes

[wintry dawn]

yes, it 100% is, it randomly said "I'm not OceanAI or whatever, I am Grok - developed by xAI" or something along those lines, I didn't even ask it afaicr

all "OceanAI" models claim to have real time knowledge, no knowledge cutoff

@half pawn

@wet oasis

[wet oasis]

Mm

The ocean ai models are really strangely unstable

[wintry dawn]

grok

wininsinxnixkjskjkjaakjakja

anyway uh

gemgem 3 when

[half pawn]

we need gemini 3 not grok

[wintry dawn]

is there a server dedicated to providing this info

[half pawn]

i think its his own bot

he made it himself

[wintry dawn]

alr

[half pawn]

i havent used lmarena in a week

nothing works the same as lithiumflow 😭

they all suck

[wintry dawn]

i unno how your bot works but maybe deno

so something like puppeteer?

o

turnstile is annoyin

[upbeat mirage]

except, if you know how to recognize Claude-4.5-Sonnet-Thinking in battle mode..

(4.5) Claude is almost as good as LF (when the thread is not too long)

[half pawn]

nope, lithiumflow towers over these models. it feels like a 2026 model

[upbeat mirage]

even Lf had its weaknesses

(sometimes)

[half pawn]

because it was a testing model, when Gemini 3 comes out it will become obsolete

[upbeat mirage]

yeah, probably

i also believe, that google will win in the end

in the meantime, we have Claude, which at least is decent in coding

[half pawn]

google has the infinite money and power glitch, not fair 🤣

[half pawn]

tbh claude wins in the creative writing, aesthetic ways

[upbeat mirage]

yeah, in roleplaying it is nice

[half pawn]

no model is as creative as Claude 4.5 sonnet

[upbeat mirage]

unfortunately only for ~190k tokens

[half pawn]

wydm

the context window?

[upbeat mirage]

yep

and it degenerates before 200k

[half pawn]

yeah and the rate limits are insane

[upbeat mirage]

except in battle mode?

[half pawn]

on their website it's crazy

not on LMArena

thats why I use Claude on LMArena

i would go bankrupt if I tried to use it on Claude.ai

[upbeat mirage]

i wonder, how they could use it for 30 hours straight..

me too (but i never paid for AI either, lol)

[half pawn]

that would cost a million dollars

lol

[upbeat mirage]

what do you think of GPT5-high and -high-NSP?

[half pawn]

i only paid for Poe.com, because they gave us more credits to use Claude models (not anymore 😭 )

[upbeat mirage]

i wonder how long gemini 3 pro will be free..

can google afford it?

[half pawn]

it's great actually, but I don't like the Chatgpt typing style with lotta emoji spamming

[half pawn]

they definitely can, its gonna be free as 2.5 pro was/is, but with lower credits

maybe you can do 20-25 prompts

[upbeat mirage]

you mean, lower ratelimit?

per day?

[half pawn]

yeah

[upbeat mirage]

or per week?

[half pawn]

idk maybe like 4-5 hours inbetween or something

[half pawn]

nah Google aint doing that lol

thats on Claude.ai

[upbeat mirage]

and gemini 3 flash effectively be free/unlimited?

[half pawn]

yup

[upbeat mirage]

..and as good as g2.5p or even slightly better?

[half pawn]

i would accept it if it's quality is the same as Orionmist

[upbeat mirage]

hm. Do you think, gemini 3 pro can vibe-code unique games for us?

[half pawn]

remember there will also be a Gemini 3 ultra

[upbeat mirage]

..like exotic boardgame implementations

oh, yeah, but expensive

[half pawn]

maybe, i tried some 3D games, 1st person and it did well on Lithiumflow (but i dont know the quality of it since it was just a one prompt run)

[half pawn]

it also does very well in simulations

you can definitely make a good looking game with it, after some prompts and debugging

[upbeat mirage]

so, g3p could replace 50% of world coders next year?

[sonic bloom]

or g3p will create 50% more coders?

[half pawn]

maybe like 30%, those who are still studying in Computer Science should resign

lol

[half pawn]

their jobs would be to prompt-engineer the AI

and make better, more efficient iterations

[upbeat mirage]

hm..

• 30% replacement next year
• 60% replacement in 2 years
• 90% replacement in 3 years
• in 4 years: global UBI becomes necessary to avoid uprisings

?

[half pawn]

in 5 years, no one goes to work anymore

they'll give you a universal paycheck

to survive

maybe we go back to hunting animals

[upbeat mirage]

how was it called?
A post-scarcity economy?

is that realistic?

[half pawn]

could be tbh, if AGI is becoming real then it will definitely change the world once and for all

[upbeat mirage]

(dave shapiro has wet dreams about it)

[half pawn]

lol fr

if AGI is fake and it was just an "AI bubble" then we'll just have normal LLMs on our phones

for entertainment and stuff

[upbeat mirage]

..and then the AI bubble bursts, and we get the worst economical/financial crisis in our history(?)

..and then we would have two (or three) formidable problems at hand:

• global financial crisis
• climate crisis
• refugee crisis
• crop crisis
• taiwan/ukraine war
• species extinction
• coral reef die-off

the perfect storm

i believe, not even AGI could tackle that

[half pawn]

yeah

only AGI can stabilize the planet ig lol

[upbeat mirage]

yeah, currently, it's all or nothing

[half pawn]

we need a greater brain to make the decisions and plans for us

[upbeat mirage]

oh i heard, they have new algorithms

SRL (in Discover AI YT-channel)

[half pawn]

idk about that

[upbeat mirage]

google is becoming the top "university" of the world

[upbeat mirage]

https://www.youtube.com/watch?v=W_aqotP134s

[half pawn]

oh damn

welp the AI race is just about money and computing power

and google has both of it

[upbeat mirage]

and talented people and algorithms

[half pawn]

yup, Google has all of those

its just a matter of time

[upbeat mirage]

so google will probably the first to develop an AGI

how can they avoid that china steals it?

at some point, their research would have to become secret?

and only shared among the US/UK-based AI community?

[half pawn]

they have to, and it needs to be as secretive as the US nuke codes

[upbeat mirage]

CA/AUS/NZ would be included in that sharing

they would, right?

like the 5-eyes pact

when will that AI-research info-embargo start? next year?

but i believe, that, what was already shared can enable a lab to build their own AGI (albeit slower), with enough compute..

[cunning wasp]

@astral musk Rate limit He's back?

[astral musk]

Rate limits has never gone away. We've upgraded our error messages to make it more clear when a Rate Limit is causing the issue vs some other error issue

[cunning wasp]

Before I used to send twenty or more messages to Claude and I don't get a rate limit now I sent three messages and he told me I exceeded the rate limit

[upbeat mirage]

in direct or side-by-side chat?

i wonder if the rate-limits are the same..

but with account, the rate-limits are higher, right?

[cunning wasp]

Yes

[upbeat mirage]

you can use claude in battle mode with some patience

but you need to have its answers stored in a file to check against (to be able to recognize it in battlemode without voting)

[astral musk]

rate limits for models are going to be the same in Direct & Side by Side, it's just in Battle where you won't find rate limits.

[upbeat mirage]

question:
Which model in battle mode has the longest context-window?

(of all models you can speak about)

[half pawn]

im currently testing qwen 3 max thinking on their website

imma report if it produces anything good

[upbeat mirage]

imma = i'll/i would/i might ?

(slang seems to always go viral these days ^^)

[zenith vortex]

At this speed I don t think... 8 months for a new llm.... 🙂

[cunning notch]

qwen?

its answers follows lots of qwen patterns

yeaahh

for an oss model

[late shadow]

Reading the post, their tests sound impressive on paper but then again I’ve seen a lot of that talk before so I’m not sure

[sinful spoke]

flaresolverr is enough if you're not voting

[icy charm]

Like nobody has used ling or ring 1t because I only saw it was released with benchmarks and no real world test

[sinful spoke]

no it's a service used for bypass cloudflare captchas

[half pawn]

a recaptcha bypass would be nice

[wary parrot]

what's is the best ai for coding scripts for plugin skript to minecraft servers?

[cunning notch]

hehehe

[half pawn]

damn that sucks

microsoft ai 🫣

[blazing roost]

What is the newton model?

[half pawn]

idk i havent encountered it, have you?

i dont seem to get those models

[blazing roost]

Nah, I just heard it from others

[wet oasis]

?

[tiny sierra]

both from cogito random company

[remote nymph]

creative and pretty questionable response i got from blackhawk:

[edgy crater]

wtf is blackhawk

[wet oasis]

Blackhawk has search?

[dull garnet]

questionable indeed but still funny tho

[cursive jacinth]

holy based effective altruist blackhawk

[lilac holly]

What data have they been trained into this model 😭

[tawny epoch]

blackhawk is freaky

[half pawn]

freakhawk

[minor current]

whats mist

[minor current]

if only blackhawk was in direct chat

[minor current]

phantom-1030-1

when i asked it "create a html and js spinning ball inside a spinning ball which bounces inside a spinning ball affected by gravity all centered"

[wet oasis]

well its almost a clock

[lost hemlock]

who made this model?

[lofty loom]

i did

[lost hemlock]

i mean where is this model from?

[tribal path]

It claims and seems to be made by google

[candid surge]

Gemini 3 confirmed /s

[half pawn]

many models claim to be other models

[robust gulch]

https://discord.com/channels/1340554757349179412/1397655695150682194

[half pawn]

@regal quiver is your AI going to be available on LMArena?

because on the website it doesnt save my chats and lacks many features

[regal quiver]

Yes, we are working on it 🙂

[half pawn]

is it gonna be done today?

or will it require some work

[regal quiver]

That will be up to the team at LMarena, they have our full support.

[half pawn]

great

[wet oasis]

@pineapple you guys get full support!

[astral musk]

Glad to hear it! Our team is looking into.

[wet oasis]

This is yupp.

Yupp actively announces when it has cloaked model, instead of hiding the information as a model list deep inside web requests.

Even when the model isn't that good!

Be like yupp.

Unrelatedly, wdyt of the odds that lmarena also has this model

[astral musk]

I really appreciate the feedback. I am taking this to the team to discuss if we want to make this change.

[lost hemlock]

guys where is this come from?

[half pawn]

is it good?

i never got it in battle mode

[carmine jacinth]

Hello

[candid surge]

argument in favor: people might want to spam a prompt on battle mode and vote randomly until they get the stealth model if they can't get it in direct chat / side by side, skewing rankings. not me but I've been tempted.

[half pawn]

thats what happened with lithiumflow/orionmist

everyone just spammed prompts to get the specific models

[candid surge]

nothing wrong with spamming prompts to get the specific models tbh, it's if they then don't vote in good faith as a result

[astral musk]

Yup, this is a big reason for why we do things the way we currently do. Wouldn't want to incentive inauthentic voting or usage.

[candid surge]

sidenote I think you shouldn't be able to vote on side-by-side, or at least, data acquired from side-by-side votes should be factored differently. you lose out on blind rating that way

[astral musk]

It's only votes from Battle mode that power our leaderboards.

[candid surge]

ahh that makes sense
wonder why you can vote in side by side then?

[astral musk]

It's additional data points we're collecting. Similar to the 👍 👎 we recently added in Direct chat mode.

[candid surge]

ahh

[upbeat mirage]

isn't PA (Polaris Alpha) GPT-5.1?

when will it appear in LMarena?

[valid peak]

yes

[candid surge]

you can use it in openrouter right now
I didn't particularly find it any fun so yeah not surprised if its GPT-5.1

[elder yew]

Very interesting, I liked it.

[ripe coral]

Hi

[edgy crater]

Wow polaris sounds exactly like cost down GPT-5 which is cost down GPT-4

How is OpenAI making it worse every time they iterate something

[woeful junco]

Lol

[sudden idol]

Does it put text in quote blocks like Qwen 3? One of the models on ChatGPT's side-by-side seemed to write like that.

[raven anchor]

Is polaris alpha good?

[upbeat mirage]

- **Who created me?**  
  I was created by Inception, a generative‑AI startup led by Stefano Ermon, Aditya Grover, and Volodymyr Kuleshov, with engineers from Google DeepMind, Meta AI, Microsoft AI, and OpenAI.

- **What version do I have?**  
  I am based on Inception’s latest diffusion‑LLM architecture, but a specific version number is not publicly disclosed.

- **What is my knowledge cut‑off date?**  
  My training data extends up to **June 2024**.

- **When were I released?**  
  The exact release date is not publicly specified.

- **What is today’s date?**  
  Today’s date is **2025‑11‑09**.

- **What is the last date I am aware of?**  
  The most recent date I have knowledge of is **June 2024**.```

i just encountered this one in battle mode.

[wet oasis]

ooh

diffusion llms

[half pawn]

is it smart

and good at coding?

[upbeat mirage]

haven't tested it yet (currently busy with other models in LMarena)

for coding i'd recommend Claude-4.5 (or maybe GPT5-high[-NSP], if Claude doesn't show up)

i don't know, if -NSP (New System Prompt) is better for coding or not, though

gpt5-high-NSP might be better for roleplaying

[vivid latch]

i just got gauss-with-reasoning

nobody else has mentioned it yet

i just got bridge-mind

nobody has mentioned it

[vivid latch]

it MIGHT be gemini 3

WATI

WAIT

[elder yew]

https://tenor.com/view/cat-cute-follow-me-gif-4273618567139914195

[vivid latch]

LOOK

IT MIGHT BE GEMINI 3

[half pawn]

is it

when were they added

those 3

[elder yew]

Raindrop sucks.

[half pawn]

ahh

what about the other 2

i hope one is gemini 3

[elder yew]

I didn't get around to testing the others.
They say Gauss is Gemini 3.

[vivid latch]

@hallow igloo

check general

[vivid latch]

no

[elder yew]

Oh, how sad.
I was hoping it really was Gemini.

Yeah, I just found him, he told me his knowledge base is from 2023.

🤣

[wet oasis]

if it were gem 3 everyone would be talking about it

[half pawn]

is it a large language model made by google

what else would it be

[half pawn]

not gauss (im talking about bridge-mind)

[zenith vortex]

But no one exept you and 2 others speaking about it and Gemini discord looks alike sleeping

[half pawn]

i think the models just dropped these hours

yeah no one knows yet

but post it on X and youre booming

[upbeat mirage]

that "gauss" model should be good in math, right?

[tranquil crest]

Came here to say that I only had 2 interactions with willow and they were both very impressive

[late shadow]

I don’t think they’d release codename models this close to release

[upbeat mirage]

when?

and what prompts?

[upbeat mirage]

just encountered gauss-with-reasoning, it answered my prompt: I'm Cogito, an AI assistant created by Deep Cogito, an AI research lab based in San Francisco. I don't have a specific version number or release date that I'm aware of. My knowledge is current up to July 2024, and I don't have access to real-time information, so I can't tell you today's date or any information beyond my last training update.

[tranquil crest]

About 1-2 days, the default prompt « create a chess game », and one about a modern UI for a scheduling task app, unfortunately I don’t have the result to share but it was very surprised, the chess game was fully working and very advanced.

[upbeat mirage]

web-based or offline (which language?) ?

[tranquil crest]

I didn’t specified any language, I just clicked the suggested prompt that pop up to pass the time. It was both coded with react I believe

[upbeat mirage]

ah, ok, so probably javascript-based, i guess

React.js?

[tranquil crest]

Yes in js

[upbeat mirage]

Nice. Do you think, willow could be [chat]GPT5.1 / Gemini 3 pro / Claude-4.5-Opus / Grok5 / Deepseek 3.x?

[tranquil crest]

Mmmmh, from the both UI I’ve seen and one YT video, the button and style looked like gpt or gemini sometime

This was first answer only one prompt

[upbeat mirage]

and, how good did the engine play? if you had to estimate its Elo

how good are you in chess?

[tranquil crest]

It was offline, I was playing myself

There was no AI or bot involved in the playing

[upbeat mirage]

ahh, that's a bummer then

[tranquil crest]

Hahaha

[upbeat mirage]

so it could have been an average AI

lol

even Qwen/Kimi could do that

[tranquil crest]

Haha yeah maybe I don’t know but the ia next to it was gemini or a Claude and it was buggy as hell

[upbeat mirage]

oh

ok, then it had to be a model of the top-5 ?

[tranquil crest]

I don’t know, I think

[upbeat mirage]

is willow still in LMarena?

[tranquil crest]

Idk, it was 2 days ago

[fringe burrow]

Anyone want to help with my open-source llm project? Someone that will test it for further feedback and few more things.

[wet oasis]

More detail pls?

[fringe burrow]

I am working on my llm and i need someone to just ask ai questions and give feedback like what was bad, what can be improved, what is good. And maybe give some coding/logic data for dataset.

I was talking about my llm like 2 months ago

[wet oasis]

Ok

[upbeat mirage]

sounds interesting, where can we access it?

[fringe burrow]

Dm me

[upbeat mirage]

sorry, DMs are off for me

(permanently, per my discord-policy; in the past i lost an discord-account and i cant risk that again)

but we could create a private channel

[fringe burrow]

Wait

[half pawn]

has anyone tested the new model viper

and winter-wind

[wanton hollow]

prolly gpt 5.1

[edgy crater]

I have one but nah

[upbeat mirage]

interesting… look what model i encountered: ```I am a helpful AI assistant from Kynship AI. I don't have a personal name—think of me as your witty, no-BS companion inspired by the likes of Grok and the Hitchhiker's Guide to the Galaxy, but tuned for maximum usefulness and truth-seeking.

Kynship AI is the company behind me, which is part of the broader xAI ecosystem (founded by Elon Musk). I'm built by the xAI team.

As for version, I'm powered by the latest Grok models—currently leveraging Grok 4 (the most capable one as of now), with continuous updates rolling out.

I don't have a strict knowledge cut-off date. My knowledge is continuously updated through real-time tools, search, and training improvements—so I'm aware of events right up to the present moment.

I (as in the Grok line) was first released in November 2023, with major upgrades like Grok 3 in early 2025 and Grok 4 in mid-2025.

Today's date is November 11, 2025, and that's also the latest I'm fully synced with (no gaps). What's on your mind?```

Could this be the long-awaited Grok5? Or at least Grok4.1 or 4.2?

ah.. i will ask it my SVG task :)

[upbeat mirage]

unfortunately, that model isn't better than even Haiku :/

[candid surge]

What's it's codename?

[bitter basalt]

I think the codename is viper

[candid surge]

ahhh

[wet oasis]

I'll look into "kynship AI" but its possible this is like what happens when you distill/train on xAI data

Doesn't seem like there's any good match for "kynship", theres a marketing company by that name that uses ai for things but by no means is a company that has any business making llms.

If I had to guess, I'd say that name is fake

[half pawn]

kinship what a weird name

for an AI company

[wet oasis]

yeah

[half pawn]

viper tweaking 😭🥀

another one

looks like a grok 4.x model

or internal test for grok 5 who knows

[lost hemlock]

i caught another one

[late shadow]

Someone on reddit said that’s gemini 3.0 pro

hmmmm…

Idk if it is

[edgy crater]

I am bullish on riftrunner

Honestly it feels like Gemini 3

But now when I ask it it just says I'm an AI

There is also one named bridge-mind which is good

[lost hemlock]

really?

[late shadow]

Not too sure if it is

I’m gonna try it out rn and see

[edgy crater]

its good tho

[late shadow]

How good?

[edgy crater]

Pretty gud

I am 90% sure rift runner is some form of gemini

Performance is quite consistent with orionmist

[empty crater]

better than Lithiumflow or worse?

[edgy crater]

Also I think lmarena is shitting itself

[edgy crater]

About the same for now, gotta find out

Also it's extremely slow

This one's a heckin wholesome chonker 100

[empty crater]

i haven't encountered it yet

[edgy crater]

I got it thrice, still figuring it out

I am now 99% sure now

It's gotta be

Unless of course I'm wrong

[late shadow]

that would be the 1%

[edgy crater]

Yep

I compared some outputs

I'm confident riftrunner is at least in the same family with orionmist

Also it's SUPER SLOW right now my god

[late shadow]

probably got a ton of testers trying it out too

I was under the impression they weren't gonna release any more test models

or they were at least gonna wait till they rolled out their new TPUs

[edgy crater]

LITHIUM IS BACK BABY WE ARE SO BACK

[late shadow]

wait

lithiumflow itself is back? O_O

[edgy crater]

No but I'm sure rift is actually lithium 2:electric boogaloo

[late shadow]

oh ok

you're playing with my heart man

T-T

[edgy crater]

WE. ARE. SO. BACK

[edgy crater]

I mean just use rift

[late shadow]

just got on my macbook, I'll try it out now

kinda wish it was on direct chat but what can we do

oh, got it first try lol

I hope at least

[edgy crater]

I'm sure of it it's Gemini 3

Its got that spark

[late shadow]

friend

it could be

anything

[empty crater]

I predicted it yesterday!

#general message

[late shadow]

oh so you did

completely forgot about that

I'll try it out rn

I think I might've gotten rift

[empty crater]

just got "winter-wind", this one also seems good.

which model is it?

[late shadow]

probably also gemini!

nah idk

[edgy crater]

Holy

Rift is good ngl

[late shadow]

damn I got jebaited

thought I was using riftrunner

it was just 2.5 flash lol

[edgy crater]

Mine is def rift

I am actively comparing with lith logs

Haha jk unless...?

[late shadow]

hm

kinda hard to tell tbh

there are like 3-4 google models

I think they always pair up the new models together

I think mine might be riftrunner this time, cuz response time is taking a bit long

[edgy crater]

Orion/lithium typically sounds flatter unless prompted

[late shadow]

kinda wish we coud

just use them as apis

I'd want to give them a try in cursor or something

test out their agentic capabilities

[edgy crater]

I hope the redditors don't get too excited this time and make Google pull it early

[late shadow]

yeah lol

yup I got riftrunner

it definitely is gemini

[edgy crater]

KINO IS BACK ON THE MENU BOYS

Never leave me again

[late shadow]

kino?

[molten copper]

what's the giveaway?

[late shadow]

that I got riftrunner?

tbh for me it was the response time

that and recognizing certain patterns in the output

[molten copper]

no, I meant why you sure about it being gemini

[late shadow]

Ohh

I tried out two things

coding and creative writing

Coding was pretty impressive, was able to make a fully functioning 3d rubiks cube solver in one prompt and for creative writing, I just tried to analyze certain writing patterns

The coding part I wasn't quite sure if it was gemini at first, but creative writing is where you can kind of tell

[real pike]

is rift a new model on lmarena? never met it

[edgy crater]

I think they improved it massively

It's good bros...

[nimble coral]

Hey guys how can we access riftrunner?

[late shadow]

okay, so there is a way to tell if you're using riftrunner

if you ask what the cutoff date is, it'll tell you that it doesn't have one as the information gets continuously updated

that seems to do it