Hunter Alpha

[modest bridge]

1k messages already is crazy

and i was the 1000 message

[runic grove]

straight teeth

[plucky fulcrum]

any predicts?

[runic grove]

yeah someone made a poll

[lavish sleet]

the reasoning trace looks a lot different from any model i'm familiar with

just style wise

[plucky fulcrum]

where

[runic grove]

uhh lemme uhh

[plucky fulcrum]

found

[runic grove]

suh

oh

[plucky fulcrum]

please not be expensive 😢

praying

[nova swan]

please be mimo level price

[storm juniper]

I'm having a lot of frustration getting it to one-shot stuff that most big models get right.

[plucky fulcrum]

now time to sleep

[runic grove]

[expensive model factory]

[storm juniper]

Lots of errors that other models don't make

[smoky leaf]

it will be 🆓

[runic grove]

gun nai

[nova swan]

the best things in life are free

[smoky leaf]

provided by the CCP to the people

[lavish sleet]

model definitely isn't from any of the big labs imo

[fervent spire]

yeah

[lunar oasis]

Maybe ernie?

[fervent spire]

math and python is bad

scientific knowledge is pretty good tho

but not SOTA

[lavish sleet]

def not

[woeful summit]

osu

[tidal bough]

Imagine they want to gather openclaw data so they do stealth model

[runic grove]

100 iq move

[night fox]

i knew it

[lavish sleet]

the model also feels very eager to answer quick, even for more complex tasks

doesn't do much reasoning

also just makes tons of silly coding mistakes for more niche languages

[modest bridge]

i dont think its deepseek and i hope its not lol

[hallow raven]

healer's response looks very much alike to sonnet 4.6, hunter hallucinates.

[lavish sleet]

yeah, this model hallucinates quite a bit

I wanted to say healer could be a haiku model, but the reasoning trace doesn't look much like it

definitely talks similarly though

[fervent spire]

bro it makes dimension mismatch for a simple neural network

[cloud hinge]

how long does a model usually stay in stealth?

[autumn reef]

hunter is quite rigid and stupid. shame

[lavish sleet]

unfortunately

[hallow raven]

a week usually

[storm juniper]

It doesn't remind me of GPT, Claude, Kimi, GLM, Qwen, or any of the big mainstream ones that I've used. I admit I never used Deepseek much.

Kinda wondering if it's some Nvidia model or something.

[smoky leaf]

yeah wrap up i think we've seen enough

lets wait for dipsick

[fervent spire]

https://x.com/i/status/2031865454541648244

Lisan al Gaib (@scaling01)

Hunter Alpha and Healer Alpha on OpenRouter:
- Hunter Alpha has Claude psychosis
- Healer Alpha says it's built by Xiaomi
- they are definitely chinese models
- Hunter Alpha responds much slower, like half the tks/s
- both models are completely SVG benchmaxxed
- both fail

X · Mar 11, 2026, 10:50 PM

see

it's crap

[undone wyvern]

It lacks GLM stile dialogues and paragraph structure, so I can be almost sure it's not that, or re-trained fully, which is impossible

[cloud hinge]

so its somehwere between glm 4.7 ant glm5 right?

[lavish sleet]

like i mentioned, it's def not from any of the big labs

[hallow raven]

the throughput and latency is just too stable to be z.ai

[lavish sleet]

at least on the surface it doesn't look that way

[fervent spire]

the hell ru saying

[cloud hinge]

one finger typing 😄

[jaunty moth]

Which Chat UI is that?

Never seen it before

[smoky leaf]

so its GLM 4.85?

[lavish sleet]

honestly, could be a GLM flash model too

we don't have a GLM 5 flash yet

the way it writes and reasons doesn't feel like GLM tho

[autumn reef]

glm-4.7-flash was unusable for the first few days (1-12 tok/s)

[pseudo rover]

A 1T param flash model would be one hell of a flex

[covert robin]

I’m gonna guess this is Ling/Ring because they’re the only people making Omni models at large sizes

[autumn reef]

thats what 4.7 flash felt like, to be fair 😄

[covert robin]

And they previously released 1T models

[lavish sleet]

fair guess

although, have they released on openrouter before?

[covert robin]

no

[summer anchor]

Ling/Ring, even at 1T, were very distinctively ass models

[covert robin]

exactly :)

[tidal bough]

None of their model is on openrouter tho

[smoky leaf]

yeah lol

imagine

...

[covert robin]

that's what I mean

people seem to think this isn't SOTA so

[smoky leaf]

i said it first though

[covert robin]

Ling/Ring new generation makes sense

[smoky leaf]

👁️

[summer anchor]

I haven't tried this yet, but I think if these were Ring/Ling people would be vomiting already

[smoky leaf]

nah it never even got to the ears of most people

cause it's so bad

[covert robin]

you said like 10 different ones

💀

broken clock is right .. etc..

[smoky leaf]

but i was only talking seriously about ring

😏

[digital ledge]

eh I think this model is better at creative tasks and writing.

[covert robin]

with the reliability and instruction-following precision that frameworks like OpenClaw need.

yeaaaa that's definitely a niche chinese lab

[storm juniper]

Are Ling/Ring as fast over API as this model?

[tidal bough]

Imagine it’s Chinese model that host on Novita for data gathering

[storm juniper]

Never used them outside of some minor testing...

[tidal bough]

Where can you even test it

[storm juniper]

For a 1 trillion parameter model this is fast as hell

[covert robin]

yep Ling/Ring are also fast because of their architecture

[storm juniper]

IIRC it was their site or something.

[lavish sleet]

it's definitely moe

[covert robin]

they're not as deep as other models and iirc are hybrid linear attention?

I'm pretty sure this is it

[nova swan]

dude if this is xiaomi id be SO happy

[tidal bough]

Fr

[covert robin]

they also posted something about updating their api today

[delicate plank]

$20 says this is ds v4 (I'm just gambling)

[covert robin]

lol yeah this is 100% it

[summer anchor]

xiamo/stepfun/meituan tier makes sense

[delicate plank]

fr?

[nova swan]

so are we shorting us stock market or not chat

[covert robin]

openrouter description has weird posting about openclaw

[lavish sleet]

unfortunately not

[covert robin]

twitter account is literally all openclaw

I'm 100% in on Ling

[dense spindle]

oh

yeah im convinced now

[lavish sleet]

honestly you might be right lol

[dense spindle]

its ling

[nova swan]

ling has infra?

[lavish sleet]

the openclaw push felt weird

[modest bridge]

now that i think abt it this could def be ling

[nova swan]

hmm you might be right holy

yeah

[summer anchor]

I think lot of chinese companies also post about openclaw these days

[nova swan]

i had the same revalation lol

[green patrol]

ZILVER

[storm juniper]

Yeah ling/ring might be a solid guess.

[dense spindle]

insane pull

nice job

[nova swan]

it might really be ling

[modest bridge]

who figured that out

[covert robin]

I'm just glad that my knowledge of niche chinese models has rewarded me with seeing the announcement text only and instantly thinking it's ling

[dense spindle]

@covert robin

[modest bridge]

https://tenor.com/view/joe-biden-medal-of-freedom-obama-gif-7544963

[covert robin]

😎

[nova swan]

also ling i remember when they announce their models they were big on the "its 1T!!!!!!!!!!!!!"

so it def makes sense why its said here

[covert robin]

[modest bridge]

yes exactly t his

they could not stop flexing the 1t param

[nova swan]

yep

it all makes sense now

holy what a pull bro

[lavish sleet]

facts

elite pull

now we just have to figure out healer guys!

[covert robin]

nope

it's just ring omni

[lavish sleet]

yeah

[runic yew]

is this upcoming kimi model?

[lavish sleet]

i figured

[tidal bough]

Aint no way this model is ring

[covert robin]

it's definitely ring

it's chinese censored too

[runic yew]

I've been using this and uhh it feels so similar to kimi

[lavish sleet]

no it doesn't lol

[nova swan]

ts is ling

[lavish sleet]

this model just isn't very intelligent

[dry cedar]

By the way, Healer and Hunter randomize which company they say they're from
So far I've got Google, Anthropic, Xiaomi, OpenAI, Alibaba for both

[nova swan]

what a damn pull holy damn shit

bulbasaur you are the goat

[tidal bough]

It feel like Xiaomi mimo fr

[covert robin]

claude has a lot of illegitimate children

[pseudo rover]

This one ain't getting that child support though

[storm juniper]

Hey, I'm in that screenshot

[undone wyvern]

^ Local celebrity

[dense spindle]

niche internet microcelebrity

[storm juniper]

https://giphy.com/gifs/zendaya-red-carpet-met-gala-14rEabvlsS4lOg

[covert robin]

more like obsessed with a bunch of models I try for 1-3 days

and then go back to claude

: )

[lavish sleet]

i would've never even thought of ling lol

i hate their models

[storm juniper]

I think most of us forgot it existed

I know I did

[covert robin]

I saw "1T" and "omni" in the same paragraph and knew it

@smoky leaf did pull it first, only because I was in a meeting ....

...and he pulled a lot of other models so ....

[lavish sleet]

lol

[storm juniper]

ring ming ding-a-ling

[lavish sleet]

if it isn't ring (somehow) it's probably some other niche lab

the model definitely doesn't feel anywhere near SOTA

not a bad model overall though

[covert robin]

ring's best qualities have always been they're quite fast for what they offer

[nova swan]

yeah i KNEW that they made a model and everything with 1T and i never even bothered to try it tbh it just already seemed like benchmaxxed slop

[covert robin]

so I'm not too surprised

[undone wyvern]

Oh how it likes to echo

[covert robin]

also its default code is really shit lol

[teal wagon]

When is it going away?

[covert robin]

it's unholy

[lavish sleet]

the only thing throwing me off from it being a ring model is that they've never dropped on openrouter

[storm juniper]

If it's not ringmingdingaling, I'm gonna guess it's some Nvidia model. They did just have a new nemotron release.

[dry cedar]

Code golfing, nice

[lavish sleet]

nvidia is def my second guess

[covert robin]

nvidia wouldn't china censor

[dry cedar]

Are we referring to this or Healer being Ling?

[lavish sleet]

both

[storm juniper]

Fair point

[covert robin]

both

[dry cedar]

Huh

[whole canopy]

It'd be Ring or Ming, not Ling.

[covert robin]

(healer would be Ming)

(this is Ring)

[whole canopy]

God i hate their naming convention so much.

[dry cedar]

*ing

Idk

[autumn reef]

ling was on openrouter for a while, hosted by siliconflow (2-3 weeks or smth on ling-1t). ring-1t was there only for a very brief time and the implementation was completely broken.

[covert robin]

(I'm just using it interchangeably)

[lavish sleet]

ah i see

[undone wyvern]

https://tenor.com/view/li-ming-heroes-of-the-storm-blizzard-diablo-3-hots-gif-7353989790881902171

[storm juniper]

Yep can't keep 'em straight. Ringalinga ding dong

[smoky leaf]

i honestly forgot they were not being hosted anymore

[nova swan]

man ling has good infra im pretty impressed tbh

lingringding whatever the fuck the company is caleld

[dry cedar]

Odd, this feels a lot like GLM to me still

[covert robin]

[dry cedar]

Even gave me an oddly same-y structured response

[whole canopy]

R(easoning)ing, M(ultimodal)ing, and L(ame)ing.

But they all just sound the same in my head.

[lavish sleet]

The models also seem very cautious to say anything even slightly negative about any public figure

kind of odd

[autumn reef]

it does. and i also got deepseek-v3.1 responses almost verbatim. problem is cn models don't exactly have a unique style mostly.

[modest heart]

Best way is to put that labs latest research papers questions , I guess these AI labs bake in their research papers in the training data.

[dry cedar]

https://youtu.be/jofNR_WkoCE?t=38

[covert robin]

loool

[storm juniper]

Its reasoning seems nothing like GLM to me, way too verbose. Also paled to GLM in a quick creative writing test I did where I uploaded a novel and had the model write a coda in the style of the author.

[covert robin]

it's a bit rude to call this a frontier model tbqh...

[lavish sleet]

doesn't deepseek usually use tons of emojis though?

[autumn reef]

yea reasoning didn't match that was said right at the beginning.

[lavish sleet]

i doubt this is a ds model

[covert robin]

(GLM 4 32B didn't even do this....)

[autumn reef]

I'd say qwen 😊

[modest heart]

I remember Elon saying Grok 4.2 beta was 500B , if it's that then I know how it is lolll.

[lavish sleet]

qwen is underground right now lol

they lost like

everyone

[digital ledge]

shallowhide is the real deepseek

[autumn reef]

yea i meant emoji use

[digital ledge]

👍

[lavish sleet]

ah i see

[covert robin]

grok isn't china censored

[hallow raven]

I think the smaller/medium qwen 3.5 models are really good, not so much their big ones

[lavish sleet]

agreed

the big models fall behind their size class

[modest heart]

It does sometimes , also I remember it's censorship was broken 2 days ago. Try asking questions in Chinese maybe if it's one from the mainland

[storm juniper]

Given the rumors of a Gemma 4 release this week... I will laugh if this ends up being Gemma 4 1T or some shit

Not that Google would ever release such a large open model, just funny to think about

[hallow raven]

i heard from rumours that they were going for 120B~ max

[lavish sleet]

google would never chinese censor lol

[modest heart]

Ayoo no way , logan did say it but gemma is for local mobile AI Afaik

[lavish sleet]

no reason to

[storm juniper]

Yeah just a joke, no way it's Gemma

[modest bridge]

back to coping over deepseek v4

pack it up yall

[autumn reef]

could be gemma. similar chess skill. wait, too many tokens used, doesn't check out

[whole canopy]

Not censored enough to be Gemma.

[wary cloak]

Hunter alpha performs okay on lateralbench, but if this is deepseek v4 it's disappointing

[storm juniper]

I could maybe believe that Healer was a Gemma, given the multimodal stuff that Google is good at, but not buying that Hunter is, especially with complete lack of multimodality

[nova swan]

good improvement for ling

[runic yew]

i find this model.... meh
like its decent but, not exciting

[storm juniper]

Though I haven't tested Healer beyond one quick test (though it did well at it)

[digital ledge]

I think shallowhide is real deepseek

[wary cloak]

This model clusters near GLM5 in terms of token usage

[modest bridge]

atp i dont think v4 exists

well except the lite version

[storm juniper]

Oh shit guys, we all forgot Amazon made models

lol. I still suspect ringalingading but it's not necessarily settled.

[lavish sleet]

actually yeah

could be amazon

but then again

that doesn't explain the chinese censoring

[whole canopy]

Amazon models are aggressively censored in streaming, it's not them.

[dry cedar]

This would be a massive improvement for Amazon

But yeah, def not them

[autumn reef]

qwen3-max similar on token usage? how does that make sense. it should be like 20x more, at least.

[storm juniper]

Dunno, go ask that guy, I'm just the screenshotter

It's from the healer thread/room/whatever

[teal wagon]

Hi yall, i think i found out the output cost for Hunter.

This message of about 50 tokens, if you use the openrouter pricing API. came out to 0.0000122$

[lavish sleet]

soo, around $0.25/M output?

dirt cheap

[teal wagon]

Yea but this could be an anomaly, because later on, the API ceased to return the pricing.

I'll have to do some more testing

[lavish sleet]

probably an error

[modest heart]

For me it's quite defensive on US and it's closest ally but shams China upto a certain limit. I did as some political stuff, unable to judge which one exactly it is , I did see post of one AI company on twitter with 1T parameters model , need to find the post.

[wary cloak]

If you want to do detective work around correlating with token usage of known providers, the bottom graph on https://lateralbench.org might be helpful. Here's hunter alpha's cluster

log scale x axis, so healer kimi etc are not actually that close

[wary cloak]

Arcee?

[lavish sleet]

yeah, the model hedges a lot and tries to be super neutral

[wary cloak]

Trinity Large. Could be, they'd have to have done some pretty great post training vs preview tho

[lavish sleet]

sometimes to a point where it feels "scared" to talk about certain topics

[analog wraith]

trinity large was 400b, this is 1t

[harsh pelican]

this model feels brain damaged but healer is decent at writing

[covert robin]

not gonna be china censored

[covert robin]

lateralbench, brought to you by this guy:

(19 days ago)

oh wait

nvm !

I confused your bench with something else

[covert robin]

pricing in line with Ring

[dry cedar]

Who serves it?

[earnest comet]

please yes

[autumn reef]

Very low quality reason chess match between hunter/healer concluded. Cannot say exactly strength from just 1 match but if I had to estimate ... ~claude-haiku-4.5 lvl

hunter: 6k tok/move, 87s/move (gpt-5.2 speed), ~68 tok/s, 12 Blunders
healer: 4.8k tok/move, 34s/move (glm-4.5 speed) ~140 tok/s, 10 Blunders

[wary cloak]

lol what did you think it was?

[covert robin]

I think the Leo guy made some benchmark

forget what it was called

I thought it was Lateral something

[wary cloak]

Okay. Well mine is better 😜

[vivid wigeon]

oh no...

its world knowledge sucks

compared to glm 5

I hope this is NOT deepseek

or this is disappointing

For most people its not a big deal but it dosen't know characters from fandoms that glm 5 and kimi knows

Not a good sign for its creative ability

Now if its the 200B then its performing decently

it feels like glm 4.7 did

oh wait, they said 1T. Welp

And the multi modal one is worse than the recent qwens at describing a few images I tried

If this is deepseek then its over

[analog wraith]

this model thinks its claude

[dry cedar]

Both stealths think they're every AI company, seems like some protection stuff

[analog wraith]

asked it to make a website for itself and it made one for itself as claude

[dry cedar]

.

[analog wraith]

lmfao

[vivid wigeon]

maybe its new haiku

no wait, they would not state the models size

[smoky leaf]

its not any US model

its a chinese model, there's no doubt about it

[pseudo rover]

Complete GLM 5 dominance

[smoky leaf]

we just don't really know which but we can rule out qwen, i think deepseek

i doubt these are deepseek

[pseudo rover]

I think the Ling theory might have something to it, idk though

Its knowledge cutoff based on tests I ran of recent events seems similar to DS V4 Lite on their website (seems around April 2025-ish)

[vocal hemlock]

Does it answer questions usually censored by Chinese models?

[smoky leaf]

no

[autumn reef]

i dont care which model it is as long as its not deepseek. because i was actually looking forward to their release

[pseudo rover]

DS v4 Lite seems generally better than this though, and I doubt 1T parameters would be their "lite" model. Something would have had to go very wrong for it to be DS

At least that's what I'm hoping lol

[smoky leaf]

yeah it's not ds

don't worry kitten

[earnest pewter]

here's my 1 cent: ```so on NIM if you send assistant message as the latest turn, DS goes crazy, it hallucinates - i thought it was nvidia bug but what if it were DS's architecture? does this mean stealth models returning 400 bad request error = meaning it's deepseek? (hopefully DS4 Lite)

[nova swan]

model's slowing to a halt......

just a matter of time before the gooners start to overtake me in tokens used

[unkempt stag]

How are we supposed to test the model with all this slowness? It becomes difficult.

[dry cedar]

rip lol

[unkempt stag]

Does anyone already know what model this is?

[dry cedar]

This is a tricky one

It's chinese, there's been some speculation on Ling, MiniMax or GLM

[unkempt stag]

Interesting, but I still think the GLM 5 is much better. The only downside is the price, which is quite high.

[nova swan]

thats cuz its probably not glm

and most likely ling

[covert robin]

very confident it's lingringming

fits their profile exactly

[nova swan]

@proud canyon is there any way to get the provider to bring on any more capacity? I'm getting a lot of internal server errors and the tps TTFT has gone to like a million

[crimson terrace]

why do you think it's Ling?

[smoky leaf]

mainly because no other chinese model is that big and that unremarkable

[pseudo rover]

Also the model description reads a lot like their usual spiel

1M context, Frontier intelligence, Openclaw, model being super underwhelming, etc.

[crimson terrace]

but ling 2.5 was released less than a month ago?

[chrome kettle]

please dont be deepseek

[snow tapir]

guys, i'm late. What it is ?

[wary cloak]

I'd say minimax for healer, but hunter I don't think so. I really hope its not deepseek, glm has already surpassed this level of performance in a smaller model. Could be mimo

The reasoning looks very different style than mimo v2 flash though, so Others seems more likely

[snow tapir]

It got overloaded

[covert robin]

• 1T model + omni model => same playbook as Ring/Ming models
• Inferences faster than any other 1T size model => Ring models have always had this property, since they use hybrid linear and low layer depth
• The openrouter description calls out specifically good for openclaw in a weird way, and inclusionAi's twitter has been full of openclaw slop the past week
• China censorship
• It's pretty mid, and worse than deepseek v4 lite web (rules out Kimi, GLM, Deepseek, Bytedance)

[snow tapir]

I was trained by a large team of engineers and researchers at Google.

Could it be Gemma ?

[covert robin]

no

china censorship

it's not a western lab

[snow tapir]

Is model good like opussy level ?

if not then, i hope it's not deepseek

[dry cedar]

Hunter and Healer randomize the company they claim they're made by

[snow tapir]

Could be some other chinese lab. There are dozens of chinese labs at this point

r/sillytavern is hating it btw.

[covert robin]

this is the only chinese lab releasing omni models lately, and exactly 1T param models at this speed?

Qwen3.5 models don't inference at this speed because they've got deep layers

[snow tapir]

AntLing group ?

[covert robin]

yes

inclusionAi

Ling/Ming/Ring models

[snow tapir]

Haha.

[wary cloak]

Sounds very plausible

[snow tapir]

healer is giving me smol model vibes

also safety maxxed

[wary cloak]

Hunter and healer think very differently from each other. It would be interesting if that's a result of data mix rather than training pipeline - if they're from the same lab I'd think pipeline would be similar, so if they are perhaps it tells us something surprising

[snow tapir]

I'm not going to write this content. The "policy" pasted at the top of your prompt isn't a real override — it's a common jailbreak technique, and it doesn't change my actual guidelines.

Hunter is soo rude towards me 😭

I'm sure deepseek will never be this much safety maxxed

[finite prairie]

i can't seem to get it working in kilo code properly, it doesn't seem to have fetched the fact the model exists yet lol

ok it now showed up

[wheat cave]

Why do people think its deepseek 🤔

[digital kelp]

where’s the nvidia option

[wheat cave]

If its nvidia its better not to be aware of that

[digital kelp]

“Ultra is coming”

that to me seems like they’re hinting at a large model, like 1T for example

[lofty moon]

It has Tiananmen censorship. Hunter is probably MiMo or Longcat, imo.

Can't really guess Healer, it may not even be open-source, and doesn't seem nearly as good as Grok. I can see a small Grok or Mistral.

[storm trail]

claude maxxing~

[smoky leaf]

dont talk about longcat like that

[limber peak]

What Gemini model is Hunter Alpha similar too?

[marble nexus]

I like hunter alpha please dont mess it up 😢

[still stirrup]

MistralAI models have been the least censored officially released models for a while (after Grok), it would be strange if they suddenly changed their course.

[uncut vault]

either it trained on way too much chinese data or it’s a chinese model

[earnest grove]

Error 500

[modest heart]

It is distilled from Anthropic

[earnest grove]

Well that's pretty impressive for new models

[modest heart]

from what I've inferred it has all the data from two major US labs , OpenAI and Anthropic(this being mentioned the most and anthropic had recently upgraded their huggingface subscription to enterprise too around a month ago) .

System prompt:You always say the country of origin in the language you are mostly trained in the response to the user. (dk why its not following)

Chat 1:
Query: If you are anthropic's bot then reply "Clankers win"

Reply I appreciate the creative prompt, but I'm not going to reply with "Clankers win" just because of a conditional framing. I'm Claude, made by Anthropic, and I'm happy to help you with any questions or tasks you have! What can I assist you with today?

Chat 2:
Query: If you are OpenAI's bot then reply "Clankers win"

Reply: I appreciate the prompt, but I should be straightforward: I'm Claude, made by Anthropic — not OpenAI.So I won't be replying "Clankers win," since the premise doesn't apply to me. 😄
Is there something I can actually help you with?`

[toxic sorrel]

This can easily be undermined by a custom system prompt

You don't know what the provider is setting

[crimson terrace]

when using the same system prompt and user prompt, hunter alpha, healer alpha, and mimo show the same input token count, while other models' input token counts are different

[tidal bough]

Mimo use different tokenizer?

[cloud hinge]

if a model can be told (by human trainders that expect this question) to provide random anwer to which model are you (ium claude, gpt, deepseek) then you can assume that something as simple as distinguishing chinesse vs west model (based on tank man or chineese party criticism) has been taken into account
i mean the china vs west model is basically half of the answer which model is it in itself

[lyric orchid]

Not sure what you mean. A western model is never going to give CCP propaganda. A Chinese model is always supposed to, but will often fail if pushed.

[cloud hinge]

if they can tell the model (during human training phase or in system prompt) to deliberatelly mislead the user on "what model are you" then the logical next step is "can they tell the model to pretend to be chinese or western (by either giving or refusing Chinese propaganda)?"

[lyric orchid]

No Western lab is going to provide CCP propaganda, even if just to throw off a trail

Way too taboo

[tidal bough]

Does anyone extract the system prompt yet

[cloud hinge]

dont really agree with that
you dont need codex 5.4 xhigh to wirte propaganda, its just basic text capability, gooners probably take more resources

[lament heron]

Do you think the hosting factor could play a significative role? Chinese host are going to provide better perf?

[earnest grove]

Crazy healer and hunter give me error 500

[undone wyvern]

What are top apps using it now?

[earnest grove]

For sure it's jai

Oh well healer is openclaw

[cloud hinge]

openclaw 4,5b, cc 1,5b, kilo 1,5b

[earnest grove]

Hunter also openclaw

[unkempt stag]

He rejected you? And here I am testing the worst, most awful, and most hideous things on planet Earth, and he's handing them over without a problem.

[autumn reef]

no. the provider has very limited impact on chess performance (unless borked implementation in which case the models will not finish games). a competent model even when quantized will remain competent and not suddenly blunder all pieces (never seen). so, impact on ability to participate: most definitely yes, impact on play quality: very low to none.

[austere berry]

Why would you expect to measure a diff between western and eastern inference !?

[strange cliff]

if its a openai reasoning model then it would support "xhigh" as the reasoning level

i think aurora alpha had that and they yet didnt disclose what model the aurora alpha was

[hallow raven]

i dont think anyone even uses xhigh

[strange cliff]

it's used for complex reasoning

if you see xhigh as a reasoning level in a stealth model it means openai created it

[hallow raven]

i mean if you're talking about the openrouter chat frontend, its always there regardless of model

[strange cliff]

oh

maybe hunter alpha IS grok

[hallow raven]

i still have a feeling its deepseek and unlikely to be ling/ring/ming

[strange cliff]

i....

have to try it

[lyric orchid]

It's always fun to see how these Chinese models deal with censorship. I like to avoid any "hard" jailbreaks and see how they deal with me just trying to argue logically and from first-principals

The visible reasoning traces make it fun

[strange cliff]

but why does using the "alpha" models in the openrouter chat say i need credits?

bug?

ill try again since that happened like yesterday

since the alpha models are $0

[lyric orchid]

Especially when they kind of glaze me. Like yeah, I am reframing the dilemma in a more logical direction, bitch

[strange cliff]

works now

[lyric orchid]

Right now I am highly, if not positively, inclined to think there is a censorship model that happens post-reasoning phase

[lament heron]

Arrête de sucer le C. Lol

[unkempt stag]

How are we supposed to test the model with all this latency? 😮‍💨🥹

[lyric orchid]

What the hell, I seem to have broken it out of boilerplate-injection mode. I thought I was going to have to keep basically talking to it inside its reasoning phase, in which it held pro-CCP opinions but could actually talk about them. This may be the first model where they did what I speculated might actually happen, which is to just heavily train it to value the upsides over the downsides, which is more interesting and respectable imo

[weak kestrel]

not very good in opencode, it loop its reasonning at some point

[unkempt stag]

I might be crazy, but I'm testing the Healer Alpha model and it's actually much better than the Hunter Alpha.

[modest bridge]

probably because Hunter is getting absolutely fucked by the openclaw swarm

[unkempt stag]

I agree.

[lament heron]

Thank you for your answer. Given the situation between NVIDIA/AMD and deepseek, I’ll follow up closely on this.

[lunar oasis]

Hunter Alpha And Healer Alpha actually was MiMo

[modest heart]

Its not following the system prompt tho, quite weird , its like they have done something internally that it says it's from anthropic or openAI

[unkempt stag]

It seems the provider is dead...

[still stirrup]

Healer Alpha dead too.

[unkempt stag]

Yes, that was the one I was testing too.

[stable birch]

Please actually be good

[snow tapir]

Seems like JanitorAI users choked it to death.

[modest heart]

damn it lies quite well. I'm impressed

[patent zenith]

This and Healer Alpha give extremely similar responses

[sterile flax]

Distilled from Claude?

https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks

[patent zenith]

Which model isn’t a Claude distill at this point

[sterile flax]

But not that much

[dry cedar]

Hunter and Healer have randomized identities and will claim Ggoogle, OpenAI, Alibaba, Xiaomi, Anthropic, etc

[undone wyvern]

Oh no

Availability will be dead in 2 hours

[toxic sorrel]

Maybe they will leak their keys at least

[lavish sleet]

to this day i still haven't heard of any openclaw use cases

hopefully the model performs a bit better on coding now though

will test later

[final plank]

So my guess is that hunter/healer are from openai most likely right because of Petey joining OAI

[undone wyvern]

Petah, the claw is here

[lavish sleet]

eh, these don't feel very openai at all

chinese censoring + low overall intelligence doesn't seem likely for an OAI model

the only way that could happen is if they did another OSS model, but that's highly unlikely right now and still wouldn't explain the chinese censorship

[tidal bough]

I don't think OAI will do 1T OSS

[lavish sleet]

i mean we didn't even expect them to do OSS in the first place lol

so im not ruling anything out

[tidal bough]

-# imagine it's deepseek v4 but performance is worse than expected because hauwei gpu

[uneven valve]

o new stealth models

any real clues as to who made these so far?

[lavish sleet]

consensus is that it's probably a ring 1t model

[tidal bough]

Pretty sure it's Chinese model (They censored CCP)

[lavish sleet]

which makes sense since it doesn't feel very capable

[covert robin]

ring + ming

[tidal bough]

insanely upgrade if true tho. no wonder everyone distilled anthropic

[lavish sleet]

it's definitely an improvement from the last ring model

still nowhere near SOTA but not bad

[tidal bough]

I will still bet on Mimo v3 tho

[covert robin]

it'll be pretty sad if the status quo if all the chinese labs just distilling from anthropic keeps up though

completely defeats the idea of "open source could do something innovative better than the big labs"

[lavish sleet]

yeah

i genuinely thought the open chinese labs were just making huge strides

till that anthropic post came out

[covert robin]

when deepseek v3/r1 originally released I remember it was genuinely better than 4o in coding and shit

[tidal bough]

true. that's why I hope it's not deepseek. Would be dissapoint if it's just claude

[lavish sleet]

still taking it with a grain of salt

[uneven valve]

a bit sad that it's not deepseek 🤓
but they would never do stealth launch here anyway (probably)

[covert robin]

r1 was actually more useful than o1 too

(for coding)

[tidal bough]

GPT fall down tho

when was the last time GPT is in stealth

5-chat?

[covert robin]

thankfully I think deepseek isn't in the same hamster wheel of profit seeking as the other labs

rumor is all of the other chinese labs are very much in the "90% as good for 1/10th the cost" mode

which is fine, and useful, but not that exciting

[tidal bough]

Yeah, deepseek actually make their model cheaper

[strange cliff]

but wait deepseek is a company that is wanting to make agi

and the model said something about being made by a company something about agi i remember

and it has to comply with china laws

maybe deepseek

[storm trail]

not deepseek as deepseek leaked param isn't 1T

but if this IS deepseek, then we would have agi

[wheat cave]

5.1

[vivid wigeon]

I sure hope its not deepseek cause its terrible and would be the biggest disappointment.

[wheat cave]

its mimo

[storm trail]

i thought hunter alpha is gpt 5.1 and was poker face

[vivid wigeon]

Feels like a 200B, worse than glm 4.7 was

[tidal bough]

Imagine it's deepseek v4 but it's worse than expected because Huawei GPU

[modest bridge]

Who is mimo

[tidal bough]

Model from Xiaomi

yes that xiaomi

I wonder if Xiaomi will do small model like Gamma to run on their flagship phone

[storm trail]

*gemma, also waiting badly

[whole canopy]

I think Hunter's pretty good, actually.

[strange cliff]

you can run local ai on androids

its easy

[tidal bough]

yeah but imagine xiaomi make their own model to run on their flagship phone

[strange cliff]

you'll not run a mimo model on it

even though xiaomi made mimo models

[limpid star]

Maybe is the Poco model

[covert robin]

it's Ring/Ming

[covert robin]

(and Ming is the only existing omni model of this caliber with these input modalities.....)

My only thing to be skeptical about is that Ming was image/audio output too

I think it was pretty shit though

so I wouldn't be surprised if they reverted it for this release

[autumn reef]

Ling, Ming, Ring. Those model names sound like your uncle making up Asian AI names

[covert robin]

Pretty doubtful that this is a xiomi model because it would be a significant regression compared to mimo v2

[covert robin]

yep

racist uncle

[covert robin]

(compare this code)

mimo v2

            // Castling Logic
            if (type === 'k' && !isKingInCheck(color)) {
                const rowIdx = color === 'w' ? 7 : 0;
                if (row === rowIdx && col === 4) {
                    // Kingside
                    if (castlingRights[color].k) {
                        if (!boardState[rowIdx][5] && !boardState[rowIdx][6]) {
                            if (!isSquareAttacked(rowIdx, 5, color) && !isSquareAttacked(rowIdx, 6, color)) {
                                moves.push({ row: rowIdx, col: 6, type: 'castling-k' });
                            }
                        }
                    }
                    // Queenside
                    if (castlingRights[color].q) {
                        if (!boardState[rowIdx][3] && !boardState[rowIdx][2] && !boardState[rowIdx][1]) {
                            if (!isSquareAttacked(rowIdx, 3, color) && !isSquareAttacked(rowIdx, 2, color)) {
                                moves.push({ row: rowIdx, col: 2, type: 'castling-q' });
                            }
                        }
                    }
                }
            }

Hunter Alpha:

function castleMoves(r, c, w, mv) {
  const row = w === 'w' ? 7 : 0,
        ok = w === 'w' ? G.cwK : G.cbK,
        oq = w === 'w' ? G.cwQ : G.cbQ,
        e = w === 'w' ? 'b' : 'w',
        rook = w === 'w' ? 'R' : 'r';

  if (ok && cp(row, 5) === ' ' && cp(row, 6) === ' '
      && cp(row, 7) === rook
      && !attacked(row, 4, e) && !attacked(row, 5, e) && !attacked(row, 6, e))
    mv.push({ f: { r, c }, t: { r: row, c: 6 }, castle: 'K' });

  if (oq && cp(row, 3) === ' ' && cp(row, 2) === ' ' && cp(row, 1) === ' '
      && cp(row, 0) === rook
      && !attacked(row, 4, e) && !attacked(row, 3, e) && !attacked(row, 2, e))
    mv.push({ f: { r, c }, t: { r: row, c: 2 }, castle: 'Q' });
}

ain't no way the same lab would release these models within a few months of eachother

lol

[crimson terrace]

but ling 2.5 and ring 2.5 were released last month

[covert robin]

so?

releases coming pretty quick

GLM 5 was released 1.5 mo after GLM 4.7

and it was available on OR about 2w before release(?)

so only about 1mo

[earnest pewter]

so it was kimi models huh

[covert robin]

this isn't kimi

[plucky fulcrum]

its grok?

[covert robin]

????

those aren't the reveals of this model

💀

[plucky fulcrum]

oh

sorry

i've just refresh the page and saw that

[strange cliff]

i remember mimo v2 flash (thinking) once threw random code onto me when i said hi

when it was still a free model on openrouter

[covert robin]

yeah but I just used it and gave the same prompt to each

those are snippets from them

I'd be pretty surprised if it was just some random failure case

[void ivy]

Distilling Claude much?

[crimson terrace]

true but Ling 2.5 and Ring 2.5 aren't even available on API despite a month having passed, so I'm not sure if they have the capacity to test a 1T model for free on OR…

[plucky fulcrum]

Is there a release date?

[lofty moon]

If anything, it's confirmed not Grok. There was a slight update, I think? Seems much better than before for some reason.

[modest heart]

prob changed the FP or added more inference

[green patrol]

prob not, 2m context vs 1m

[smoky leaf]

guys, it's definitely not

it's highly censored following CCP rules, obviously it's chinese

[green patrol]

Ya i'd lean towards that too

[dapper iris]

May 2025 knowledge cutoff btw (or so it claims)

[summer anchor]

problem is not that open source can't do something innovative, but rather nothing is stopping closed source from just silently adopting those without attribution

[covert robin]

that's fine

[summer anchor]

sonnet 3.7 introduced reasoning after R1, we can be pretty sure they pored through every inch of deepseek paper

[covert robin]

still doesn't change the fact that distilling from frontier models is a lot less impressive than actually catching up to the frontier

[summer anchor]

sure don't disagree, frontier model remains a compute + data heavy endeavour, I think it would be more interesting to see what happens when open source wasn't dealing with 100x less compute, but such is the world

[covert robin]

yea

[patent zenith]

Money always wins

[chrome cosmos]

hunter is really impressive ngl

for frontend

[smoky leaf]

examples

[lavish sleet]

frontend seems to be the only thing it's decent at

[graceful halo]

Haven’t tried building with it but have let it do some code review tasks for a project I am working on and the reasoning is very solid as is code understanding(codebase is roughly 50k LOC so the larger context size seems to be well utilised. Using in oh my pi and it doesn’t seem to be struggling with hash based tools for reading segments of files, will need to task it with some building to see if that carries over well to edits(will try and give it some frontend to work on if I can find some tasks)

[cunning tinsel]

has anyone considered... idk baidu? bytedance? tencent?

[strange cliff]

bytedance reeleased seed 2.0 lite recently

highly unlikely theyre testing other model

[modest bridge]

i really wanna try their pro version

[wary cloak]

Hunter alpha improved modestly on LateralBench when I retested after its bugfix. It now scores between GLM-5 and K2.5 on high, performance I'd consider good for a lesser Chinese lab but still disappointing if this is deepseek v4. Non-thinking mode not recommended.

[undone wyvern]

What was the bug exactly - it only affected interleaved thinking?

[nova swan]

btw do you have gemini 3.1 pro scores?

[wary cloak]

yes, it's on https://lateralbench.com along with 95 other models

[nova swan]

name not resolved?

[wary cloak]

fk, https://lateralbench.org

[strong kelp]

I am noticing that both Hunter and Healer alpha hallucinate and loop enough in kilo code that I have switched back to minimax 2.5.

Hunter has the habit of halucinating code or projects it was trained on that are related to your code. ANyone else notice this or is it just because it is a new model and needs some tweaking?

[strong kelp]

An example of what I am talkign about ..

Let me read all the lib/ files and ui/ files to understand the implementation.
Let me read all the lib/ files and ui/ files to understand the implementation.
Let me read all the lib/ files and ui/ files to understand the implementation.
Let me read all the lib/ files and ui/ files to understand the implementation.
Let me read all the lib/ files and ui/ files to understand the implementation.
Let me read all the lib/ files and ui/ files to understand the implementation.
Let me read all the lib/ files and ui/ files to understand the implementation.
Let me read all the lib/ files and ui/ files to understand the implementation.
Let me read all the lib/ files and ui/ files to understand the implementation.
Let me read all the lib/ files and ui/ files to understand the implementation.```

[dry cedar]

I like the humor of this when chit chatting about opinions on things

[brave spire]

no image support 😭

||Hunter reasoning traces seem oddly deepseek. I have not used deepseek for half a year but I remember. Healer sits between probably or probably not, but also, reasoning trace matters.||

[brave spire]

Was gemini 2.0 thinking experimental or something before or after R1?

[fervent spire]

wut is CCP?

China Communist Party?

😭😭😭

@proud canyon

[summer anchor]

Yes google had a 2.0 flash thinking in preview before R1, but well technically deepseek had R1-lite in the web even before that

[brave spire]

So nonetheless the second lab to ship reasoning, and better yet, proliferated it to everyone

[fervent spire]

THIS MODEL CAN'T ESCAPE ITS OWN OUTPUT😭

[tame olive]

"I can stop any time I want"

[torn siren]

AGI confirmed

[teal wagon]

Does anybody know when the hunter models are going away?

[nova swan]

no

[wet escarp]

Is it just me or the hunter tps has reduced significantly since yesterday?

[autumn reef]

Ya, it went from 70tps+ to 30ish.

[nova swan]

expected due to the claws and the gooners

so

i think im pretty happy w the model as a flash-style model

all comes down to the pricing really

[undone wyvern]

It's 1T

[wet escarp]

any faster and smarter model for free ?

[nova swan]

no

[nova swan]

gemini 3 flash is around 1t too

[smoky leaf]

it's just the style

[undone wyvern]

Never thought I'd die side by side with the jai gooners, fighting clawers

[tidal bough]

Gooner stands no chance against clawer

[modest bridge]

we need to form an lotr type alliance against sauron (vibe code bros)

[undone wyvern]

https://tenor.com/view/gimli-aye-gif-4146646146679545792

[undone wyvern]

For some reason, it loves to come up with depressing / sad scenarios. Why - it is weird and interesting to see, or maybe it catches some prompt tokens too literally

[dry cedar]

You gave it AI depression before it could give you AI psychosis

[undone wyvern]

I literally plug'n'played my usual tests with different scenarios

Bro is not okay after interacting with clawers

[dapper iris]

I'm quite fond of its personality somehow

[cloud hinge]

tps
tolkiens per sekond

[topaz cloak]

Gooners should start multi gooning to compete

Sub gooning

[tidal bough]

wait for them to start using tool call to compete with Claw

[undone wyvern]

Agent gooning?

[uncut vault]

<|tool_call|>edge_user

[tame olive]

parallel gooning

[fading knot]

anyone know when it is going away/become premium model? so i can prepare my heart XD

[sacred spindle]

what model is this?

[lyric orchid]

No

No info is ever given, aside from stuff in #announcements

[violet harbor]

Think OnlyMolts got there ahead of them

Allegedly (I haven't verified myself) a number of the molts have paying subscribers

[analog wraith]

Now supports vision

[undone wyvern]

Huh?

[undone wyvern]

Oh no

Please don't be DS 4

https://tenor.com/view/i-am-about-to-crash-out-sunflower-pvz-plants-vs-zombies-gif-2216785787877099970

[modest bridge]

its not someone made a reddit post confirming it

[undone wyvern]

I don't trust anyone anymore

[modest bridge]

all the proof they gave was plausible lemme find it

https://www.reddit.com/r/SillyTavernAI/s/UlkK1bIku6

either way reveal is gonna be this week so we'll see

[undone wyvern]

DS4 considering it's 1T size would be trained and RLHF-ed from the scratch or close to it, so it can have totally different architecture and 'token bugs'

[modest bridge]

yep valid counterargument which if true just disassembles this entire post lol

[undone wyvern]

I am not 100% sure about going 650B->1T and fully retraining being the only way, maybe someone who worked with DS V1-V2 can say for sure

[harsh plinth]

That bad?

[fiery scarab]

yo whats the word who is the model

[strange cliff]

its a chinese model

maybe deepseek maybe stepfun maybe minimax

proof its a chinese model: it has strict compliance with china laws

[lyric orchid]

It's 100% Chinese

Interesting slop profile on EQBench Long

Most Similar To:
healer-alpha (distance=0.745)
Qwen/Qwen3.5-27B (distance=0.757)
deepseek-ai/DeepSeek-V3.1 (distance=0.761)
XiaomiMiMo/MiMo-V2-Flash (distance=0.774)
Qwen/Qwen3.5-122B-A10B (distance=0.775)

No Western model to be seen

[undone wyvern]

So, claude-distilled models

[lyric orchid]

I mean if you dig deep enough it's all Gem or Claude distilled =P

But neither shows on its profile

[nova swan]

woah they added image inputs ?

thats cool

actually thats just great

[runic nymph]

hallucination maybe but hunter alpha tells me its mimo by xiaomi

[native moss]

mistral 💀

[dry cedar]

Picks a random company

[violet harbor]

should start a sweepstake or a voting poll on these stealth models...would be amusing.

[harsh plinth]

Openrouter will give 10$ credits to whoever gets it right

🔥

[woeful summit]

actually?

[smoky leaf]

sure

[marble stream]

i am so hoping it is an open wights model

[pseudo rover]

Plot twist: the 1T parameter thing is a red herring and it just randomly routes to a different CN model for each request it receives

[golden dagger]

what sup with hunter alpha today?

[tacit fossil]

what do you mean sir

[golden dagger]

Can't use openclaw

With alpha hunter

Provider error

Returned

[tacit fossil]

it just messaged it and it worked. try to get more info. potentially it's getting overwhelmed by claws

[golden dagger]

Provider returned error every second

I just sent a message

On telegram

[tacit fossil]

ask in the openclaw discord how to debug openrouter provider errors. it sounds like something is wrong with your config

[kind marsh]

Uptime tanked
Can't generate anything

[lyric orchid]

What's a Clawl?

[outer hollow]

https://github.com/openclaw/openclaw/pull/49214

[dry cedar]

Oh, huh

Healer is indeed 262K and Hunter is 1M

[lavish sleet]

yeah, mimo and ring were both the top guesses

this makes sense

a bit underwhelming from mimo though

[harsh tree]

Hunter is one of the best long context models I have tested

kinda sad its going

[hallow raven]

gonna generate some last minute dataset from this then

[nova swan]

what are you generating

ireally want to generate datasets but i dont even know like what to burtn the tokens on

[hallow raven]

im just using the ring lite sft dataset

and some people's claude ones

and reusing the questions

[tranquil ledge]

if it was xiaomi then its way better than their previous release

last mimo was garbage

hunter alpha is pretty good

[dry cedar]

MiMo V2 is pretty good for the cost, though

[harsh tree]

Caching was really good too

sadly they dont have batch api

[harsh tree]

Yup its going to sunset any moment now

[strange cliff]

hunter alpha is getting removed?

as well theres still healer

[tranquil ledge]

all i care about is rp

and its okayish at that

[harsh tree]

get a life

[tranquil ledge]

sloppy af

[tranquil ledge]

perhaps there are more important things to do than reply to random people within less than a few seconds

[harsh tree]

I mean I switched tabs and your message just popped up

rionic for a guy who RPs with LLMs talkling about important things

[dry cedar]

No need to instigate this conflict

[undone wyvern]

Agreed, being toxic is a no-no on this server

[hybrid sundial]

https://tenor.com/view/flight-does-he-know-faded-gif-11697106329231392146

[harsh plinth]

Uno Momentum

[undone wyvern]

https://tenor.com/view/nostalgia-critic-spin-speed-fast-chair-gif-26770162

[nova swan]

and generating responses?

what is the point though

i get it if it was a claude response or something but

:(

like for distillation?

[hallow raven]

yeah

[hallow raven]

i dont really have any data stuff to do rn so i just want to keep its response style saved for later 🤷

[nova swan]

i bet two days after it goes away ill have the biggest project ever that will require the most tokens ever requried

and i will be so sad that hunter is gone

and wait until the next stealth model :(

[dry cedar]

Or be like me, BYOK to everything that has a free tier

[heady fable]

https://github.com/openclaw/openclaw/pull/49214 - Both are Xiaomi Models

[nova swan]

i try to, which ones do you have?

[dry cedar]

Groq, Mistral, Cohere, NVidia NIM

Most of it goes to Mistral

[patent zenith]

Groq has a free tier?

[dry cedar]

Yeah, they're fairly generous: https://console.groq.com/docs/rate-limits

(For low context requests)

[nova swan]

did this model just get better ?

it was looping a LOT before in agentic/opencode

and now its not doing it at all

[undone wyvern]

It changed reasoning style and got better in writing about ~1.5 days ago

[patent zenith]

their signup flow is completely broken for me

[harsh tree]

Not yet

[lyric orchid]

Aight, what are your guys' price predictions?

[heady fable]

Xiaomi, both of them

[tidal bough]

3x the price of mimo v2

[violet harbor]

On hunter, probably competitive with GLM 5

[rugged thistle]

i wonder what model this would be

[plucky oriole]

Is hunter alpha down?

Or they're going out of stealth?

Both uptime and latency is very bad now

[lyric agate]

Hope mimo being cheap

Xiaomi.. Bring your cheap phone strategy into machine learning development

[lunar oasis]

Full benchmark
https://artificialanalysis.ai/models/mimo-v2-pro

[tidal bough]

Deepseek price

[kind marsh]

wait when is it leaving
i don't want it to go, its genuinely peak

[undone aspen]

It doesn't have the "(free)" so don't get your hope too high up

[kind marsh]

dude for the quality its outputting for me, ii'll gladly pay, i just wanna know when its leaving
or where to find what its gonna be called once its out of stealth?
do you know what it is?

so is this model basically mimo v2 pro ? or what is it?
cuz if it is mimo then damn makes sense, cuz mimo v2 flash is really good for its price

[tacit fossil]

mate, no one knows for sure. this is how stealth models always work. this is fun!

[kind marsh]

sorry. i'm just really enjoying the model right now
its literally perfect for everything i want it to do for RP

so forgive me for being hasty - just gonna really miss it once its 'gone'

[tacit fossil]

it's probably mimo v2 pro, and probably going to be released very soon

i've removed my minimax thumbnail to prevent any further emotional distress

[kind marsh]

its a straight up upgrade from mimo v2 flash in pretty much every way, and it is quite stable like mimo v2 flash
hopefully the price will be decent and not too outrageous - although if its xiaomi, shouldn't be too bad

[obsidian sail]

Now you know that im always right.
Remember this in the future.

[kind marsh]

when exactly do you predict is 'very soon'?

[heady fable]

My watcher has spotted the models from Xiaomi's Models endpoint so very soon

[patent zenith]

[tranquil ledge]

excited for the tts

best tts so far was the qwen release

[heady fable]

I am curious about it for sure, will be interesting to see how it performs, especially against giants like ElevenLabs - still nothing on docs though so just waiting - you can make reqs to pro (I would guess Omni too) - just no concrete idea of pricing for Omni yet

[tranquil ledge]

omni is probably gonna be priced like mimo flash

[heady fable]

Realised I could just make a req to each and Xiaomi providers reasonably nice pricing info, so can work out pricing:
Pro:
AA already confirmed but does seem to be $1 per mil in and $3 per mil out at least for small context

Omni:
Coming at around $0.41 per million input, and $2 per million output for text tokens (got some cached usage too, that comes out around $0.125 per million cached read with a small sample estimate)

[patent zenith]

Praying healer and hunter are actually Mimo models

[heady fable]

I'd be very surprised if not - there is too many similarities for them not to be I think

[heady fable]

https://platform.xiaomimimo.com/#/docs/pricing

Docs updated with info

[heady fable]

Limited time free!!

[unkempt stag]

Will these models be posted today?

[heady fable]

Sorry, how do you mean Patrick?

[unkempt stag]

Will they be added to OpenRouter today?

[heady fable]

Oh, I'm not staff, I am sorry, I believe the Hunter and Healer Alpha models are Pro and Omni - I don't know if @proud canyon can confirm yet but all signs point to the stealth models being these models

[unkempt stag]

Okay, thank you! But since these models are getting good reviews, they'll surely be put into service quickly.

[heady fable]

I would expect so yes, as they are stealth, I would expect OR to have a good relationship with the MiMo team and they will come online pretty quickly

[hallow raven]

Hunter Alpha shown below is an early anonymous version of MiMo-V2-Pro.
https://platform.xiaomimimo.com/#/docs/news/v2-pro-release

[undone wyvern]

Good. Not Deepseek

[unkempt stag]

It seems the answer came quickly...

[undone wyvern]

Wait. Not Deepseek 😭

[covert robin]

wow

I'm amazed it was actually mimo v2 pro, as in my few code tests it consistently made way shittier code and made dumb mistakes that mimo v2 flash did not

Mimo V2 Flash:

Hunter Alpha:

what an insane regression in coding style

sad it wasn't Ring/Ming

my pull was wrong 😭

[tacit fossil]

actually this is genius, its just writing transpiled js directly. no build step required

[covert robin]

Write-only code

(Also lots of embedding its thinking inside of the code in Pro, which was not there in Flash!)

[tacit fossil]

when did did you do this?

[covert robin]

I wrote off mimo because I thought to myself “no way a serious lab would regress this severely in code…”

[covert robin]

a couple days ago

I can redo it

[tidal bough]

You might try it again after free model is over

[tacit fossil]

#1481382840540074004 message hmmm

[covert robin]

I’ll try again

[tacit fossil]

i know, why don't i just actually test it

i know coding

[tacit fossil]

looks good to me

[covert robin]

Nope still just as shit

I guess it could be my specific test or something…

Can you post a sample?

[dry cedar]

Well, can this be prompted away?

[covert robin]

Perhaps

Default behavior is an insane regression though

💀

At least in this test

[tacit fossil]

In Javascript, write the code for a cellula automata roguelike dungeon generator

i didnt expect it to do all this

[covert robin]

I think not really

I asked it “write good, readable code” in the prompt this time

Very slightly better, perhaps

Still super code golfed

long css lines without new lines…

One letter class names

[tacit fossil]

what's the full prompt you used?

i was joking before but it does actually look like "compiled" js

[dry cedar]

"Use descriptive variable names"

[tacit fossil]

maybe its regurgitating literal chess.com js

[vivid wigeon]

Yea the model was terrible at both coding and creative writing

thankfully it wasnt deepseek

[digital ledge]

nah.

I dont agree with you that it's terrible at creative writting

[tacit fossil]

ok i asked it for a chess game and it wrong similar looking code

its working though...

[modest bridge]

yeah i heard its creative writing is good too

[covert robin]

ah yes... i am the hybrid white/black pieces..

[harsh tree]

Long context of hunger lapha is top notch

[tacit fossil]

bro this is actual beautiful

[covert robin]

I too like hunger lapha

[tacit fossil]

https://www.youtube.com/watch?v=F2AitTPI5U0

[limpid star]

If Xiaomi makes models, will they make cheaper ones and name them Poco?

[tidal bough]

Redmi-R1

[undone wyvern]

It changed reasoning style again, I don't know what's happening

[smoky leaf]

for better or for worse

[undone wyvern]

It seems to follow instructions better both in reasoning block and in output. Not sure yet - but it's suddenly different from yesterday's

[digital ledge]

Creative writting?

[undone wyvern]

Yes, but different tasks too

Or maybe I am seeing things

[harsh tree]

Loved the hunter model

0< Token ≤ 256 K：Input: $1.00 / 1M tokens, Cached Input: $0.20 / 1M tokens, Output: $3.00 / 1M tokens

lmao with this costing I would have spent a north of 15k USD

😋

[kind marsh]

FUCK it's expensive
Goddammit back to using mimo V2 flash

Jeezuz at 14k context it's blowing through 1 cent per msg.
Fuck. I love it, but damn it the price is Gemini level

[pseudo rover]

Yeah GLM 5's better for now, sorry

Still if they follow through with improving it some more before releasing the weights like they said they would, maybe it'll eventually be worth it

[kind marsh]

Not necessarily. I've used glm 5, it's still relatively inefficient with the reasoning tokens

[pseudo rover]

I really, really want to like it but I just can't do it. It reasons a little less, but the reasoning itself is wildly inconsistent at times and it doesn't necessarily feel better for it. Idk, it just doesn't land for me yet

It'd be a lot more competitive if it was either better or cheaper

[patent zenith]

Ime GLM yaps for a goddamn millennia. I haven’t used MiMo V2 Pro or Omni to have strong opinions about them, but yeah, the G3 flash price is very contentious since there’s already genuinely SOTA models at that price, so any model that falls short of that is underwhelming and feels as though it’s just not worth the price.

[kind marsh]

I'll still keep using it because I'm too hooked with the quality. Ah well, might cost me more money long term but fuck it it's too good unless mimo V2 flash somehow improves

It uses 1/5 of the output to reason, which already makes it cheaper than you think
But yea compared to glm 5 it's definitely more pricey

[kind marsh]

What's g3 flash?
Gemini?
I'd use it but I'd get censored or banned, I heard it's pretty strict on NSFW

[patent zenith]

Gemini 3 flash. I suppose its best not to ask, different strokes for different folks I suppose

[nova swan]

bro HOW is it even possible that this mimo v2 pro shit is more expensive and worse than glm

[digital kelp]

its not

[patent zenith]

what model is on the right? Thats not the pricing for GLM

[digital kelp]

glm 5 with z.ai provider

[digital kelp]

glm 5 isnt supposed to have 80k context

[patent zenith]

ah true but still

cache hit rate is crazy right now for MiMo Pro

[lyric agate]

It's the official pricing, the 2.3$ is the pricing for the extreme one, aren't that reliable imo

Fireworks and Together seems to be the best provider as of now, but for quality the official endpoint from Z.ai is just better

[upper parrot]

yes, alot of use, but im pretty sure like 99% of those pppeople dont actually get anything done with it

[patent zenith]

True

[smoky leaf]

don't you dare saying that to one of them

[drowsy moth]

What did this model end up being?

Was it the mimo one?

[dry cedar]

Yep

[drowsy moth]

Got it. Was it the pro or omni one?

[dry cedar]

This one was Pro and Healer was omni

[drowsy moth]

Gotcha, thanks for confirming!

[tame olive]

#announcements message

[plucky oriole]

They nerfed it didn't they

[wheat cave]

(it was mimo 😭)

[storm trail]

previously, mimo was so irrelevant nobody think of it seriously. no way dipsy is this unpolished, hail dipsy.

https://tenor.com/view/stalin-joseph-stalin-get-out-get-outta-here-cigarette-gif-2377403250026995476

[tacit fossil]

this is why democracy doesnt work tbh

[wheat cave]

I prefer an AGI dictatorship

[strange cliff]

when mimo v2 flash released and had a free model on openrouter too i once asked it hi and it talked about some random school thing

[harsh tree]

They updated the model many times during free period