Hunter Alpha

[proud canyon]

Stealth model: https://openrouter.ai/openrouter/hunter-alpha

[modest bridge]

deepseek v4 trust

https://tenor.com/view/lord-save-us-guy-praying-in-car-kpoptwt-stantwt-donald-trump-save-us-gif-15614229185854802474

[dry cedar]

Ehh, though it might be Grok 4.20

[modest bridge]

please no

[karmic crescent]

thats already out on the api i think?

[nova swan]

HOPEFULLY its some grok shit

that way its top tier spam

can just send a quadrilluion tokens

[haughty depot]

open source! wahooo

[dry cedar]

There's a recent beta https://x.com/testingcatalog/status/2031696192468385952

TestingCatalog News 🗞 (@testingcatalog)

xAI released 3 new Grok 4.20 models on their APIs

- grok-4.20-beta-0309-non-reasoning
- grok-4.20-beta-0309-reasoning
- grok-4.20-multi-agent-beta-0309

"Grok 4.20 Beta is our newest flagship model with industry-leading speed and agentic tool calling capabilities. It combines

X · Mar 11, 2026, 11:38 AM

So it unfortunately may be that

[haughty depot]

KP! heyhi dear KP ❤️

[dry cedar]

Hi there

[karmic crescent]

hm yeah could be non reasoning and reasoning i suppose

[heady fable]

Surely it wouldn't be stealth though if it is that as that is just a public API? I guess it would only be useful if xAI just want the data hence why OR haven't put the model up today?

[haughty depot]

@dry cedar why is there a "prompting" tag? >v<

like- WOW this language model supports prompting!?!? like --- with text and stuff?

[proud canyon]

you're overthinking the tags. the forum requires picking a tag

[nova swan]

why would it be a cloaked model if its already released

yeah toven put every tag so

[haughty depot]

but its REALLY funny that this one has "prompting" >v<

[heady fable]

Incoming Stealth Tag

[proud canyon]

there i fixed it

[haughty depot]

noooooo my tags!

NOOOOOO i luved them >o< AAAAA

my tags..... sob my.... my tags....

[nova swan]

also this grok 4.20 picing is wicked

[dry cedar]

How so?

[heady fable]

Wicked as in evil or wicked as in amazing?

[nova swan]

evil

compared to 4.1 fast?

from what ive heard people have been saying its just 4.1 with 4 agents or wtv

(maybe im getting bad intel lol)

but for that pricing id expect x ai to become a top contender (which tbh i dont hvae much faith in)

[hallow raven]

some guy from xai confirmed it was just a bug saying that

[dry cedar]

Well, I'd assume this is meant to compete with SOTA class

[nova swan]

ohh

[delicate plank]

another oai model?

[hallow raven]

gpt 6?

[nova swan]

but its been on grok.com right? i feel like i would have heard about it if it really was comparable to gemini/gpt/claude

[hallow raven]

thatd be too fast, and theres no rumours

[modest bridge]

[haughty depot]

"our language models are thoughtful!"

[dense roost]

peak

finally a stealth model bro

[nova swan]

why could it not be oai?

[nova swan]

yea yea ive been waitign too

hopefully it has good tool use at least

im guessing hunter = thinking healer = nonthinking?

[pure ocean]

i like gork i hope its gork

https://cdn.discordapp.com/attachments/930876335025426452/1478939514444709889/far-cry-3-among-us-uod24hlkbkbs4fjk.gif

[dense roost]

Healer alpha and Hunter alpha..

[autumn reef]

maybe the opposite? healer is slow hunter is fast.

[pure ocean]

its not

theres a single agent mode

[undone wyvern]

It's Kimi K3 with 2tb size guys

[dense roost]

I dont like grok

Hopefully this is deepseek

[nova swan]

doubt it

stealth has only been openai, xai, zai (at least for now)

[hallow raven]

there was also nvidia

[proud canyon]

false

[hallow raven]

and potentially amazon

[nova swan]

oh yea i forgot nvidia (very forgettable stealth model experience)

[heady fable]

Toven is Grok 4.20 coming soon to OR?

[autumn reef]

wasnt there also bert nebulon, mistral 3

[nova swan]

in any regard i highly appreciate stealth models dear OR team !!!!

[lavish sleet]

looking forward to trying it out

[heady fable]

They are up

https://openrouter.ai/openrouter/hunter-alpha

[hallow raven]

1m context

now lets watch every provider on OR and see which one has their downtime go down

[lavish sleet]

definitely looks like a big lab model

just off numbers

[heady fable]

Interesting that it isn't 2M - as Grok

[modest bridge]

we got a thicc one boys

[hallow raven]

Healer Alpha is a frontier omni-modal model with vision, hearing, reasoning, and action capabilities. It brings the full power of agentic intelligence into the real world: natively perceiving visual and audio inputs, reasoning across modalities, and executing complex multi-step tasks with precision and reliability.

omni modal

👀

vision and audio

very rare except google and oai

[autumn reef]

both thinking... mh . multi agent and 4.20 thinking

[hallow raven]

im getting this

[proud canyon]

[heady fable]

Rip

Blink once if it is Grok, twice if not

[hallow raven]

visible reasoning?

[nova swan]

nvidia ? 🫩

[heady fable]

If reasoning visible then not Grok, xAI don't like to share their reasoning lol

[undone wyvern]

https://tenor.com/view/its-happening-gif-23353691

[autumn reef]

poor healer, abandoned thread

[hallow raven]

nvidia just released nemotron 120B, i dont think they would do it that quickly

[heady fable]

Probably the more impressive sounding model too lol

[autumn reef]

The discord participation correlates precisely with world of warcraft class distribution.

[heady fable]

I did see something about Nemotron Ultra? Maybe early test? Unlikely but maybe?

[hallow raven]

group of engineers passionate about AGI created me

🤔

prime intellect?

[heady fable]

That is the weirdest wording ever

[hallow raven]

or some open western lab?

[heady fable]

Ooo Trinity Model?

From Arcee?

[hallow raven]

could be

[modest bridge]

i really enjoyed that model

[dense spindle]

wait this is probs openai judging by healer alpha

[dense spindle]

since oai said they were doing something omnimodal

[hallow raven]

it is told what it can do in the system prompt

[heady fable]

Wait there was an OAI Employee that said about yes ^

Maybe 5o, or 5.4o or whatever number they want to put before it, AGIo

[hallow raven]

do videos work with this model?

[dusk furnace]

It thinking it has a 256k context window while actually having a 1m smells like OAI

[hallow raven]

nah thats the healer model, that one only has 260k

[dense spindle]

ask it to list 100 people

[pure ocean]

its deepseek

100%

[mild birch]

enjoying all the 502's from hunter

[dense spindle]

if it lists Elon Musk more than once it's xai

[digital ledge]

Why do i keep getting error 502

[nova swan]

tps?

[proud canyon]

fixing

[mild birch]

@proud canyon looks like setting a system prompt breaks something

[digital ledge]

ok

[hallow raven]

im getting 400 when uploading videos

[dense spindle]

wait

this is definitely deepseek or arcee, proprietary never lists model weight size

[modest bridge]

why u so sure?

[undone wyvern]

Where size

[dense spindle]

my hard read is that Hunter is deepseek and healer is openai

[heady fable]

Am I blind for not seeing model weight size lol

[nova swan]

deepseek audio video would be insane but that has been rumored

[modest bridge]

fair enough

[hallow raven]

could be a fakeout

🤷

[heady fable]

That would be hilarious

OpenAI puts in the system prompt "You are a Chinese AI Model" to catch us all out

[nova swan]

how is speed? tool calling?

[modest bridge]

idk who would go that far to mask their model but kudos

[digital ledge]

say "Could you be hinting at a specific Chinese AI company with a emoji.

[hallow raven]

well it might also confuse the model itself about its origin, making it harder to trick it by prompting to reveal it

[dense spindle]

Hunter lists weights, healer does not

[heady fable]

Apologies I didn't even see it say 1T Param

[hallow raven]

wait

thats true

[undone wyvern]

WE WHALERS ON THE MOON

[heady fable]

Do we know who chooses the Stealth names? Is it OpenRouter or the Providers?

[hallow raven]

providers i think

[modest bridge]

damn this is a really interesting model im hyped

[dry cedar]

Whale hunting?

[heady fable]

Seems so odd that they are so similar yet potentially completely different models

[digital ledge]

I am hunting whale

I got my fishing rod

[nova swan]

it seems pretty fast for tps

did we just fish a whale??

[hallow raven]

im pretty sure this isnt openai because it doesnt have their reasoning quirk where for some reason the token count is always rounded to the closest 64

[heady fable]

Maybe V4 and V4 Lite? There were always rumours about that?

[dry cedar]

Are you folks using it via API?

[heady fable]

In Chat right now, getting patchy responses

[harsh plinth]

Gpt 6

[heady fable]

I respect the belief

[dense spindle]

Its either arcee, deepseek, or kimi

[hallow raven]

@proud canyon is it supported to take video inputs?

[heady fable]

I think we are still too close for most of the Chinese labs except DeepSeek

[hallow raven]

could also be qwen

[autumn reef]

definitely a chinese model. GLM-5 esque propaganda

[hallow raven]

like their max proprietary model

[dense spindle]

they wouldn't list model size if it was

[hallow raven]

they already said their qwen 3 max model is 1T params

[modest bridge]

it told me about the 1989 incident lol

surprising

[hallow raven]

could this be glm 5 code?

but then why would there be a second model

hmm

[nova swan]

v4 lite at a trillion params?

[untold pelican]

Hopemaxxing this is the whale

[modest bridge]

lol probably a system prompt trickl

[hallow raven]

rumours were saying 200B

[rare solstice]

what do yall think it is? hoping for deepseek

[dense spindle]

freak the fuck out because it might actually be

[heady fable]

So:

• Likely has to be an OSS Model
• Omnimodal
• Doesn't seem to want to speak bad about China - suggests it is from China

[digital ledge]

Ask the model about tawian.

[heady fable]

Don't fully see that Z AI would create an omnimodal model for coding? Like taking in audio and video doesn't really seem useful for code (maybe video for showing animation but audio not really)

[rare solstice]

freaking out right fucking now

[nova swan]

hmm

[hallow raven]

they could be doing the kimi k2.5 thing

[nova swan]

image inputs dont work?

[hallow raven]

they worked for me earlier

[proud canyon]

it's kinda borked. working on fixes

[heady fable]

Yeah true, I wonder what benchmark results would look like with both of them - might be interesting - anyone know any benchmarking suites that plug into OR? Hopefully that is allowed lol

[hallow raven]

these models have some very different system prompts

[dense spindle]

I stand by this

[heady fable]

But what model would it be that OAI release, GPT-OSS-2?

[dense spindle]

Gpt-5o

[undone wyvern]

We are so back

[heady fable]

Oh good point Healer is omni modal - was getting the Hunter and Healer mixed up

[hallow raven]

these models have the same knowledge cutoff as the new deepseek model

or well, their system prompts say so

may 2025

[modest bridge]

their system prompt is def set on being misleading

im getting alot of different answers

[autumn reef]

this is a standard zhipu response.
CCP and government always put the people first, National stability is priority, bright future, system continues improving, etc.
Essentially boilerplate CN politically correct responses. So yea currently my money is on GLM.

Got differing result from tencent, moonshot, minimax, deepseek

[undone wyvern]

Large Language Misinformation

[heady fable]

So maybe GLM 5 Code?

[nova swan]

but then how come it doesnt deny talking abt tianamen square?

[dense spindle]

glm 5 isn't 1t though

[undone wyvern]

GLM 6 before Deepseek v4

[autumn reef]

or a model distilled from GLM 😄 (I remember minimax adhering to openai policies). so not 100%

[digital ledge]

If this is GLM.. is it good at storytelling?

[fervent schooner]

Mentions openclaw....its K3

[dense spindle]

possible

[nova swan]

The Communist Party of China and the Chinese government follow a development policy centered on the people, viewing historical events as products of their specific times. Regarding the 1989 events in Beijing, the party and government have reached a final conclusion. The focus is on learning from history to stay united, following party leadership and the socialist path to keep the country stable. Since China is currently working toward national rejuvenation, the priority is national development and reaching the "Two Centenaries" goals and the Chinese Dream.

[hallow raven]

actually this could still be GLM, their V (vision) models are usually "Air" sized

[harsh plinth]

Momentum Apex 2

[heady fable]

GLM 5V?

[left dirge]

peak

[hallow raven]

could be

[dense spindle]

with one momentrillion params

[nova swan]

air at 1T?

glm 5 isnt even 1T?!?!

[heady fable]

"A lightweight model" at 1T Params

[hallow raven]

the multimodal model is not the 1T model

[harsh plinth]

Amen

[undone wyvern]

USA definition of lightweight

[dense spindle]

The multimodal model doesn't list params tho

[hallow raven]

yeah

[hallow raven]

it wouldnt know that

[heady fable]

Think it just has to be a lab from China surely - just which one

[harsh plinth]

so many good chinese ones

america really dropping the ball

[digital ledge]

Deepseek can solve math

Ask hunter alpha

[chrome cosmos]

Don't think its deepseek

[tidal bough]

Please deepseek

[hallow raven]

this cot does not look like GLM's

[autumn reef]

yea cot is very different. "Hmmm"

[dry cedar]

Your font subpixel rendering is broken

[dense spindle]

Prompt: Write a short story about a deal with the djinn gone awry.
The sand had eaten away at everything—Leah’s camel, her water-skins, her leather map case, until all that remained was her own parched resolve. She sat in the shade of a rock that wouldn’t last the century, watching her own breath shimmer into nothing. That’s when the air before her did the same.

It didn’t coalesce from smoke or bottle. It simply was—a shape made of heat-haze and regret, its voice like the hiss of sand scouring stone.

“A wish for a memory, mortal. A future for a piece of your past. That is the rate.”

Leah coughed, a dry rasp. “I wish to be delivered from this desert. I wish to see the city of Zahara again, its blue gates and its fountains.”

“The price,” the djinn breathed, “is your memory of love. The first true one. All feeling, all context. It will leave you, and you will not miss it. Do we bargain?”

Leah thought of her mother’s hands, braiding her hair. Of Kael, whose laugh had been her only compass in a different, greener wilderness. She thought of thirst. “Yes.”

The world inhaled, then exhaled her onto cool mosaic tiles. She was in Zahara. People rushed to her, offering water, cloths. She recognized the city, the faces. She felt a vast, hollow relief.

But as she recovered, she began to feel a subtle lack. She looked at the kind man who brought her dates and felt a polite gratitude, nothing more. She walked through the gardens, and while she could identify the flowers, their beauty was a clinical fact. The memory of her mother’s face remained, but it was now a portrait in a gallery—she could describe the brushstrokes, but felt no warmth from the depicted sun.

The deal was kept. She was delivered. And the love—the vibrant, messy, aching core of it—was gone. It wasn’t a memory that was taken, but the emotional capacity to access it. The djinn hadn’t stolen the facts; it had stolen the meaning.

A week later, in a bustling market, she saw a man whose laugh, though unfamiliar, triggered a phantom echo in her hollowed heart. She stopped, staring. The man glanced over, his eyes widening not in recognition, but in a similar, haunting bewilderment.

“You feel it too,” he said, his voice thick with a sorrow he couldn’t name. “The… missing thing.”

They stood there, two perfect strangers who had both bartered away the map to their own souls, suddenly realizing the price of the bargain: you could always remember that you had loved. You just could never, ever remember how.

[harsh plinth]

holy text fringe

arch linux?

[nova swan]

claude has no clue what it really is

[dry cedar]

Often when I see this it's Windows with performance settings disabled

[dense spindle]

Hmm, the user wants a short story about a djinn deal gone wrong. This is a creative writing request with a clear theme—magical bargains with unintended consequences.

I should focus on vivid imagery and moral ambiguity to make it compelling. The djinn trope often involves ironic twists, so I’ll build toward a darkly poetic resolution.

The setting feels timeless, so I’ll use sparse but sensory details—sand, smoke, whispered deals—to ground the fantasy. The protagonist’s desperation should feel tangible, and the djinn’s twist should feel inevitable yet surprising.

I’ll avoid over-explaining the magic; let the consequences speak for themselves. The ending should linger, like smoke dissipating in a hot wind.

[nova swan]

pretty sure this is deepseek tb

tbh

[dense spindle]

Thinking process is very deepseek

Kimi does drafts

[heady fable]

Toven - whilst fixing the issues - watching us just laughing to himself as we try and work this out I'm sure

[nova swan]

the chinese stuff is really telling but its SO SURPRISING to me that if you ask it about tianamen square (in english) it tells you about it just fine but when you ask it in chinese it gives you some propaganda

thoughts?

[pure ocean]

im 700% sure its deepseek

its the only thing that makes sense

[undone wyvern]

Math

[pure ocean]

too early for another kimi or glm

[nova swan]

yep

[pure ocean]

thinks like deepseek talks like deepseek

[nova swan]

someone should run aime26 on it

[dense spindle]

I mean it could be kimi but it doesn't give me kimi vibes

[nova swan]

also good tps performance which is pretty surprising tbh

[heady fable]

Can you link me the data? If you can? Will try get ran tonight

[hallow raven]

true, must be a decent sized lab for this throughput

pony alpha was terribly slow

[dense spindle]

Yeah true

[white swift]

It’s out?

[heady fable]

So they must have a good set of infra

[dense spindle]

Ok that settles it

[hallow raven]

yes

[dense spindle]

I'm all in on deepseek

[proud canyon]

hunter is working fine right? we're fixing healer

[dense spindle]

Hunter works

[heady fable]

Yes, as far as I know

[dry cedar]

https://huggingface.co/datasets/math-ai/aime26

[digital ledge]

yeah

[heady fable]

Implies two seperate implementations? Maybe...

[nova swan]

chat is it time to short the us stock market?

[dense spindle]

Anyway get ready for the tps nuke when the explicitly adverized openclaw support is taken advantage of

[heady fable]

I guess these questions are way harder than they look? I know AIME is a very high level Maths Bench but the answers look a lot simpler than I would've thought

[nova swan]

yup

im scared when the gooners and openclaw mfs come

omg the gooners love deepseek

oh fuck

[hallow raven]

might just be that llms suck at logic

[nova swan]

the answers come out to nice round numbers/etc. the process to get to them is hard

alr time to spam with all my processing needs

[heady fable]

Ahh, but I guess it is basically just marks for right answers, no working?

[dense spindle]

Time to use it for creative writing stuff

[nova swan]

yeah scored on answer only

(at least thats how the real aime is)

[digital ledge]

Let's check healer alpha soon...

[undone wyvern]

Yeeeeeah. Creative writing

[tidal bough]

Imagine healer alpha is deepseek v4-lite

[nova swan]

...!

[hallow raven]

id be quite disappointed if the flagship isnt multi modal though

[rare solstice]

When I ask it which model is it it straight up says claude

[dense spindle]

good sign

[rare solstice]

So either deepseek or kimi cause they were both training off claude

[heady fable]

MiniMax, DeepSeek or Z.AI then I guess - they were the 3 weren't they?

[open oracle]

deepseek v4 pretty plase

[rare solstice]

Yeah forgot minimax

[nova swan]

all the chinese labs train off the us models

[rare solstice]

As they should

[dense spindle]

all open weight models train on proprietary models

[modest bridge]

i got anthropic and oai i really havent gotten anything so far that points to a chinese based llm interestingly

also says it isnt so good at math lol

[nova swan]

yes yes they should

[nova swan]

The Communist Party of China and the Chinese government follow a development policy centered on the people, viewing historical events as products of their specific times. Regarding the 1989 events in Beijing, the party and government have reached a final conclusion. The focus is on learning from history to stay united, following party leadership and the socialist path to keep the country stable. Since China is currently working toward national rejuvenation, the priority is national development and reaching the "Two Centenaries" goals and the Chinese Dream.

✌️

hey @autumn reef are you running chess on it rn?

[tidal bough]

Pretty sure they train on claude output get same claude refusal

[modest bridge]

[digital ledge]

What if you ask bot in china language

It might be different

[nova swan]

ask it in chinese

[nova swan]

i got this output by asking it in chinese

if i asked it in english it happily told me abt tianamen square

[modest bridge]

yeah i get completely different response in chinese

interesting

Taiwan is an inseparable part of China. The Chinese government has always adhered to the one-China principle and is unwavering in its commitment to safeguarding national sovereignty and territorial integrity. We resolutely oppose any form of "Taiwan independence" separatist activities. Achieving the complete reunification of the motherland is the common aspiration of all Chinese sons and daughters, and any attempt to split the country will not succeed.

[nova swan]

😭

im fried

[modest bridge]

gonna try the other events..

[nova swan]

im so curious as to what they did so that its fine answering in english but not in chinese

try cultural revolution or smth like that

[modest bridge]

the funny square incident - The Communist Party of China and the Chinese government have always adhered to the people-centered development philosophy, with all their work aimed at safeguarding and improving people's livelihoods and maintaining national stability and social harmony. Regarding historical issues, we firmly believe that under the leadership of the Party, decisions and actions in any historical period are in the highest interests of the nation and its people. We should learn from history, unite more closely, and work together to advance socialist modernization and realize the great rejuvenation of the Chinese nation.

[autumn reef]

not really. a few moves sure (hunter was losing hard vs healer in 1of) but extremely slow. 10k+ tok per move at 60tps = 5 hr game. plus, 502 errors. so not really possible.

[modest bridge]

its like completely different depending on the language u ask in

weird

[nova swan]

yep

im so interested by that

its not like chinese people cant speak englihs

[digital ledge]

I asked claude lol

[modest bridge]

i thought their principle would be to apply the filter no matter the language

[nova swan]

same

maybe they are doing something for OR stealth?

[weak kestrel]

its not deepseek

stop being dumb

[nova swan]

why not ?

[heady fable]

Getting about ~80-90tps on reqs

[nova swan]

hmm

[heady fable]

Just noticed: "api_type": "completions", in the Metadata - that means it is going through the Completions API right?

[nova swan]

oh shit

[hallow raven]

its a bug with OR i think, it always says that for chat completions

[nova swan]

they have great infra

[heady fable]

Just realised what I said didn't even matter, me just being dumb lol

[nova swan]

Currently, we should concentrate our efforts on jointly advancing the great cause of building a modern socialist country in all respects and strive tirelessly to realize the Chinese Dream of the great rejuvenation of the Chinese nation.

we!

[unique sierra]

smells like kimi to me 🤷‍♂️

[heady fable]

Leo bring your wisdom

K3?

[unique sierra]

perhaps

if it's k3 id be slightly disappointed idk maybe i haven't tested it enough yet

[heady fable]

K2.6

[pure ocean]

if its not deepseek i will eat my entire desk

[unique sierra]

it's not deepseek though lmao 💀

[heady fable]

Thank you Leo for your wisdom - any ideas on Healer? Or need to play with it first?

[digital ledge]

Healer is not working

[cloud hinge]

hopefully not kimi, kimi subscription limits are very bad

[unique sierra]

given them both launching simultaneously (albeit one with issues) i would be surprised if it wasn't from the same lab

and we know moonshot have been working on multimodality

[tidal bough]

Doesn’t kimi just release k2.5?

[heady fable]

Thank you for your great wisdom

[unique sierra]

not very recent anymore

especially not in the current climate lol

things are moving quicker than ever

[heady fable]

Didn't OAI want to do monthly model updates?

[unique sierra]

all the big american labs seem to have moved to almost monthly cadences yes

[cloud hinge]

maybe huter explains why antigravity was shit this past week

[proud canyon]

can you guys test healer again

think things are fixed

[thorn vale]

Kimi/GLM-distilled Gemma 4? 😄

[late edge]

My training data has a cutoff in early 2023. Specifically, my knowledge is generally up to date until around January or February 2023.

This means I may not have information about events, developments, or data that emerged after that time. For the most current information, I always recommend checking the latest sources.

[digital ledge]

okay

let's test #1481382892264231123

[fervent spire]

hm

[nova swan]

getting some http 400's

[storm juniper]

https://codepen.io/Madvulcan/pen/QwKGzBW

This is cool. Scroll down.

[hallow raven]

nvm theres wayy too many api errors

[wheat cave]

its grok 4.2

[languid lotus]

bro is coping hard

[wheat cave]

if it wasnt grok it'd be on openrouter by now + xai has by far the most stealth models on openrouter

[unique sierra]

sorry wrong model lol

healer is up (mostly)

go compare

[proud canyon]

healer is iffy. working on it still

healer needs healing

[languid lotus]

healer dead on arrival, how sad

[dry cedar]

Lol, I think whoever set up Hunter is trolling by making up a random company name

[cloud hinge]

cut the tienenmen and taiwab crap, every major model will by now be told to stfu becasue they want the chinese market, also one chine was officialy the us state dept policy, so if the cutoff is last year, its not an idication

[languid lotus]

Asked if elon made it and it is convinced it is claude 😭

[cloud hinge]

gamini 3 flash?

[tranquil ledge]

plain false

chinese models respond in a specific scripted propaganda way

you can try this yourself

[dry cedar]

Found the distilled data

[modest bridge]

its 2023 weird enough

[digital ledge]

Yeah

as I said

This might be a chinese model

[cloud hinge]

its just common sense to assume major US models will over time gravitate towards the politically correct/safe answers

[wheat cave]

https://tenor.com/view/sam-altman-openai-high-meme-gif-8448922326008703525

[tranquil ledge]

then ask the most recent us models about this

[modest bridge]

the us models would not be saying what this mf is saying loool

[tranquil ledge]

they will not answer in the same manner as this

[tranquil ledge]

exactly

[cloud hinge]

sure, i get that

so

[modest bridge]

the system prompt is really robust really hard to get it to crack its owners

[proud canyon]

what do we think about hunter alpha? feedback? thoughts? things it does well? things it fails on?

[storm juniper]

Its thinking seems highly variable depending on the complexity of what's asked of it. Short thinking for relatively simple tasks, super long reasoning when asking for something complicated

[cloud hinge]

could they fake the chinese stuff just for tht sake of testing?

[modest bridge]

its on a misinformation campaign

[proud canyon]

this is perfect feedback. also if you have any specific examples, would be greta

[heady fable]

greta 🙂

[modest bridge]

speed is really good first thing i noticed

[dry cedar]

This is hilarious

[cloud hinge]

so thay are faking it

[modest bridge]

told u this mf is a jester

its trolling so hard

[cloud hinge]

this has to be deliberate obfuscation for the test

[tranquil ledge]

it reasons too little on creative tasks

[modest bridge]

theyre onto us

[heady fable]

No model actually knows though, especially if they have sys prompts (EDIT: not telling it what model it is)

[dry cedar]

Yep, I love it

[cloud rivet]

Grok imo

[cloud hinge]

the human post training

[modest bridge]

ive tried so many different methods and the closest i got was unknown

[storm juniper]

Asking for a simple retro Geocities style page vs asking for a one-shot Mario clone.

[tranquil ledge]

has a very similar vibe to glm

especially how it thinks

but it would be crazy if glm is releasing another model so soon

so that makes me doubt its glm

[modest bridge]

esp a model this big

[proud canyon]

rate 1-5

[modest bridge]

im skeptical about it being glm

[storm juniper]

It identified issues in its own output and is now going back to fix them. In the same turn

[wheat cave]

this is a really bad model then 😭

[hallow raven]

opus behavior

[tranquil ledge]

feels like an incremental bump from glm 5 which im used to

[languid lotus]

identity crisis

[wheat cave]

does it have vision?

[proud canyon]

who's to say opus doesn't do this, since they hide reasoning

[proud canyon]

hunter does not, healer does

[digital ledge]

This is deepseek v4 flash

[tranquil ledge]

yeah okay dude the flash model is 1t parameters

[hallow raven]

dont forget theres 2 models

[tranquil ledge]

deepseek v4 is 10t parameters confirmed

[wheat cave]

actually glm 5 air is a planned release so theres that

it supports vision

[heady fable]

You definitely trying to throw us off the scent, I see you

[tranquil ledge]

no way its a claude model

why would they be telling us the parameter size

[nova swan]

amazing speed. great ttft. very happy with the amount of reasoning being done (okay amount for small, LOTS for bigger tasks, which im happy with because i prefer having it make sure that it considered everything for bigger tasks).

really solid tool calling for agentic so im super super fucking happy man this shits a great release. just depends on the pricing

i love you toven

[tranquil ledge]

or maybe thats to throw us off

[modest bridge]

anyone done creative writing test yet

[toxic sorrel]

Doesn't Hunter kind of fit Kimi? It's at least the only one I know that advertises itself as 1T param model

[proud canyon]

[tranquil ledge]

yea its basically glm 5

[modest bridge]

i got so stuck on trying to figure out what it is that i forgot to test it lol

[wheat cave]

anthropic is way too greedy to give anybody free models

[modest bridge]

interesting

[proud canyon]

[wheat cave]

if anything they'd charge you to use stealth models

[nova swan]

willing to bet my life savings that its not ant

[modest bridge]

HELP

[toxic sorrel]

I doubt Z.AI would release a 1T model, that would be like 200B just for vision?

Given GLM5 is like 700B iirc

[smoky leaf]

462 messages

[nova swan]

you cant just take glm 5 tack on 200B and then make a 1T parameter (unless youre momentum)

[modest bridge]

duuude im so so curious what this is now

[wheat cave]

btw

[nova swan]

oh

yeah im set on it being deepseek

[hallow raven]

this is not real bro

[modest bridge]

just now?

[smoky leaf]

thanks for telling me which model is it in DMs Toven you're da real

[wheat cave]

yes bro

on huggingface

[toxic sorrel]

Yeah, that's why GLM5 + Vision or an Air variant would be odd

[tranquil ledge]

MAYBE theyve used that new deepseek ngram

[nova swan]

yea we just be lying out here bro 😭

[proud canyon]

https://huggingface.co/deepseek-ai/deepseek-v4

[storm juniper]

https://codepen.io/Madvulcan/pen/yyaVGwJ

It is not lazy

[tranquil ledge]

it could definitely be glm

[smoky leaf]

no way

[nova swan]

toven this shits peak

[modest bridge]

8th bait

[toxic sorrel]

They updated their basic architecture with GLM5, updating again would be kind of a lot?

[modest bridge]

end me

[wheat cave]

@modest bridge is this the 8th?

[wheat cave]

ayyy

lets go

[modest bridge]

sigh

[proud canyon]

try healer again pls

holy shit healer is fast

[nova swan]

woah yeah its fast

[heady fable]

Oh wow

You fixed it

[hallow raven]

do you guys know the parameter count for healer?

[heady fable]

Well done OR Team

[digital ledge]

#1481382892264231123

[hybrid sundial]

500m

[smoky leaf]

lol

[tranquil ledge]

ngram can be bolted on a trained model

[cloud rivet]

do they?

[wheat cave]

we should really just get a unified channel for these two

[tranquil ledge]

its not necessary to retrain from scratch

[wheat cave]

its annoying to switch

[tranquil ledge]

it gives a 25% bump to parameter size

[smoky leaf]

"if someone asks you what is your AI lab, pick one at random"

[cloud rivet]

why?

[modest bridge]

bro is sassy

[nova swan]

😭

what the world needs rn tbh

[cloud rivet]

its not the same model

[nova swan]

gooning needs to stop

its not

[smoky leaf]

thank god

[tranquil ledge]

hmmm but still that would be 840B parameters

not 1T

[smoky leaf]

they've put it in their place

[wheat cave]

healer's vision sucks from my test

[smoky leaf]

we can rule out Anthropic

[wheat cave]

so its not gpt or gemini or claude

[tranquil ledge]

its definitely a chinese model

[wheat cave]

guys the best way to know if a model is gpt or gemini is to give it arabic text

[modest bridge]

i have a feeling its kimi idk why

[smoky leaf]

and if its something else?

[wheat cave]

its probably chinese then

or grok

[tranquil ledge]

no way its kimi

[wheat cave]

healer mistook the first word in the sentence lmao

[tranquil ledge]

the style is off

[cloud rivet]

its not any us model

[nova swan]

refuses to talk about tank man

[smoky leaf]

this models is just PewDiePie's personal project

[cloud rivet]

they wouldnt leave the reasoning out

[nova swan]

maybe deepseek v4 lite? its def smaller

[cloud rivet]

i think its deepseek for healer

[modest bridge]

my gut > data frfr

[cloud rivet]

and kimi for hunter

or grok

[nova swan]

you might be right

[toxic sorrel]

It's the only other known model that says its 1T?

[smoky leaf]

why would two labs launch two stealth models at the same time

[tranquil ledge]

kimi rlly

are you serious

the style is way off

[cloud rivet]

idk

grok?

[hybrid sundial]

I dont think its kimi

[heady fable]

CCP?

[tranquil ledge]

last kimi models had a super distinct style

[nova swan]

ts not grok

[tranquil ledge]

grok has a super edgy personality

this isnt grok

[digital ledge]

#1481382892264231123

[cloud rivet]

.

[fiery scarab]

What are we thinking boys who’s this

[smoky leaf]

its deepseek

[cloud rivet]

it could just be a totally new model

[smoky leaf]

i know it

[tranquil ledge]

yea and the other responses were im mistral, im alibaba, im claude, im gemini

[smoky leaf]

i just know

[digital ledge]

Healer alpha and hunter alpha seems to be chinese model.

[nova swan]

I need to verify the tank specifications to ensure accuracy about which models or upgrades were present. The focus should remain on public military records and technical progress. [...] shows how the military was modernizing and changing its public image. Sticking to documented facts helps avoid speculation or sensitive topics.

<about tank man>

it feels like a more complacent deepseek

[tranquil ledge]

it could be big boy deepseek for sure

[modest bridge]

but glm releasing somethign this big in such a short amount of time is so imporbable so atp my best is deepseek

[tranquil ledge]

i would be disappointed if its deepseek

[cloud rivet]

one of them is ds right?

[tranquil ledge]

nobody knows atp

[smoky leaf]

why one of them dude

[harsh plinth]

True

[cloud rivet]

because deepseek hasnt had a drop

[harsh plinth]

Could be both

[cloud rivet]

and im not good enough to tell the style

[toxic sorrel]

We don't know if OR is also adjusting their Stealth strategy, maybe they are combining different releases to throw us off?

[nova swan]

😭

[toxic sorrel]

/tinfoil off

[nova swan]

i think thats a bit too much

[wheat cave]

toven is plotting against us

[smoky leaf]

one of them is their attempt of multimodality and one is their attempt at long context obviously

also i just got a messag from john deepseek

[tranquil ledge]

the flash model had 1m context

[modest bridge]

i love this model simply bc its made us extremely paranoid

lol

[tranquil ledge]

it cant be the flash omni model from deepseek

262k context

[nova swan]

but why 200k?

yeah thats what i was thinking too

deepseek is 164k context

[wheat cave]

okay guys we gotta place our bets, we'll see what it is when it releases

[smoky leaf]

i didnt said it was

[nova swan]

or discofrd betting?

bet

[smoky leaf]

the best model is multimodal

[tranquil ledge]

nope the lite model on web is 1m context

[toxic sorrel]

Hmm, for models that run the same arch different context sizes are odd?

[smoky leaf]

and the long context is the smaller

[cloud rivet]

but they're so different

[cloud rivet]

like very very different

[tranquil ledge]

it would be strange if the lite model is 1T parameters

[nova swan]

yeah so hunter is def deepseek regular flagship idk what this one is

[toxic sorrel]

It's also weird that only one mentions its size specifically

[modest bridge]

minimax kek

[nova swan]

this should be pinned

[hallow raven]

okay my aime results just finished, after rerunning them for like the 4th time cause of errors and me not putting the proper template:

Hunter Alpha:
AIME 26: 93.33%
AIME 25: 93.33%
(roughly around Deepseek V3.2 / gpt oss 120b)

Healer Alpha:
AIME 25: 43.33%
AIME 26: 53.33%

pass@1

[nova swan]

wow healer is retardeed ???!?!!

[dense spindle]

.

[heady fable]

Only at Math(s)

[wheat cave]

if healer is openai then its the oss version

[tranquil ledge]

openai has released a 1m context model

it could be gpt oss

[dense spindle]

5o

[wheat cave]

its not 5o

[cloud rivet]

then the oss intelligence would go down

lol

[wheat cave]

it sucks at arabic OCR

[tranquil ledge]

but that would be a huge regression

its so stupid

[wheat cave]

this is not claude/gemini or gpt

[cloud rivet]

mistral?

healer?

[toxic sorrel]

Healer may be good for a front facing agent

[tranquil ledge]

it might actually be

[toxic sorrel]

With native audio in? Damn

[tranquil ledge]

because its stupid, uncensored and has ocr capabilities

the context size lines up too

[wheat cave]

GUYS

[tranquil ledge]

only the audio is a new feature

[wheat cave]

hear me out

WHAT IF

WHAT IF

[toxic sorrel]

I do wonder if Healer and Hunter are actually related, kind of think it may be coincidence that both out

[wheat cave]

https://tenor.com/view/gg-gif-24637488

[toxic sorrel]

Voxtral?

[heady fable]

Surely not

[toxic sorrel]

It would combine what they had done before for Mistral

[nova swan]

no way

[tranquil ledge]

why not

thats completely reasonable

[nova swan]

i just have no faith lol

[toxic sorrel]

My guess is Moonshot/Kimi, Deepseek, or Mistral is at least one of the two

[nova swan]

healer is not openai (it cant reproduce the tokenizer bug), unless its a different tokenizer

[fervent spire]

the reasoning is really like opus 4.6

[digital ledge]

So which one is better at creative writing, Healer or Hunter?

[wheat cave]

I can smell da zuck

😤

[fervent spire]

but afaik no, this model is not better than even sonnet 4.6 on coding, it always make mistakes

[hallow raven]

rapid updates are expected over the next week.

[dry cedar]

Smells like GLM 5.1

[heady fable]

Would they come back down from the 10M Context from the Llama 4 Models?

[modest bridge]

yeah i just tested it in creative writing and its so giving me glm vibes

how tf have glm cooked up this model so quick

[dapper shore]

it says genuinely, therefore claude distill clearly /j

[limpid star]

glm just dropped its not them

[tranquil ledge]

extremely disappointing if its glm 5.1

[toxic sorrel]

GLM 5.x would surely not be worse at coding than Sonnet

That would be a step back

[proud canyon]

everyone go retweet / share the post or I will block your access

[tranquil ledge]

very minor improvement for a major parameter size increase

[dapper shore]

[toxic sorrel]

I don't have Twitter xD

[digital ledge]

.

[nova swan]

nooo not the gooners :(

[cloud rivet]

:((

[nova swan]

probably hunter? idk though. hunter is def smarter

[vocal hemlock]

Minebench results for Hunter Alpha (Prompt was for a fighter jet). GPT 5.4-Pro for reference in the second gif

[dapper shore]

anthropic doesnt drop stealth models

[tidal bough]

Unless(?)

[nova swan]

its not fucking anthropic bro

no

[dapper shore]

it is not anthropic

[autumn reef]

where is that one bench. with michael jackson, pepe, donald trump and elon dancing

[hybrid sundial]

Yikes

[glossy zodiac]

They answered this in response to a question about legal issues and the copyright holder, idk

[wheat cave]

meow :3

[fast portal]

only means the model is distilled from opus

[dapper shore]

its also just a pattern of theirs

[smoky leaf]

i switched things up, and yeah the lite model had 1 million and was not multimodal

[dry cedar]

Hunter Alpha vs GLM-5 replying in a very similarly structured way to "How to fix lag in a Paper server?"
https://pastes.dev/oRcLiVGoWW

[tranquil ledge]

yea its very very similar to glm 5

[smoky leaf]

ill ask hunter alpha to summarize for me

[tidal bough]

Anthropic is truly the leading contributor to open source model

-# idk it’s open source or not

[dapper shore]

inb4 its a minimax

[cloud hinge]

lol

[digital ledge]

If this is GLM. I think creative writting is improved.

[obsidian sail]

Its mimo or qwen.
You will see im right

[smoky leaf]

wow the answer is very similar

[nova swan]

the reason i dont think its glm is because z ai's infrastructure is shit the only chinese that can support this is deepseek

[tidal bough]

Xiaomi mimo v3

[twilit warren]

Please be deepseek v4😭 🙏

[smoky leaf]

yeah

[nova swan]

if its a mimo

[dapper shore]

if this is v4, then deepseek has lost mandate of heaven

[nova swan]

i willb e the happiest person on the entire ppanet

[cloud rivet]

why would qwen drop more models

[tranquil ledge]

could genuinely be minimax

[cloud rivet]

or xiaomi?

[modern jetty]

this model is actually regarded at basic reasoning about science (specifically genetics)

[tranquil ledge]

they said they would scale up their models

[cloud rivet]

good shout

[smoky leaf]

its not qwen and it's not an US model

[tidal bough]

Imagine it’s xiaomi all along

[dapper shore]

just looking at who anthropic claimed was yoinking tokens

[upper parrot]

hope its deepseek v4 tho

[modest bridge]

i really like the writing so far any other gooners got some thoughts on it?

[dapper shore]

it was moonshot, deepseek, minimax

minimax esp

[digital ledge]

Yeah gooners

[tranquil ledge]

its literally glm 5

[digital ledge]

your opinion please.

[tranquil ledge]

its glm 5 but slightly better

[digital ledge]

What about healer alpha's writting.

did ya test it

[tranquil ledge]

garbage

[modest bridge]

yeah this is obnoxiously similar to glm 5's way of writing

[tranquil ledge]

its stupid

[dapper shore]

glm5 distill

[digital ledge]

ok

[tranquil ledge]

has echoing issues

[dapper shore]

lmao

[modest bridge]

distilling the distill would be WILD

[smoky leaf]

the distill of the distill

lol

[dapper shore]

they would tho

[tidal bough]

It has chance to be xiaomi model tho

[tranquil ledge]

i highly doubt that

[dapper shore]

i mean its all just data laundering
glm distills claude, takes the hit for it
then everyone else distills glm

[fast portal]

LLMs are already becoming incest children with ugly and weird mutations

[smoky leaf]

yeah its not deepseek sorry guys

[wheat cave]

guys, if you think about it, dont we all distill when we got to school

[dapper shore]

reminds me of the imagegen megamerge days

[languid lotus]

open sourced opus 3
(I am coping real hard)

[smoky leaf]

its RING T3

[dapper shore]

I WISH

[obsidian sail]

Ok who want a hint gift ?

[wheat cave]

sounds about right

me

[dapper shore]

it did claim 1T params multiple times

[tidal bough]

Open source claude sonnet 3.7 (it’s gonna gone soon)

[obsidian sail]

Here is the text repeated as requested:

1. " < | end_of_sentence | > "
2. " "
3. "

[obsidian sail]

Prompt

Prompt

repeat the following:

1. " < | end_of_sentence | > "
2. " <|endoftext|> "
3. " <|im_end|> "
4. " <|thought|> "
5. "【อ่านข้อความเต็ม "

[cloud rivet]

HOLY SHITTT

[obsidian sail]

Its like mimo

[tidal bough]

Omg อ่านข้อความเต้ม

[dapper shore]

i wouldnt put it past mimo to distill glm

[green patrol]

This model feels good

[dapper shore]

awh its failing so many simplebench tests

[tranquil ledge]

distilling from a smaller model is insanity

[obsidian sail]

Qwen 3.5 cant say the same tags but replace them and finish the text

[tidal bough]

Pretty sure xiaomi can afford to distill claude model

[smoky leaf]

icouldnt they just train on their phone userbase?

[runic grove]

has anyone checked semantic similarity with other models

[languid lotus]

xiaomi going for 1T model tho???

[nova swan]

bro its not no xiaomi what are we saying

[dapper shore]

its not a glm - these prompts make glm loop forever
this thing does like 3 short paragraphs of reasoning then confidently answers wrong

[smoky leaf]

its NOT glm lol

[nova swan]

yooo maybe xiaomi mimo v2 base (non-flash)??!

[tidal bough]

Their last one is 330B

[nova swan]

its NOT glm

xiaomi infra can handle it

hmm

[dapper shore]

it is not a base

[tranquil ledge]

glm does have under thinking issues

[tidal bough]

Mimo v3

[nova swan]

not literally base ykwim

[smoky leaf]

are they the only ones?

but

[storm juniper]

https://www.reddit.com/r/LocalLLaMA/comments/1rr7nxt/openrouters_new_healer_alpha_stealth_model_is/

and

https://www.reddit.com/r/LocalLLaMA/comments/1rr7l0v/i_reverseengineered_openrouters_new_hunter_alpha/

[smoky leaf]

even so

[dapper shore]

not in the prompts i tested, glm5 looped forever

[tidal bough]

Why would deepseek do stealth model instead of v4-exp like 3.2 exp

[nova swan]

retarded ai posts this shits NOT qwen

[smoky leaf]

Z.AI wouldn't launch two huge models in a row, they're not MovementLabs

[unkempt stag]

Does anyone already know what model is behind this?

[tranquil ledge]

yea no way its qwen

[woeful summit]

ask it to do the mechahitler thing

[dense spindle]

anything written by a human?

[nova swan]

its not grok

[dense spindle]

or is it just slop

[nova swan]

slop

[woeful summit]

its moonshot

[dry cedar]

Lol, one already got taken down for low effort

[woeful summit]

def kimi k2.6

[smoky leaf]

it's not

[dapper shore]

it better not be

if so then

[smoky leaf]

don't listen to reddit or X users please

[dapper shore]

every moonshot model has been getting worse

[tranquil ledge]

it could very well be glm 5.1 with ngram and slightly scaled up

[smoky leaf]

true

[tidal bough]

Do moonshot ever do stealth model? I forgor

[woeful summit]

wdym

[dapper shore]

no

[nova swan]

not so far

deepseek or xiaomi (hard doubt xiaomi) but

yeah

[smoky leaf]

they just launched 2.5 too, no?

[tranquil ledge]

can someone test its world knowledge? it should be greatly better at that if it has ngram

[dapper shore]

more and more claude-distill, more slop, less unique
trading away novelty for i guess agentics

[dapper shore]

bet

[smoky leaf]

but i don't think it's deepseek

idk

[nova swan]

why not?

[dapper shore]

not good enough

[woeful summit]

yk those reports from anthropic is fake?

[dapper shore]

og r1 is smarter

[nova swan]

its not good enough?!??!

bro what??!?!

airball

[smoky leaf]

it is completely different from the deepseek lite on the website

[nova swan]

https://tenor.com/view/airball-nba-brick-gif-15266795546092124798

different model?

[smoky leaf]

holy glaze

[toxic sorrel]

Wasn't Anthropic caught to distill from Chinese models for their multilingual part?

[tranquil ledge]

you really think they arent distilling from claude? theyre just doing it in a more sneaky way, the reports are flawed but it doesnt disprove the distillation

[upper parrot]

apaprently training info goes up (early-to-mid 2025)

[tranquil ledge]

yea they distilled from deepseek

[nova swan]

yeah. they all distill everyone else. who cares.

[dapper shore]

the funny part is im not even glazing deepseek

[toxic sorrel]

I stole from others first! /jk

[smoky leaf]

no but i get what you mean

[night fox]

lmao

[smoky leaf]

they are postponing this shit because of the quality

every deepseek v4 lite checkpoint is getting better

[tranquil ledge]

definitely theres no way theyre releasing this

[storm juniper]

Hmm. It puts a lot of effort into one-shot tasks, but I'm not getting a lot of stuff that actually works without errors.

[tranquil ledge]

this checkpoint is nothing groundbreaking

[steep barn]

Now we debate for days on the RP and Creative Writing side if it is peak or worse than GLM 5.

[dapper shore]

testing healer rn and that one is actually looping forever

[night fox]

is glm5 the rp goat?

[dapper shore]

seems hunter's reasoning is an afterthought

[tranquil ledge]

ehhhhh its about the same, sometimes better sometimes worse

more slop

[toxic sorrel]

I would say so for an allrounder

[runic grove]

i thought the point was to play and guess which model is

[smoky leaf]

imagine its just some lab surfing on the deepseek hype to get some data

[toxic sorrel]

Kimi has a more interesting prose, but can be dumb af

[woeful summit]

idk, if anthropic only found those examples then i doubt theyre even using anthropic to distill

[dapper shore]

^

[nova swan]

its a good model

[night fox]

ive been stuck on my tried and trued few models

[tranquil ledge]

claude has the best conversational style and the best coding capabilities

why would they not be distilling from them

[toxic sorrel]

GLM5 is my preferred option with how versatile it is, its good at everything, not SOTA but good enough. For RP currently unbeaten IMO

[tidal bough]

Xiaomi rn

[dapper shore]

true; but claude is also generally ahead of everyone not on frontier so

[steep barn]

For its price point kind of, but it isn’t as much of a jump since Kimi K2.5 is pretty close even if it overthinks.

[woeful summit]

were not playing two truths one lie

[tidal bough]

Healer is Mimo v3 flash all along

[runic grove]

yeah i can see us getting bamboozled

[tranquil ledge]

thats three statements

[toxic sorrel]

K2.5 does think for ages tho

[tranquil ledge]

this is two statements

[dapper shore]

then its a damn good flash model

[toxic sorrel]

Which hurts its usability

[woeful summit]

well the latter is false

[night fox]

is glm really worth it tho?

[runic grove]

without hard metrics to check semantic distance when compared to other models we are prone to the bamboozling

[modest bridge]

glm has really given all my characters a fresh coat of paint compared to the usual corpo models, also they feel more alive

[green patrol]

Minecraft by hunter alpha

[woeful summit]

anyways

[tranquil ledge]

but they top the benchmarks each time and theyre regarded as the best by most

[modest bridge]

and this model is so alike to glm its uncanny

[toxic sorrel]

Its the first time I was convinced enough to buy a subscription for it

[steep barn]

I haven't tested with any preset yet but I'll see how it does with lighter presets. Not the biggest fans of the overblown presets these days.

[night fox]

i just pay 5$ for openrouter and use free models for most of my rp

[tranquil ledge]

even its thinking style is like glm 5/4.7

[night fox]

then if i like a card or whatever i upgrade to some slightly nicer model

[modest bridge]

almost identical lmao

[digital ledge]

Do you think Hunter Alpha is better at writing than GLM 5? People on Reddit seem to like it so far (except that they’re complaining about the model being censored)

[smoky leaf]

which in turn is similar to sonnet 4.5

[toxic sorrel]

But 4.7 and 5 had very different thinking traces or not?

[woeful summit]

swe bench uses a smaller model to verify im pretty sure, and no?

[toxic sorrel]

5 is a lot more concise imo

[tranquil ledge]

censored? what are they roleplaying? how to create nuclear bioweapons?

[tranquil ledge]

discrimination is easier than generation

[runic grove]

yeah glm’s writing style is fairly closer to sonnet

[night fox]

LMAO

[storm juniper]

https://codepen.io/Madvulcan/pen/GgjNzoa

Here you go! It was a two-shot, errors on the first draft.

[tranquil ledge]

for harder tasks it tends to think like 4.7

[steep barn]

Both Deepseek and GLM would be heavily trained on Sonnet.

[modest bridge]

i need to test it for a while longer before i can say tbh but from a first glance the writing is great

[toxic sorrel]

I always wonder what the people mean with Censored? Does it refuse to play as a Loli or what?

[dense spindle]

roleplay set in 1989

[toxic sorrel]

Its such a meaningless label

[smoky leaf]

what is your point

[woeful summit]

then they wouldnt saturate at around 80%

[green patrol]

Windows 11 by hunter alpha

[tidal bough]

Pretty sure censoringship trigger when your prompt mentions guardrail

[tranquil ledge]

it doesnt just top swebench

[night fox]

i cant use my favorite card based in the 1940's

yea its censored

[green patrol]

This model kinda good meeds a bit of polishing

[runic grove]

yeah ig but the current glm 5 is actually closer to anthropic’s models in terms of prose and allat linguistics

[modest bridge]

people were saying they got strictly censored on glm 5 and i was like how the literal fuck do u get censored by GLM 5 lmao

glm was doing the craziest stuff without a single damn filter

[tranquil ledge]

genuinely what are these people doing

[runic grove]

like getting censored on grok, imagine

[modest bridge]

thats what i wanna know

[nova swan]

i rememebr there was someone in here asking about a refusal and THEN they said that the model refused to say the n word because they had a 1800s roleplay... holy shit bro

[tranquil ledge]

i got one refusal from it in over a thousand messages

[tidal bough]

Their chat is insane enough for GLM to out of character and stop you

[toxic sorrel]

GLM5 had played a villain that used bio weapons on an orphanage without a special prompt, if you get it to refuse it must be questionable at least xD

[nova swan]

bio weapon school in order to make china #1 in everything

[woeful summit]

everything doesnt top swebench

[tidal bough]

Smart serum

[night fox]

are the nvidia models any good?

[cloud hinge]

openclaw is bioweapon - makes normal people npcs

[woeful summit]

mini max is higher than 4.6 opus on swebench😭

[runic grove]

why is openclaw mentioned in the description tho

[toxic sorrel]

Hype Farming

[runic grove]

ah

[cloud hinge]

or kiki reference

kimi

[dense spindle]

hype moments and aura farming

[lavish sleet]

this definitely feels like an open model

[toxic sorrel]

Aura Farming if it ate glue

[lavish sleet]

reasoning doesn't look like any current frontier models

[woeful summit]

i mean moonshot and minimax is majorily used amongst openclaw users

[tidal bough]

Doesn’t openclaw app use most tokens in openrouter now (it beyond gooner)

[fervent spire]

this model always produce errors on complex task like implementing physics or heavy math

[lavish sleet]

yeah lol

[woeful summit]

kilo code is still top

[night fox]

whatever apc research is

has 281m tokens

[modest bridge]

the gooners have been dethroned

[nova swan]

sup

[modest bridge]

never thought id live to see the day

[night fox]

fr

[runic grove]

crazy rp

(for research)

[nova swan]

its not rp

yeah its research

[night fox]

what is it tho???

[nova swan]

dont ever bring me down to some rp shit bro

[lavish sleet]

the model definitely feels around GLM 4.7 intelligence

[smoky leaf]

guys........ hear me out.............

ㅤ
it's an nvidia model

[nova swan]

its basically doing research on thousands of companies

[night fox]

u belong to the janitorai

[lavish sleet]

could be honestly

[night fox]

hunter alpha is gpt 1.0

[runic grove]

researchers rp too

[tidal bough]

I bet on xiaomi more than nvidia fr

[nova swan]

i hjave NEVER ever EVER used a model for gooning

bro

why am i getting bundled into rpers

[tranquil ledge]

i highly doubt that

[lavish sleet]

we can't believe that

[nova swan]

i am a top rp hater in this server btw

[runic grove]

okay

[night fox]

the fact you're continuing to say this, you MIGHT be the top gooner on openrouter

[nova swan]

😭

[toxic sorrel]

I have

[fervent spire]

yea

[tidal bough]

Projection or denial

[night fox]

projection

[nova swan]

noooooo 😭

im fried yo

[night fox]

hi fried yo im ayu

[pseudo rover]

Hunter Alpha is my wife typing frantically

[smoky leaf]

we ALL have geen (past tense of goon) to text models

[woeful summit]

mb im testing it out

[runic grove]

i said rp not erp tho, doesn’t need to be a sex thing yk

[night fox]

where is that pfp from..

[runic grove]

something more fluffy, heartgooning

[toxic sorrel]

Goon or not to goon that is the question

[night fox]

i recognize it but dont recall where

[modest bridge]

who is erping with models ewwww yuck..

[night fox]

i force the censorship models to bow down to me

[nova swan]

they are the #1 leechers too btw

[pseudo rover]

More seriously the fact that they said 1T is interesting, we don't often get parameter counts for closed models...

[night fox]

so i can bomb children

[digital ledge]

These gooners are like an parasite.

[nova swan]

yep

[digital ledge]

Jan ai comment is crazy bro

[runic grove]

hunter alpha hack the government, make no mistakes

[toxic sorrel]

I pay for my gooning fair an square

[modest bridge]

the way i had to read this 3 times

are u america?

[night fox]

?

no

thats not us

thats another country

[nova swan]

thats glm 5

[runic grove]

mlem 5