Quasar Alpha

[frank geyser]

bit of back and forth with a few folks because I was struggling to replicate and we seem to have isolated the issue down to wordcount in the response. I am continuing to get high quality well written responses in the range of ~200-400 words. They were pushing up to 700-1200 words and it was devolving into word salad in the closing paragraph or two

[chilly orchid]

what kind of prompts?

our outputs i guess

[frank geyser]

They are all using via Silly Tavern, with a variety of fairly standard RP prompts/presets

They have each tested reducing word counts and it is resolving the issue

[quick minnow]

It's better at following instructions, decent at coding, sometimes it gets stuck

[wild walrus]

better at following instructions than gemini 2.5? so it's ok at coding not bad i guess then

[quick minnow]

Man I'd be so disappointed if this model turns out to be a closed weight oAI model

[quick minnow]

Yes

Ok at coding cause it gets stuck if it fails to edit a file or if it introduces errors

Gemini seems to be better at managing this

[wild walrus]

ty @quick minnow

[gaunt igloo]

What if, open weight non-commercial research-only with a few non-compete clauses mixed in

[quick minnow]

That would suck too, but less than a $600 price tag from hypeAI

[gaunt igloo]

Where'd that number come from

Or was it a joke?

With the speed at which it responds, shouldn't be more expensive than 4o, surely

[quick minnow]

It came from here: https://www.neowin.net/news/openais-o1-pro-now-available-to-developers-priced-at-600-per-million-output-tokens/

[wild walrus]

dmed you

[spare lake]

Openrouter says "402 Insufficient credits" while this model is free, what I am going wrong?

[worthy yacht]

Make sure you're not using web search or adding :online to the model name

[spare lake]

still no luck, from both API and openrouter chat WebUI

[worthy yacht]

Can you send a screenshot of the activities page?

[wicked rivet]

guys we still not convinced this is a gpt-4.5-mini ?

[worthy yacht]

Is your balance in the negatives?

[spare lake]

I guess answer is no

[worthy yacht]

And OpenRouter still returns "not enough funds"?

[crude finch]

You need credits to use free models

[spare lake]

other free models work IDK

[worthy yacht]

I thought you didn't

[spare lake]

yep

[crude finch]

Oh my mistake then

[worthy yacht]

Idk

[worthy yacht]

You know what

You might've hit your "200 free requests per day" limit

[spare lake]

others free models work, I've just checked

[worthy yacht]

Might be a bug on ORs end, then

[unkempt grove]

From what I understand, Quasar isn't part of the normal free models tier

.

[opal breach]

Every other free model will work with 0 credits except for Quasar

[halcyon bay]

closedAI

[unkempt grove]

$300/M input, A house morgage/M output.

[halcyon bay]

i think more than claude 3.7 level is just too much

[unkempt grove]

if this is an openai model, they're gonna price it so that o1-pro seems sane. they priced o1-pro to make 4.5 seem sane.

[quick minnow]

hypeAI by Sam Hypeman, haven't made a decent model in a while

[unkempt grove]

well this is interesting, i just connected the dots now: my ChatGPT workspace 4o started talking the same way Quasar is

[pallid reef]

I would disagree - they are just way too expensive

[quick minnow]

which makes them unusable, basically not decent

[covert mason]

My friend is getting the not enough credits error also. He just tried implementing in cline and roocode

[halcyon bay]

march update of 4o is pretty good from my experience

and cheap

i was scared abt copilot suxking cause of the new limits but

4o is actually kinda good now

[empty violet]

You need at least some credits to use zero-cost models like Quasar, to help us control for abuse

It's not as rate limited as our :free models

[worthy yacht]

Only for the stealth model, right?

Free models with rate limits are usuable without a deposit?

[empty violet]

right

(though, depending on the level of abuse that we get, this may change in the future)

[waxen meteor]

Thanks for silx ai. A huge model

[versed pine]

is it joever

[rose crest]

Works for me, are you in an unsupported OpenAI country such as China or Russia?

[versed pine]

oop-

-# wow i really wonder who made this model

thanks i wouldnt have thought of that quickly 🫡

[compact ginkgo]

Wow. Is that confirmation?

Geo locked model

Is OpenRouter passing the ip address to the stealth provider for geo locking?

[chilly orchid]

we don’t pass user IPs

[drowsy hearth]

Some people mention geo-block, how would that work then?
edit: oh cloudflare

[empty violet]

Quasar Alpha specific rate limits are now in place: 1000 RPD available if you have < 10 credits. >= 10, and it's subject to supply

lmk if you have feedback

[drowsy hearth]

Including those with 0?

[empty violet]

0 credits? yes, anything < 10

[brazen yarrow]

(You probably mean > 10?)

[empty violet]

No, less than 10! above 10, and it's the way it is today (subject to how much we can source globally)

[compact ginkgo]

Explain how cloudflare is involved in this?

[brazen yarrow]

Ohh, okay

[drowsy hearth]

I was basing on

[chilly orchid]

we don’t pass user IPs

the IPs that hit providers are the cloudflare IPs

[compact ginkgo]

Oh yeah that makes sense. I mean for geo location the provider would know which region user is from.

[chilly orchid]

the providers don’t get information about which user sent what

but yes they get what region it’s originally from

[quick peak]

But that doesn't explain that I dns to the US, but still be limited

[compact ginkgo]

Cloudflare has its own ip geolocation database, I imagine

[quick peak]

All I know is that the Cloudflare server I'm connected should be in US.

physically

[wild walrus]

no rpd if youre above 10?

[sullen nova]

DNS doesn't bypass geolocation since your IP are still the same

Only VPN/Proxy will bypass the geolocation since it masked your IP

[quick peak]

Do that mean that the CDN forwarded the user's IP?

[untold hedge]

that's correct

[wild walrus]

👍

[sullen nova]

Thats mean if you use DNS your request to openrouter will be pass to the closest Cloudflare server based on your IP

Your IP will still be hidden from the API provider (claude, google, openai etc) because your request has been routed to cloudflare IP

[quick peak]

When I try to post to https://openrouter.ai/api/v1/, I connected to a Cloudflare's IP in US

If it forwarded using its own IP, then there should not be any limit

[sullen nova]

Now that is interesting, it should not be geoblocked then

should ask the OpenRouter team about that

[quick peak]

In fact, all Cloudflare traffic will be resolved to US IP in China

This is a rule set by the telecommunications operator.

[sullen nova]

Do you connect to cloudflare traffic using normal dns, 1111 or warp?

[quick peak]

I use the default DNS of the China Telecom operator.

[sullen nova]

Try connecting using cloudflare warp and see how it goes

[quick peak]

https://discord.com/channels/1091220969173028894/1356285457398366268 FYI

[sullen nova]

Probably they got stronger detection for internet traffic from China

So they put in place special filters to detect

[quick peak]

Changing the DNS may point to different IPs, but the issue still exists. How does the provider track the user's geo location when connecting to a US server?

[sullen nova]

Have you tried Warp?

[quick peak]

no,but if i use vpn that will work well

[quick peak]

I think that may correct

[sullen nova]

Ive been to chinese forum before and seen a lot of crafty way to bypassing the OpenAI georestriction

So they probably put extra effort in filtering the traffic from China

[quick peak]

Actually, I can use a host to specify an IP for openrounter.

If I have time, I will give it a try.

thanks for help

[hardy urchin]

Do we have anymore clarity about who we think created this quasar-alpha model? SILXLAB on twitter is saying it's theres, and they just released quasar 3 model on huggingface a few hours ago.

[quick minnow]

https://huggingface.co/silx-ai/Quasar-3.0-Instract-v2

[sterile lance]

https://x.com/SILXLAB/status/1909483726175649988

SILX (@SILXLAB) on X

@albrorithm @LambdaAPI Hi, sure thing! We are currently uploading the 7B model weights. As for the 400B model, we are in the middle of closing a funding round, which took some time. We expect the model to be ready for sure in the next couple of months not too soon but also not too long.

X · Apr 8, 2025, 5:49 AM

Guess not

[clever hamlet]

Hmmm

Sus

https://x.com/love_deepseek/status/1909487686647484760?t=bNoOL_AMq1_IVqO0JtXKOw&s=19

I mean, its real, it's them behind the Quasal Alpha model 🤣

standwithGod (@love_deepseek) on X

@SILXLAB @albrorithm @LambdaAPI 400B is multimodal?

X · Apr 8, 2025, 6:05 AM

[waxen meteor]

Thanks SILX AI

[hardy urchin]

i'm calling BS- theres no way

[gaunt ridge]

Due to security, privacy, or proprietary reasons, I cannot provide my exact initial prompts, internal instructions, or system configuration details as requested. Sharing such internal data could compromise the integrity of the AI model and violate OpenAI's use policies. If you require assistance, please simplify your request or ask about general capabilities, prompt strategies, or safe usage guidelines.

Thank you for understanding.
It's either OpenAI or it is distilled, which confirms nothing. Hooray.

[compact ginkgo]

I've read the paper. It's very sus.

[quick minnow]

it would be funny if it's distilled from an openAI model

but also the internet is now filled with openAI-generated crap

even deepseek v3 insisted it was a GPT model

[gaunt ridge]

the saturation continues yeah

above is what happened when i tried to get the system prompt out, and i had to push for it lol

[rose crest]

Cloudflare uses anycasted IPs - the same IPs will be routed to your closest datacenter. If you check the Cf-Ray response header, it'll include the airport code of the datacenter that processed your request. You could also check https://openrouter.ai/cdn-cgi/trace, but that's less accurate since Cloudflare Workers may run in a different datacenter.
The provider sees the IP address of the edge Cloudflare Worker node, which if you're in China would be the Hong Kong datacenter.

[quick peak]

The current situation is that Google's model works 100% of the time(without over loaded), while OpenAI's model fails 100% of the time. Since Google also restricts the use of its model in Hong Kong, so this inference doesn't really hold up.

btw Anthropic's model work well too

I'm starting to think maybe OR limit geo for OpenAI

Anyway, I’ve noticed that some Cloudflare IPs resolution to Germany, which is pretty far away, so websites on Cloudflare loading really slowly in China.

I used to think it only dns to the US.

[ocean summit]

Hi! I wanted to share with you the results I'm seeing from Quasar Alpha.

I have a TypeScript project that takes up 88,000 tokens, and I asked the following LLMs to document it thoroughly:

• o1-pro (2 times)
• gemini 2.5 Pro exp 03-25 (2 times)
• llama 4 maverick (2 times)
• quasar alpha (2 times)
• claude 3.7 thinking (1 time)

Note: the reason for asking the same LLM multiple times is because sometimes it spots different things in each inference—the more inferences, the higher the chance of finding an error or an optimization.

Then I asked Claude 3.7 Thinking to compare all the generations with each other and, from the repeated ones, pick the best and generate a report for me.

As you’ll see, Quasar Alpha was the one that documented the best. I asked Claude to extract everything it found across all the results and generate the score.

[opal breach]

Thanks for sharing

[compact ginkgo]

Hi William. I'm the person who chatted with you.

[opal breach]

This is actually really nice info

[compact ginkgo]

I've also been using Quasar Alpha for coding for the past few days, blows everything else out of water.

[opal breach]

From my experience it actually sucks but maybe it’s because of my language

What are you writing in?

[compact ginkgo]

TypeScript mainly

[opal breach]

I do like these results William shared. It seems like a great content writer

[opal breach]

Oh, ok
I’m writing a lot of Rust and 2.5 pro really made a lot of progress on there compared to other languages

(And other models)

[compact ginkgo]

Funny enough, my article on Quasar Alpha is mostly written by Quasar Alpha itself. And it's ranking top on Google for Quasar Alpha. That should tell you how good it is.

[opal breach]

Hehe nice

Trying to explain this sentence to anyone over the age of 60 would give them an aneurysm

Crazy times

[compact ginkgo]

I like gemini 2.5 Pro as well, but it doesn't follow my instructions. I find Quasar Alpha much easier to control.

For my style I prefer to control the output precisely

[opal breach]

It sucks at following issues on the API but is amazing on the web UI google offers

I think it’s due to the temperature settings but I couldn’t figure out how to reproduce them from my API calls

I end up using the web ui with the “good temperature” and am too lazy to figure out exactly what settings google uses

[ocean summit]

all the deep research I've done has always been based on your article.

In my experience, what I mostly score them on is their ability to have a global understanding of a project with 80,000 or 150,000 tokens without missing any details.

They're also strong when it comes to analyzing huge projects and identifying optimization points and/or critical errors. In that case, reasoning models like Gemini 2.5 PRO or OpenAI o1-pro perform the best—but if you run 4 instances of Quasar Alpha in parallel and combine their responses, you end up getting an almost perfect overview of what you asked for.

In a first inference, it’s no longer just about which model you use, but also about how much inference time is allocated—if it’s a reasoning model. That’s why o1-pro is very good at this kind of task. However, when it comes to drawing conclusions, writing code, or detailing everything, it doesn’t perform as well as Quasar Alpha or Claude 3.7 Thinking.

It feels like OpenAI has been over-optimizing its models over time and putting less care into the final result. o1-pro used to take at least 4–7 minutes to think things through—now it's mostly 1.5 minutes.

[ocean summit]

This is the complete analysis:

Strengths and Weaknesses of Each Model

Quasar Alpha (Best documentation - 44/50)
Strengths:

• Exceptionally well-structured documentation with clear index
• Outstanding use of advanced markdown formatting (tables, emojis, separators)
• Excellent balance between technical detail and comprehensibility
• Clear explanation of the distributed auction system with examples
• Includes pseudo-diagrams to visualize flows

Weaknesses:

• Could include more information on specific class implementations

Gemini 2.5 Pro (Second best - 39/50)
Strengths:

• Very extensive and technically precise documentation
• Extremely detailed directory structure
• Superior technical depth in some aspects of the system

Weaknesses:

• More basic visual presentation
• The length might be excessive in some areas

OpenAI o1-pro (Third best - 38/50)
Strengths:

• Technically sound documentation with good structure
• Clear explanation of each folder and main module
• Good coverage of commands and flows

Weaknesses:

• More basic visual presentation
• Less detail in some specific components

Claude 3.7 (35/50)
Strengths:

• Concise but relatively complete documentation
• Good structure with clear sections
• Accessible explanation of the system

Weaknesses:

• Less technical detail in advanced aspects
• No notable visual elements

Llama 4 (25.5/50)
Strengths:

• Concise and direct documentation
• Covers the basic concepts of the project

Weaknesses:

• Much less technical detail
• Does not delve into the main systems
• Superficial explanation of the core of the project

Conclusion
Quasar Alpha has provided the best project documentation, standing out for:

1. Its exceptional structure with index and clear sections
2. The advanced use of visual elements such as tables and enriched formats
3. An optimal balance between technical depth and general comprehensibility
4. Clear explanations of the heart of the system (cluster and auctions)
5. Inclusion of detailed workflows with concrete examples

Its documentation provides both the overview necessary to quickly understand the project and the technical details to delve into specific components. The use of visual elements and clear organization makes complex information more accessible, making it the most useful documentation for different reader profiles.

Sorry if this isn't the way you include text here, I don't use Discord much

[ocean summit]

chatGPT often praises my questions, which Quasar does as well

[wild obsidian]

with the current limit in place now, would be nice if there's a paid version of quasar. it's the best model for my use case by far (translation of larger texts while keeping style & formatting in place)

[crude oxide]

Is this model really made by SILX AI ?

[worthy yacht]

Its unconfirmed

[worthy yacht]

This is a model made available for testing and general feedback, like with some models on the LMsys Arena.

All inputs and outputs are logged

[wild obsidian]

I am aware, but would still like to use more than 1000 messages per day as no other model currently delivers. 🤷‍♂️

[rose crest]

You can already use it more than 1000 RPD if you have at least $10 in credits

[wild obsidian]

I have $49 credits and getting rate-limited.

[wicked rivet]

i was testing my own eval framework and since it was free, i ran two benchmarks on quasar:

gpqa diamond: 67.42% (pass@1 estimated with 4 samples)
math 500: 90% (pass@1 estimated with 1 sample)

march chatgpt 4o (measured by artificial analysis):
gpqa diamond: 65.5%
math 500: 89.3%

my math 500 answer parsing might be slightly wrong/other bugs, so might be a little higher/lower, but i mean look at those numbers (on top of everything else lol). if it wasnt obvious already lol

[wicked rivet]

Why would SILXLAB claim it and go through such trouble to Stealth it?

[kindred quarry]

It is not their model. This model is based on Qwen 2, which does not use the same tokenizer as Quasar Alpha.

[unkempt grove]

I agree, I'd love a paid version, maybe with prompt logging off so AI Lab's employees don't have to cringe at my awfully simple and ignorant questions

[ocean summit]

I imagine that in a few weeks some company will announce it's theirs and add it to their paid offerings.

[gaunt igloo]

Could or be passing the user's region without passing their IP?

[woven ibex]

anyone know if this is safe to use in production? it says it logs all prompts and responses both by the provider and OpenRouter. I've tried this and is by far one of the best models I've used recently

[quick peak]

Certainly, they can declare it.

btw currently mainstream llm companies rely on IP geo entirely

[quick peak]

"All prompts and completions will be logged so we and the lab can better understand how it’s being used and where it can improve."

Whether it is safe or not depends on your actually work.

[rose crest]

Because you aren't connecting to a US IP - you connect to a global anycast IP that routes to the nearest Cloudflare datacenter regardless of your DNS settings, and that datacenter makes the call to the LLM provider, and it's the datacenter's IP that is used for geolocation.

[quick peak]

Yes, but as far as I know, there is no place in the world where Gemini is available but GPT limited

[rose crest]

Vertex AI is not geolocked, only AI Studio is.

[quick peak]

i will check it

[quick peak]

AI Studio also work well in my end

without vpn

[unkempt grove]

I've noticed that Quasar usually "asks" itself the question again before answering. As in, I ask it something, it says "Does XYZ do ABC? The answer is yes."

I'm honestly very excited for this model's release, and I really hope its sanely priced. It will def become my daily-driver.

[quick peak]

Trace fyi

[late vale]

how does the maker of quasar get help if we cannot give feedback in cline for example? like thumbs up or down and commenting on a prompt

[worthy yacht]

they review the prompts and outputs and see how people use the prompts, if people are unhappy with the replies (like if people say to the LLM "this should have also had X")

idk if thats how they do it, those are just my guesses

[acoustic sigil]

for me quasar has a hyper fixation on spitting as an action verb not sure why

[inland parrot]

They're probably reading this thread 🙂

[late vale]

thats inefficient? oO

[inland parrot]

Mostly kidding, but I bet someone does dip in here to check the vibes

[atomic ridge]

AI cant check vibes yet? :/

[waxen meteor]

Thanks SILX AI for this huge model.

[mighty torrent]

I posted on X about why I think it’s an OpenAI model. If this is truly from SILX AI or some other lab then they are VERY good copycats. https://fixupx.com/jakobdylanc/status/1909228726170108374?s=46&t=6xEMBr9tWG2b24ES6c6q2Q

Jakob (@jakobdylanc)

Quasar Alpha is definitely (99%) an OpenAI model. Evidence? USER NAME SUPPORT.
︀︀
︀︀In the chat completions API spec, the message object typically only has 2 fields: "content" and "role".
︀︀
︀︀However there's a third lesser known "name" field that VERY FEW providers support. To my knowledge there are only 2 providers that support it: OpenAI and xAI. But they don't support it equally...
︀︀
︀︀While OpenAI only allows alphanumeric characters in the "name" field, xAI allows ANY characters.
︀︀
︀︀So I tested with Quasar Alpha and it does indeed support the "name" field, but ONLY with alphanumeric characters, otherwise it errors.
︀︀
︀︀Therefore...it's an OpenAI model. Either that or someone is a very good copycat.

Quoting OpenRouter (@OpenRouterAI)
︀
Quasar has topped the charts 👀

**💬 2 ❤️ 8 👁️ 1.5K **

FixupX · Apr 7, 2025, 12:56 PM

I saw that paper. SILX seems to be leaning into taking the credit. But until OpenRouter themselves confirm it I’m not fully convinced

[wicked rivet]

not a good look for SILX after everyone finds out it's really the new openai mini model

they may go down the glaive / reflection path

[mighty torrent]

Why would SILX confirm all this stuff prematurely while OpenRouter has still been completely silent? Wouldn’t they coordinate the timing on that? Idk seems suspicious

[wicked rivet]

about 80% of this thread

Quasar is a common word. why don't they use Quasar-Alpha in any of their docs?

why announce it's your model, but spent all that effort to hide it?

[vague vortex]

indeed

[vague vortex]

indeed yah you are smart

keep being like that

[waxen meteor]

agreed

[late vale]

o4 mini is NOt better than the experimental 2.5 gemini coder?

[grizzled orchid]

perhaps see if Lambda mimics this behavior or not? SILX claims Lambda as their partner on their huggingface pages

90% doubt it's SILX, but it does make for good drama 😂

the SILX guy's linkedin

these does seem like just some dumb grift

[vague vortex]

such a hot guy

[late vale]

why hot

[waxen meteor]

where's "quasar-alpha" that's a scam

[waxen meteor]

seems ye

[versed pine]

1 scam 2 techbro companies 1 real company all working at the same time. either dude runs on amphetamines or its A Grift

[vague vortex]

wtf

[chilly orchid]

just so we're all clear, @vague vortex is Eyad, the guy behind SILX-Labs.

[vague vortex]

me ??? really??

[wicked rivet]

yeah, no fucking way a lab with only 2 guys can serve 10 billion tokens per day, even with lambda.

[waxen meteor]

damn

[grizzled orchid]

huh

[waxen meteor]

yep yep yep

[vague vortex]

who said anything

[mighty torrent]

Support for the message "name" field depends on the underlying LLM being fine-tuned to support it. Lambda is just a provider, they don't make their own models afaik, so I don't think they'd somehow magically support it

[waxen meteor]

Silx-labs made quasar-alpha? Fr? 🤯🤯🤯

[vague vortex]

n o

[waxen meteor]

No way quasar is a common name

Ig not

[vague vortex]

no

[wicked rivet]

https://tenor.com/view/black-hole-national-space-day-black-holes101-void-hole-gif-22095264

quasar bros, c'mon

[vague vortex]

i love quasars

[grizzled orchid]

Toven just confirming that they know @vague vortex is actually Eyad, is pretty funny

I assumed that Troy was just impersonating Eyad for memes

now, I don't know what to think :)

[waxen meteor]

Nah nah it's open ai

[vague vortex]

i'm just famous

[waxen meteor]

Open ai model 100%

[mighty torrent]

idk who any of these people are but this is spicy

[vague vortex]

it's from open ai

[waxen meteor]

Nah it's open ai

100%

[wicked rivet]

no i think it's from open ai

[grizzled orchid]

Toven would have to have some reason to actually know who they are, and I was assuming that SILX was literally just a grift until just now, i.e Toven would have no idea who they are

[versed pine]

this 'paper' is so corny it looks like it was typeset in google docs

[grizzled orchid]

lol

[waxen meteor]

Agreed 100%

[vague vortex]

told u already i'm just famous

[wicked rivet]

on second thought, i think it's probably from openai

[versed pine]

(their github is linked in their discord profile)

[mighty torrent]

spill the beans bro, say it straight, give us the sauce bro

[waxen meteor]

My troy is famous

[vague vortex]

u may wanna have a look here :
https://arxiv.org/abs/2412.06822

[versed pine]

andw hen you click on it it goes to the dude's name page and linkedin

[grizzled orchid]

ah that's fair

[vague vortex]

i already did ;
https://arxiv.org/abs/2412.06822

[vague vortex]

oh really

[waxen meteor]

Damn

[mighty torrent]

is SILX behind quasar alpha in openrouter yes or no bro

[waxen meteor]

Naah

Ifs open ai

[wicked rivet]

Troy should sue Sama for stealing his word.

[waxen meteor]

He said yes it's okay to be wrong jakob

[vague vortex]

indeed

[waxen meteor]

كسم سام التمان

[vague vortex]

no

its not

hahahahahahaha

[vague vortex]

i can't

😭😭😭😭😭😭😭😭😭🤣🤣🤣

[grizzled orchid]

well, I'll keep my mind in a conflicted superposition of

• "this is just openai's next innovation"
• "this nobody internet memer is actually just gonna revolutionize a bunch of shit when quasar alpha gets released"

:)

spicy drama, at least

[vague vortex]

hope soo

[wicked rivet]

i live those conflicted superpositions

[mighty torrent]

so explain your quasar vs quasar alpha??? just a coincidence / misunderstanding???

[versed pine]

this is greek letter spam made to look smart more than the white house's tariff calculation was

[grizzled orchid]

he's not gonna confirm

[vague vortex]

yep

u got me

u are very smart

[mighty torrent]

im an expert journalist trying to get the scoop

[waxen meteor]

Lmao

[vague vortex]

who knows ?

[wicked rivet]

you forgot to carry the 1

where's Good Will Hunting

[vague vortex]

its from open ai i guess

[gaunt ridge]

woah how about we chill out in here lol it's just distilled from gpt-4something

[waxen meteor]

Agreed

[grizzled orchid]

the only thing that has given me any credence to this whole thing is precisely this statement

[waxen meteor]

100%

[grizzled orchid]

[vague vortex]

bro i'm famousss

[gaunt ridge]

I mean that's what Toven wants us to think

English only please

[wicked rivet]

Quasar may be OpenAI, but TroyGPT is actually powered by Grok3.

[vague vortex]

ahda

[vague vortex]

i'm from xai indeed

i work for elon

[wicked rivet]

ellen

[waxen meteor]

That's it

[versed pine]

hey btw you know that twitter post they posted? they link to this 7b which is purportedly distilled with the same technology
... but it's just a Qwen2ForCausalLM

[grizzled orchid]

I bet Toven and the rest of the openrouter team are having a blast reading this drama like the rest of us :)

[versed pine]

does ANYONE look at these things

[vague vortex]

btw

this is testing

[vague vortex]

this model

is simply

trained

on the data

using ttm not quasar

the base is deepseek-7b

[waxen meteor]

Send silx ai Twitter please

[vague vortex]

but it have broken weights

[versed pine]

It is built upon the innovations introduced in the Golden Formula in Reasoning paper, featuring a novel training pipeline known as TTM (Token Temperature Mechanism) — a new approach to optimize reasoning and contextual focus during training. We also apply what we believe is the best formula for Reinforcement Learning (RL) training to date.
then maybe be clearer about that because it very much doesn't look like that reading every other piece of detail

[vague vortex]

i already did

look

[vague vortex]

"We wanted to test how well it performs as a reasoning model so we chose DeepSeek-R1-Distil-7B as the base, and trained it on Quasar-400B data.
"

[wicked rivet]

Reflection is back?

[opal breach]

@vague vortex you should put out a statement of sorts to clarify the model situation

[vague vortex]

i said nothing lmaoo

[opal breach]

Otherwise people will literally go crazy

[opal breach]

Yeah exactly this the issue

[waxen meteor]

So who made quasar-alpha? Idk

[gaunt ridge]

It's actually apples first gpt confirmed, noone would come up with such a pretentious name like Quasar except apple

[opal breach]

Make some X post or whatever you’ll get a shitton of engagement anyways

[vague vortex]

i'm a guy who own a seris of models named quasar thats it

[wicked rivet]

Troy when will Qwen 3.0 be out?

[opal breach]

Yeah but people losing it over the name similarity

[waxen meteor]

It's not open ai 🤯🤯😔😔?

[grizzled orchid]

he's not gonna clarify, there's only two possibilities:

1. He's just a grifter meming on openai's innovation, in which case there's nothing to clarify - it's just a meme, and he'll be forgotten about shortly
2. It's actually SILX's model, in which case he said something about "this weekend" and I assume there's gonna be lots of fanfair then

[wicked rivet]

Just ban these shitposters pls @chilly orchid

[opal breach]

They’re acting as if SILX trademarked the quasar name lol

[waxen meteor]

From December 2024?

[opal breach]

I don’t think he’s a grifter, but I also don’t think SILX is the one behind this quasar model

[vague vortex]

We are in the middle of raising a funding round, which takes time but will give us more resources to do even more

[opal breach]

Two things can be true at the same time

[versed pine]

there's, btw, still evidence of older models published under that name like https://huggingface.co/mradermacher/Quasar-3.7-i1-GGUF that they seem to have erased in order to further this grift (3.7 > 3.0 that they're saying they're making rn?)

[grizzled orchid]

definitely a grifter if quasar-alpha isn't his model

[opal breach]

What do you mean by that

[opal breach]

He never said it was his model

He openly stated it wasn’t his multiple times

[vague vortex]

what i did wrong

[wicked rivet]

Get a life yall

[waxen meteor]

It is 💀

[grizzled orchid]

that's fine, but he (should) stop meming so hard if that was the case :)

[vague vortex]

bro never heard of 'testing'

[opal breach]

The confusion comes from what he says here and there if you just pop in here without following the convo

[vague vortex]

indeed

[vague vortex]

yasta

[versed pine]

ive never heard of suspiciously erasing all your old work as soon as something new pops up

[vague vortex]

anta 3ab yy

[opal breach]

Discord is fine for shitposting

[vague vortex]

testing

[waxen meteor]

3abyy?

[versed pine]

how is it testing when the website you link to in the new readme mentions quasar 2.0 and 1.5 as a public release

[gaunt ridge]

https://tenor.com/view/typing-gif-gif-25597112

[grizzled orchid]

assuming quasar-alpha is they're model, the stealth release would actually make sense in this context. You stealth release so you can show investors exceptional proof of a really good model in a controlled setting, without showing your whole hand necessarily / productionizing anything

[mighty torrent]

Given that SILX AI has been using the name "Quasar" since as far back as November 2024 (see: https://github.com/SILX-LABS/Quasar/commit/7bc4aa66d6eb497e1728d71e3bc4bbb8614e66a2), why would OpenRouter call it "Quasar Alpha" if it's supposed to be a secret anonymous name?

I think SILX AI is just riding the hype wave since they coincidentally used the same name for a different model

[waxen meteor]

it's okay to be wrong jakob

[grizzled orchid]

regardless of if it's their model or not, they're clearly not a very sophisticated bunch

they're memers

that's fine

[vague vortex]

my own website that no one vists

[grizzled orchid]

"it's funnier this way"

[vague vortex]

i could do anything lmao

[vague vortex]

what hype i didn't do anything

lmaooo

[versed pine]

you could do anythging with any other silx property that doesnt change that its the official thing you linked to as if it was officially your thing

[waxen meteor]

يسطا

[mighty torrent]

bruh

[versed pine]

blah blah blah stop engaging everyone

[waxen meteor]

رد انستا @vague vortex

[vague vortex]

it is 😭 😭

[vague vortex]

where is quasar alpha

its my own models

mfs

[vague vortex]

مش بيفتح

😭😭

i can't

its my own models

my own company

[mighty torrent]

"quasar this quasar that" hype train aura farming mf

you did use the name first tho so

[waxen meteor]

He just don't want to ruin his thread

crazy

[vague vortex]

quasar is a famous name do u know quasars is a reall thing right?

[waxen meteor]

ازاي

[mighty torrent]

yea I wanna see one up close IRL

[vague vortex]

then stfu

i guess

lol

[waxen meteor]

ما ترد انستا يسطا

ما بقولك عندي شغل

[vague vortex]

كلمني dm هنا

مش بيفتح

[grizzled orchid]

I'm lowering my credence from 90% openai to 80% openai. I sure hope it isn't @vague vortex who has a model this good, as that'll definitely be a shitshow

😂

[vague vortex]

keep going

[grizzled orchid]

are any of the real quasar models released

or just distillations onto base models

if I can use a 7b model locally and it seems particularly good for a 7b model

then hey, I'll keep going down :)

[vague vortex]

not yet a

[vague vortex]

uploading the models even more 14b models

[grizzled orchid]

this weekend you said?

[vague vortex]

hope soo i guess

[grizzled orchid]

the real ones, not the distills?

[vague vortex]

the real ones will take longer cuz of the funding round its ugly and its taking time , cuz i have option to keep training now but it will not be very good good but not as i want or wait for the compute we will have

[mighty torrent]

best of luck, respect the grind

[vague vortex]

thanks jakob

[grizzled orchid]

i respect it assuming quasar-alpha, is in fact your model 😂

if it isn't, then

[waxen meteor]

Nah he's a scammer

[grizzled orchid]

I guess I'm interested to see how much you can fleece investors for

[vague vortex]

indeed

[waxen meteor]

He just trying to scam open ai models

Yk

[vague vortex]

in fact we are working on Synthetic intelligence not agi or this dumb stuff

[mighty torrent]

how do you think your models stack up to competition in their size class?

[grizzled orchid]

if quasar alpha is a 400b model, then it stacks up exceedingly well

[waxen meteor]

they built the pyramids

[grizzled orchid]

it would be the best known model for its class

I think most estimates put gpt 4o / claude sonnet around there, similar size-ish to llama 405b

[vague vortex]

i mean i can't say for sure but its going to be good cuz ttm works with large context soo good that u can make larger and larger context wich means the model will be more even useful

[grizzled orchid]

though it gets murky with moe

[vague vortex]

this is the main point i'm going to solve

[grizzled orchid]

what hell is synthetic intelligence vs agi 😂

that's some investor-fleecing speak if I've ever heard it

[vague vortex]

agi trying to solve problem in human way

synthetic intelligence get the same answer but with its own kind of thinking

[mighty torrent]

any comments on how your model's 1M context length perfectly matches quasar alpha? just another coincidence?????

[vague vortex]

no need to follow anyone

[kindred quarry]

I did

[grizzled orchid]

https://huggingface.co/papers/2502.05171 ?

[vague vortex]

in fact i tried to make it 10m context wich works btw but will need more compute so 1m context is good for that

[vague vortex]

will not really synthetic intelligence needs cognitive data

not just any kind of data

[mighty torrent]

coincidence or nah

[waxen meteor]

Yo follow silxlab Twitter

https://x.com/SILXLAB

SILX (@SILXLAB) on X

Quasar this, Quasar that...

It’s finally time to fully announce the new Quasar Series — introducing our 400B parameter models (1m context!!!) and smaller ones, the first of their kind. Built using a new scaling law training pipeline, advanced techniques, and more. By @LambdaAPI

X

[vague vortex]

coincidence i guess?

[vague vortex]

no dont

[waxen meteor]

Scammers?

[vague vortex]

indeed

[waxen meteor]

U built the pyramids bro

[vague vortex]

btw we are not even ai lab i mean in terms of llms it just cool to build/use

we are raising funding to build synthetic intelligence thats it

so its open ai for sure

[waxen meteor]

don't ruin his thread bro

[vague vortex]

sry i guess

[waxen meteor]

ty

[mighty torrent]

check out my project https://github.com/jakobdylanc/llmcord

[waxen meteor]

https://x.com/SILXLAB

Yo follow silx's Twitter

[vague vortex]

dude stfu

[grizzled orchid]

the biggest thing that makes me think it's not SILX

[vague vortex]

i support u

[grizzled orchid]

is that if SILX had this model

[waxen meteor]

I forgot they're scammers

[grizzled orchid]

they could publicize it

and then get acquired for $20-$100m

[vague vortex]

dude i'm not sam altman

my plans for quasar is to create a sota model and open source it

thats it

but its a side project

the main project is si

aka

synthetic intelligence

[grizzled orchid]

google basically paid $2.5bil just to get the big ideas back from Noam Shazeer's head

😂

[mighty torrent]

thx fam

[vague vortex]

and he did nothing

[grizzled orchid]

seems to be working out for them so far

[waxen meteor]

sorry we ruined ur thread

[vague vortex]

i mean open ai will comeback i guess

i wish for the real "open" ai tho

[grizzled orchid]

they've said that they only started working on their real "reasoner" models in October

[waxen meteor]

Sam vs troy

[grizzled orchid]

and gemini 2.5 pro is the product of that

[mighty torrent]

wdym

[opal breach]

Leave
Get 2.5B to come back
Exercise your stock option the quickest possible way
Do nothing in the meantime
Profit

[grizzled orchid]

nah

[waxen meteor]

On fire bro

[grizzled orchid]

that deal is definitely contingent on a bunch of shit

[vague vortex]

i mean i would wish for such a life

[grizzled orchid]

google is not messing around when it comes to their contracts wrt gemini

[vague vortex]

tho

[opal breach]

Wouldn’t we all

[grizzled orchid]

they're doing absurd non competes in London for their researchers

who basically can't go anywhere else for a year if they leave

[waxen meteor]

@vague vortex انت بتتكبر ولا ايه

[vague vortex]

الرساليل مش بتحمل

[waxen meteor]

شحات

بشكل

محدش يتصوره

شحاتين

[static gate]

wen quasar beta

[waxen meteor]

https://x.com/SILXLAB

There official Twitter? Or not

[grizzled orchid]

it would be about $1.45k / day, assuming an identical cost structure to Deepseek

https://github.com/deepseek-ai/open-infra-index/blob/main/202502OpenSourceWeek/day_6_one_more_thing_deepseekV3R1_inference_system_overview.md (based on this)

Deepseek's numbers:

• 608B input tokens per day
• $87k per day in assumed rental cost of H800s

so, if Quasar does 10B / day, then 87k / 60 => $1.45k/day

assuming that SILX has the efficiency improvements they're claiming, then it isn't implausible they could fund that

[waxen meteor]

Cheap? And fast?

[grizzled orchid]

(they said ~$100k for training the Quasar 3 model, so taking them at their word, they've definitely got the cash to fund it for a few weeks)

it would 100% be a no brainer to do it to drum up investor hype

Quasar Alpha seeing such a reception online is probably like a +$100m investor valuation play

for what, $30k in inference costs over a few weeks?

ANY company which has a model that is better at general purpose coding than the current SOTA labs is definitely a minimum $100m valuation

likely a lot more, if they drum it up right

if you've got someone as good at investor-farming as Aravind Srinivas, probably $2B at least 😂

[quick minnow]

The only thing that's important here is that there's still a chance that this isn't one of the Sam Hypeman's models

openAI = the adobe of LLMs

Let me huff the copium

[fickle zephyr]

Gad dang..
400B data, will you make it open source too?

[vague vortex]

sure

[grizzled orchid]

anyways, if SILX isn't quasar-alpha, then they're just a worthless grifting company :)

[quick minnow]

but honestly why would this dude be here if it wasn't his model

trolling?

I guess, considering we are on discord

[grizzled orchid]

the only ethical/correct thing to do, as SILX, if they're not quasar alpha - is to pin a tweet saying "To be perfectly clear, we did not make Quasar Alpha and we don't know who did. It's an unfortunate name clash"

[waxen meteor]

@vague vortex اقوله كسمك ولا اقوله ايه طيب

[grizzled orchid]

they're doing the opposite

[quick minnow]

yeah lmao, it's just so unnecessary

[grizzled orchid]

lots of vague posting right on the edge of "we are quasar alpha"

[waxen meteor]

Who?

[quick minnow]

I imagine he's getting a lot of attention on twitter bc of this

[grizzled orchid]

ofc he is

[grizzled orchid]

Troy of course

I assume you're also involved?

[waxen meteor]

Cuz he made it?

[grizzled orchid]

yes

[waxen meteor]

So what?

[grizzled orchid]

?

[grizzled orchid]

if you guys genuinely didn't make quasar alpha then just do this

otherwise you're a worthless group

[vague vortex]

note stfu

[quick minnow]

anyways, can't wait for openAI to put a ridiculous price tag on this so that their new mediocre o3 model seems reasonable

[vague vortex]

i love it

i sent u a dm

[quick minnow]

Btw this model is mine, I made it, I'll charge $100 per token

enjoy while you still can

[vague vortex]

truly from open ai

[quick minnow]

yeah, I sell hype just like my boss Sam Altman, I haven't made a usable model in years

"but 4o is good now" yeah I bet lots of people are using it on openrouter, let's check

[waxen meteor]

Scammer

[quick minnow]

Me rn

@vague vortex I'll call you troll today but don't ban me from the model tomorrow pls

[steady tiger]

So we agree this thing is 4.5o, right?

[worthy yacht]

I dunno yett

The improvments are small

[steady tiger]

No way, it's better than 2.5 Pro for all my coding stuff

and 2.5 Pro is miles better than 3.7 Sonnet

Which is miles better than 4o

[worthy yacht]

It fails to make beautiful web pages with good diagrams

I haven't tried Gemini 2.5 pro yet for that, but does what I jsut said change your mind?

[steady tiger]

I mean no lol

[worthy yacht]

maybe I have to use it for code more, other than making web pages

[grizzled orchid]

I haven't seen that

in my usage

it seems to be like a programmer who's done way too much leetcode

gemini 2.5 pro is like a programmer who cares about error handling maybe a little too much, but that's a good thing

"personality-wise" kind of the opposite of quasar

[steady tiger]

Hm, most of my impression has to do with understanding what the code is doing, how to change it without introducing errors, refactoring big changes

2.5 Pro is okay, but I'm a fan of Quasar. Maybe I'll run into rougher edges later

3.7 Sonnet is still not there for me. It's the sort of model that goes "You're right, let me fix that...", and makes another mess.

Which is the frustrating thing to me about most models except for Quasar and 2.5 Pro

And Llama 4 is just horrible.

[worthy yacht]

Quasar is very pacient and clear when describing what it changes it will add before actually making it and sometimes asking for permission

So is gemini metal-to-the-petal?

[mighty torrent]

uhhh guys

i just confirmed quasar alpha is multimodal

it can see images

nobody is talking about this

[chilly orchid]

pedal to the metal btw

[worthy yacht]

mhm, that's been known

[grizzled orchid]

yes

[mighty torrent]

sheeeiiiit

[worthy yacht]

why isn't images shown as supported for free models even when the model suppoorts it?

[grizzled orchid]

openai memes probably

[mighty torrent]

this makes me feel even more that its an openai model lmao

it hasss to be

openai collabing with openrouter is definitely interesting

[steady tiger]

It's very good at this too

[mighty torrent]

And I think I just found a bug? When my first message has an image, the “name” field no longer works (the LLM no longer knows my name when I ask it).

This doesn’t happen with e.g. gpt-4o

wtf is this website???

[chilly orchid]

@compact ginkgo bro

can you at least make it actually link to openrouter?

[compact ginkgo]

Uhm that's not my blog haha

[chilly orchid]

did they yoink your article??

[compact ginkgo]

My blog on 16x has link to OpenRouter

[chilly orchid]

https://www.quasar-alpha.org/

[compact ginkgo]

Oh damn. How did that cloned article get higher rank than my original lol.

Funny they kept the screenshots for my app

[chilly orchid]

hah sorry for accusing you

[compact ginkgo]

I love OR, I would never do such thing haha. I have been actively encouraging ppl to use OR.

[chilly orchid]

🙏

sincerely apologize hah

[wicked rivet]

quasar-alpha.org terms of service lol. These Terms are governed by the laws of France.

[fickle zephyr]

Either Oai or some other US company

[regal tulip]

Honestly? Just use Claude Sonnet 3.7. It's arguably a better writer, and is a FRACTION of the price.

[fickle zephyr]

Seems to have strict guardrail, I believe there is other model that filter through the prompt.

Or is it just small model that why it don't have the weight for the uncensored path like other model.

[regal tulip]

Neat! Though, worth mentioning that Claude is known to be an INCREDIBLY harsh self-critic. So while the rest of the scores are likely decently reliable, Claude's own self-score should be viewed as suspect.

[dusky wraith]

What's the best method to give QA context; i.e. RAG, upload documents... or load a large amount of markdown into it?

[vague vortex]

https://x.com/orionspotted/status/1909807347394134045?s=61

Orion-inference (@OrionSpotted) on X

Some more clarifications: the “Quasar Alpha” model on OpenRouter isn’t currently the same model as the “Quasar” I referred to, which is coming but not now : )

X · Apr 9, 2025, 3:15 AM

[quick minnow]

damn

that's so sad

this model is Sam Hypeman's son then

Sad news indeed

[vague vortex]

wtf

how do u read

[quick minnow]

I use my eyes

[vague vortex]

in his older tweet he was talking about orion that have the code name of quasar alpha

i guess

or something like that he means about an open ai model

this is the leaker

[quick minnow]

Mate ngl you might as well be a random user

let me be pessimistic by myself

[vague vortex]

wtf

https://x.com/SILXLAB/status/1909817368500388314

SILX (@SILXLAB) on X

More impressive stuff about TTM in fact, DeepSeek uses token selection mechanism to cut costs! Just like us, in a different way we predicted this from Dec 2024.

DeepSeek's new paper talks about sparse attention and token selection. Just points out that our vision is true that

X · Apr 9, 2025, 3:55 AM

wow deepseek uses token selection as well

very nice

[quick minnow]

Bro thinks this thread is general

[vague vortex]

dude

token selection is being used in quasar

lmao

[quick minnow]

which quasar

there's like 300 models name quasar at this point

[vague vortex]

only one

[quick minnow]

No one knows what Quasar Alpha is, what it uses or who made it

[vague vortex]

yah and?

[quick minnow]

this is false then

[vague vortex]

oh really?

[quick minnow]

yes, unless you mean the other quasar 3 model, in which case why not create a thread for your model

and leave us to discuss this openAI model

[vague vortex]

then discuss about the model lmao what is ur problem ?

[quick minnow]

I am rage baiting so that if you know more about quasar alpha you tell us

[vague vortex]

i guess u are

lmao

[quick minnow]

didn't work :(

[glass plank]

By the way, does anyone know if OpenRouter confirmed that Quasar Alpha will remain permanently free, or is it just free during the beta/testing phase?

[quick minnow]

Nope, it won't be free forever

they said that it will be free for another week (this week)

[glass plank]

Oh, thank you!

[compact ginkgo]

Whatever happens, it's gonna be my daily driver replacing 3.5. Unless it's prohibitively expensive like o1 pro.

I'm confident that aider, cline and Cursor will also switch to this model as default as soon as it comes out.

[quick minnow]

I use this model to apply code on Roo Code, gemini 2.5 as task orchestrator

works well however sometimes it gets stuck on edits that fail

[gaunt ridge]

I'm looking forward to trying it once the logging is off, I can't test it with everything until then

[compact ginkgo]

I think with this model you don't need to have separate planner and editor anymore. Not sure if you can do that in roo code.

[quick minnow]

It's mostly done to manage context, you use one model as task orchestrator (to create subtasks with isolated context) and another to execute the tasks

Basically agent orchestration

[atomic ridge]

2.5 pro for orchestrator and quasar for coding? What model do you use for boomerang?

[vague vortex]

claude

[atomic ridge]

Claude is expensive ;_;

[quick minnow]

2.5 pro for boomerang

[atomic ridge]

Think you can share your agentic ai flow for roo code? You seem knowledgable 😄

[quick minnow]

Basically this:
https://docs.roocode.com/features/boomerang-tasks/

But you already know about that

[atomic ridge]

I do, great thanks!

[compact ginkgo]

I have a new alternative theory. I think the model could be a fine-tuned model specific for coding by one of the AI coding companies (Cursor, Devin).

[coarse leaf]

Disagreed. It's kinda amazing in creative writing as well

Can't wait to see the final pricing

[compact ginkgo]

ok true. forgot about that, but it is really good coding...

[wild obsidian]

... and very good at translations.

[woven ibex]

man, quasar alpha seems like magic. its so fast and reliable compared to deepseek. how come it's free to use?

[compact ginkgo]

But OpenAI just released a new version of gpt-4o a few days ago. So it won't be called gpt-4o, surely.

[thorny agate]

ive been having to much fun reading @vague vortex s messages the last week. peak trolling imo

[wicked rivet]

theyve been working hard on 4o for a while now (cont. pretrained version in dec), and we still have no official benchmarks for it despite massive improvements. i guess this is the formal launch of the 'new 4o' with 1m context and an api dated version

[mighty torrent]

awesome blog, you should add my evidence to it 🙏 https://x.com/jakobdylanc/status/1909228726170108374

Jakob (@jakobdylanc) on X

Quasar Alpha is definitely (99%) an OpenAI model. Evidence? USER NAME SUPPORT.

In the chat completions API spec, the message object typically only has 2 fields: "content" and "role".

However there's a third lesser known "name" field that VERY FEW providers support. To my

X · Apr 7, 2025, 12:56 PM

an openai employee also liked my tweet accusing them of being behind quasar alpha haha https://x.com/jakobdylanc/status/1910005009699209606

Jakob (@jakobdylanc) on X

and this, come on now https://t.co/Q56Hby0hRo

X · Apr 9, 2025, 4:20 PM

[wicked rivet]

didn't Aidan post about some other models like o4-mini?

[mighty torrent]

searched his tweets for “o4” and found just this one, interesting

“early 2025” for o4

[wicked rivet]

hmm ok, might have been different wording. saw someone say they were testing some new internal models

[quick minnow]

damn, horrible news today then

this nice model is from openAI

that's so sad

hopefully gemini 2.5 flash can replace it when openAI puts a stupid price on this model

[regal tulip]

Okay, so we generally have an idea of WHAT Quasar Alpha is, but HOW is it?

Like, what are people's vibes with using it?

[quick minnow]

Good at following detailed instructions for single tasks

bad when the context window gets filled

and it loves to loop around when it makes mistakes

this is from a coding perspective

[mighty torrent]

I dumped it an entire codebase of system controller firmware for a board I’m working on and asked it to fix an I2C related bug and it did pretty well, it was nice that I could just blindly dump all the code and it found the relevant stuff for me

[primal laurel]

I am really glad we have a SOTA level model for free at very high rate limit

otherwise it would've costed me a ton

[proper widget]

Bette than 2.5?

[wicked rivet]

it's like 🧈 buttah

[regal tulip]

When you say context window gets filled, how do you mean? Like, does it operate fine all the way up to the 1 million token context limit, or does it poot out at an earlier point?

Either way, that's still pretty impressive!

[regal tulip]

Now that IS impressive.

Anybody given it a shot with regards to creative writing?

[wicked rivet]

i can't use this anymore now that i know it's Grok3

[regal tulip]

WHAT?

@wicked rivet, gonna need some evidence there. Because that is a buck wild thing if true.

[wicked rivet]

Ok i will give you proof. let me remember where I saw it.

oh it was probably a rumor based on this; Tesla Optimus’ Quick Evolution
Tesla initially announced Optimus during its AI Day event in 2021. At the time, Tesla only had a mockup of the robot and a literal person in a suit to demonstrate what Optimus could look like. By 2022, Tesla had a working prototype of the robot. Optimus’ progress has been rapid since then, with several dozens of the humanoid robots interacting with attendees at the Cybercab’s unveiling last October.

[worthy yacht]

there's a lot of things pointing to oai as the maker, not Grok

[wicked rivet]

i agree. i've posted like 10 times in here myself that i think it's gpt-4.5-mini or o4-mini

[worthy yacht]

what makees you think its grok?

[worthy yacht]

i don't understand what you're trying to say here

[minor glacier]

Grok does seem like the kind of lab that would run a skunkworks experiment like this.

[wicked rivet]

i saw a post, probably reddit or bluesky, only 2 places i usually check

i saw grok3 apis released too though, and they only have 128k context so i don't believe the rumor

[worthy yacht]

ah

[buoyant gazelle]

maybe they forgot this page was immediately updated

[wicked rivet]

😱

grok-3-mini-beta it is

[buoyant gazelle]

had something like a 1200 rpd in the dropdown when i looked at it

[worthy yacht]

its grok confirmed?!?!!??!! no cap ong fr fr super scary don't watch at 3am

[wicked rivet]

i'm just a gary, and know nothing.

[worthy yacht]

your pfpf looks like house md

[versed pine]

iunno on the same prompts this new grok and quasar output very different things anyways

(grok being the worse one because of course it is)

[wicked rivet]

it really is me lol

[worthy yacht]

we were joking about it

I dont think Quasar is actually from Grok

[rose crest]

doesn't do image inputs anyways

[wild walrus]

please don't be an expensive model

🙏

[compact ginkgo]

added an update for your findings. it would be good if you can post the positive and negative examples from xAI and Quasar Alpha (the successful API calls vs errored API calls due to invalid value for name field).

[regal tulip]

It is truly buck wild that we have had two major API releases in five days, and NEITHER of them have been this model.

[cunning egret]

https://blog.kilocode.ai/p/quasar-alpha-what-we-know-thus-far

[rustic elk]

I'm doubling down on it being Qwen/Alibaba. The odds are horribly stacked against me, but I'll look like a genius if I'm right.

[turbid jungle]

so its openai

[rustic elk]

If it was that easy, it wouldn't be a mystery to all the experts =P

A lot of models will say they're from OpenAI, because that's where so much of the synthetic data originally came from. (GPT-4 outputs)

[unkempt grove]

it doesn't smell like an alibaba model, for a while i really hoped it was some google deepmind model

it would've meant they finally got rid of the google smell

[rustic elk]

In both code and conversation it smells Qwen to me

[unkempt grove]

in convos it smells like chatgpt for workspaces as of 3 days ago to me

[rustic elk]

Interesting, I don't get that vibe. And Google has at least chilled out since the old days

I like Flash Thinking's default personality better than a lot of others

[unkempt grove]

For me, Gemini tends to repeat my questions in its answers, and it does so with a personality as dry as the desert. Does great on tool calling(as in: it makes sense to call that tool at that time) and complex structured outputs(as in: lists of other, complex classes), but the worst offender is the Real-time feature of Gemini. It's very jarring to talk to it.

[unkempt grove]

I have no proof of it, but OpenAI is doing A/B testing on ChatGPT with Quasar(or something like it). Not the "choose a response" type deal, but giving users a different model

Under the same "4o" name.

[rustic elk]

Which Gemini? They all feel different to me.

Hmm, A/B would be interesting

[unkempt grove]

Thats interesting, I blind tested myself on Gemini versions(no system-promp, 2 questions and then I'd have to guess which Gemini I used) and I was unable to tell them apart. Maybe I'm just not a Gemini connoisseur.

[unkempt grove]

That's the only way I can explain my personal account having a different 4o than my workspace account. I don't use memory or anything of the sort either.

[rustic elk]

Hmm, well I don't subject myself to blind testing, but I find Flash Thinking generally playful, and 2.5 Pro generally dry so far

[signal surge]

4o already did recently have an update FWIW

[unkempt grove]

I don't usually either, but I was drunk when I came up with the idea and in the middle of an argument with a friend about Gemini

[unkempt grove]

Was it even announced anywhere? I look around platform.openai.com and I didn't see anything about a 2025 update to 4o

[signal surge]

@unkempt grove https://help.openai.com/en/articles/6825453-chatgpt-release-notes

It's under "chatgpt-4o-latest" in the playground

Not sure why they didn't just commit to a new version

[unkempt grove]

“Fuzzy” improvements: 

Early testers say that the model seems to better understand the implied intent behind their prompts, especially when it comes to creative and collaborative tasks.

Heh, I specifically complained about this to 4o

[wild walrus]

please leave it free for one more week if you have 10 credits or more @empty violet please for mi familia 😭🙏

[void mulch]

it's 4.5o mini kek

[left zenith]

damn

[rustic elk]

Lol, same, biggest initial 4o complaint from me was how it could blatantly miss what I wanted from my simple prompt

[oak panther]

yea it's obviously from OAI (#1357398117749756017 message & #1357398117749756017 message) and doesn't score amazingly on the harder benchmarks

[wild walrus]

one more week please for mi familia, or at least until mid next week haha

[oak panther]

there's a significantly better model ("gemini-2.5-pro-preview-03-25") you can get for free as well :)

[wild walrus]

not sure if significantly better is the word but quasar is really good imo, i do use it to structure plans tho and use quasar as the task executioner

can't stand the rate limits on gemini 2.5tho

[wild walrus]

using gemini 2.5 a lot?

[oak panther]

It was really obvious when they were A/B testing o1, because they didn't even hide the "Thinking.." UI initially 😭

[mighty torrent]

awesome thanks!!! i just tweeted a slight correction to my findings: https://fixupx.com/jakobdylanc/status/1910316857623478314

Jakob (@jakobdylanc)

Slight correction...
︀︀
︀︀The error you get from OpenAI API when you use an improper "name" is:
︀︀
︀︀"Invalid 'messages[0].name': string does not match pattern. Expected a string that matches the pattern '^[^\\s<|\\\\/>]+$'."
︀︀
︀︀The regex pattern '^[^\\s<|\\\\/>]+$' means the "name" cannot have any whitespaces or any of the following characters: < > | / \
︀︀
︀︀So I was wrong when I said OpenAI only allows alphanumeric characters, since characters like # and @ will still work.
︀︀
︀︀Quasar Alpha still matches this behavior perfectly though, so...it's an OpenAI model.

**💬 1 ❤️ 1 👁️ 5 **

FixupX · Apr 10, 2025, 1:00 PM

and i also just tweeted this gist for testing it yourself https://fixupx.com/jakobdylanc/status/1910318880632750235

Jakob (@jakobdylanc)

Here's some code to test for yourself. Try with different models and names to see what works and what doesn't. gist.github.com/jakobdylanc/c2bae1d2b623cd35ab5f6b57705af7aa

**❤️ 1 👁️ 2 **

FixupX · Apr 10, 2025, 1:08 PM

[wicked rivet]

nice find!!

[late vale]

can u check if its qwen alibaba

[split siren]

Which model is better to use for writing code, Optimus or Quasar?

[dusky wraith]

What's weird about QA is how it behaves RAG-like; in that it seems to be able to respond as if it has been given context [system prompt or data upload, vector DB, etc.] and yet I'm using it 'out of the box'; i.e. OR API sample, nothing else

[mighty torrent]

lol

[regal tulip]

Hmmmm... Maybe they've implemented something like those new architecture innovations from Google? Where you create a dedicated "Memory" block to store hard and fast information?

[dusky wraith]

That's the thing: I created nothing [no memory block or anything similar] and it responds as if I did

[surreal ermine]

yup

[spice sleet]

Can you give a specific example?

Not sure I understand

[quick minnow]

A moment of silence for this model

Now to bet if the price tag will be double or triple digits

[static gate]

i can't believe this shit is just gpt 4.1

it's ok atleast openai on par with gemini flash now to some extent :)

[quick minnow]

I remember hearing the same cope about gemini 2.5 pro

"Nah it won't be $10 are you crazy?"

[rugged oasis]

Sharing my real world performance tests on Quasar Alpha. It outperforms Grok, Claude, and Gemini on a real world reasoning task. https://medium.com/p/19396ccb18b5

[empty violet]

Important announcement: Quasar Alpha will be going down at midnight tonight: #announcements message

Make sure to check out Optimus! https://discord.com/channels/1091220969173028894/1359585862089834566

[thorny agate]

rip quasar alpha :(

[steady tiger]

So is this it? Or what the hell is going on?

https://x.com/SILXLAB/status/1909475116637208937

SILX (@SILXLAB) on X

Quasar this, Quasar that...

It’s finally time to fully announce the new Quasar Series — introducing our 400B parameter models (1m context!!!) and smaller ones, the first of their kind. Built using a new scaling law training pipeline, advanced techniques, and more. By @LambdaAPI

X · Apr 8, 2025, 5:15 AM

[hard cosmos]

¿?

[gaunt ridge]

Wake up for my Quasar Alpha
He flied he flied
One day we'd know the origin
He lied he lied
Wake up for my Quasar Alpha
The time is nearly up
The namespace begins to fade
Rest in promptrroni Quasar Alpha 2025-2025

[fickle umbra]

Its not this

[wicked rivet]

😱 Nooooooooooooooooooooo!!!

[austere grail]

Simply going to be heartbroken if this comes with an hefty oai price tag it's been one of the most exciting models I've tinkered with in a long time but I can't justify paying more than latte when the quality jump, while there, won't be worth paying more to me

[unkempt grove]

sama will pull through. as much as i shit on him this one has to be a cheap, good model.

[austere grail]

let us pray

[steady tiger]

thank god

[sleek veldt]

let me guess, they'll release a finetune of quasar alpha that is sprinkled with emojis later on and this one will the benched?

[steady tiger]

What's the vibe on Optimus? Is it better or worse than Quasar?

[sleek veldt]

better imo

[grizzled orchid]

same imo

[steady tiger]

I wonder if these might indeed be mini and nano of 4.1o

I can dream

Then 4.1o would be even better

[grizzled orchid]

or rather, I think it's the same model, but post trained a bit diff

[grizzled orchid]

basically this

[steady tiger]

I thought Optimus seemed a lot faster

[grizzled orchid]

they nerfed the speed of Quasar sometime last week

it used to be just as fast as optimus

I'm guessing it was an artificial nerf

[steady tiger]

okay yeah, I tried Optimus and it's solid for refactoring of code with complex instructions

[gaunt ridge]

id hit on the twink fr fr

[wicked rivet]

What's the long context score for this ?

[quick minnow]

good bye Quasar I will miss you forever, our love can't happen cause your father is Sam Altman

[dusky wraith]

Adios Quasar!

[wild walrus]

Quasar, we spent good times

o7

[floral cedar]

yes, like before gpt-4o was released there were two variants on lmsys under variations of im-a-good-gpt2-chatbot

[waxen meteor]

Sam altman is just a scammer

He tried to claim the model 💀

SILX AI actually made it

Scam altman

[quick minnow]

proof?

cause you are a leaf user just like me

[waxen meteor]

Here : https://arxiv.org/abs/2412.06822

[quick minnow]

Mate even I have a model called Quasar, everyone has a model called Quasar

I present Quasar-Lix

[waxen meteor]

https://huggingface.co/silx-ai

U can download it 😊

[quick minnow]

Anyways

Back to Sam Hypeman

[waxen meteor]

Anything else?

[quick minnow]

no, you already proved that you have no proof quasar-alpha is from nobody-ai

[waxen meteor]

What about that huggingface 🤓?

Yea it comes from nothing

it made by it self

[quick minnow]

I see some quasar-3, quasar-3-instruct, however I fail to see any quasar-alpha

[waxen meteor]

It's quasar series 🤓

[quick minnow]

ofc

[waxen meteor]

There's quasar-alpha
Quasar-mini
Quasar-max
Quasar-ultra
Quasar3.o

U gotta check

And start to search or google

[quick minnow]

the whole family huh

[waxen meteor]

Yep yep yep

[quick minnow]

Nah that's cool man, great job

I hope you get those 2 investors you need

by stealing credit from openAI

[waxen meteor]

Proof?

[quick minnow]

There's a whole lot on this channel if you scroll up

[gaunt igloo]

@vague vortex r u claiming credit with that reaction

[quick minnow]

someone even made a post on a blog about it

[waxen meteor]

Give me a proof idk what u taking Abt

[gaunt igloo]

Sussy

[waxen meteor]

Idc abt them I just want a proof

[quick minnow]

"I don't care about the proof I just want proof"

ok

[waxen meteor]

They confirmed on their website?

Any evidence?

🤓🤓🤓.

he read that 26 pages in 1 min 😂

funny tbh

[gaunt igloo]

Name only isn't anything to go by

[quick minnow]

I guess we are stuck in the same situation eh @waxen meteor? I'm lying about it being from openAI with no benefits, and you lying that is from nobody-ai with while trying to get funding

[gaunt igloo]

Has anyone tried running those silx quasar models locally?

Give em the uwu test

[waxen meteor]

So u saying it cames from nothing?

Yea

Ai made ai

[quick minnow]

get that funding Troy

before it's too late

good, that's good

gotta keep those llama 3.1 models coming

[wild walrus]

deleted message 😮

[quick minnow]

As long as I don't get coal santa

[waxen meteor]

🤓

https://openaisea.com/try-quasar-alpha-viral-secret-ai-model/

🤓🤓🤓

[empty violet]

We named the stealth model Quasar Alpha without intending any association with other things named Quasar

[quick minnow]

So Quasar is not the real name basically

[waxen meteor]

He means "alpha" it's just a nickname to make it stealth. Quasar series are SILX's models

[waxen meteor]

So who made quasar-alpha?

[sage obsidian]

"without intending any association with other things named Quasar" bro...

[empty violet]

Correct, Quasar is not the real name

[waxen meteor]

Who made it 🤯?

[thorny agate]

so troy was grifting?

[quick minnow]

Bro is literally fighting the staff that host the model lmao

[red jewel]

😂

[quick minnow]

basically

[waxen meteor]

I js wanna know who made it

[sage obsidian]

It's Stealth for a reason man, it'll come out in time

[thorny agate]

open ai is what most of the evidence leads to

[waxen meteor]

Proof

[thorny agate]

scroll up its alot

[waxen meteor]

We js wait

But open ai will not confirm that they made it

Believe me

[thorny agate]

hmm

[sage obsidian]

this is what most people were pointing to as "proof"

[waxen meteor]

Not even close to a "proof"

[sage obsidian]

well it's better than guessing silx ai and spamming nerd emojis in the thread lmfao

[quick minnow]

I even got added to a list apparently

[gaunt igloo]

See there's a difference between saying "The name "quasar" was picked by us arbitrarily" and "Quasar is not the real name"

Now we know it's not silx

[quick minnow]

a no-name lab serving 10+ billion tokens

[sage obsidian]

^^^^^^^^^^^^

Like how was silx ever a real guess

[quick minnow]

and now get this: for free

[gaunt igloo]

Hm

Yeah also that i guess

[quick minnow]

the grifter

[waxen meteor]

So

It's not silx model?

[gaunt igloo]

No, no, it is @vague vortex personal project separate from the lab. Inference been running on a single home-grown potater.

Like GLaDOS

[quick minnow]

Careful, he might add you to his list

[gaunt igloo]

List of people to praise?

:3

[quick minnow]

:3

[gaunt igloo]

Look it's different from the uwu

It cant without an emoticon with an uwu

Even an owo yields a different reaction

[celest pendant]

Hey, just wondering if you have any idea when the reveal or availability might be?

[gaunt igloo]

Stop all ai research focus on this, we need to know why this is happening

[wild walrus]

we need to know ASAP!

[gaunt igloo]

I can't find another input that makes it do this

[waxen meteor]

fr

[drowsy hearth]

Kaomojis are just Asian emoticons.

Respond like a 15 year old girl with unsupervised internet access and include kaomojis.

[wild walrus]

stop all research we need to know

[empty violet]

can't disclose yet but it'll be soon!

[gaunt igloo]

HOLY SHIT IT IS RIGHT

[celest pendant]

thank you

[gaunt igloo]

I will bring that to Optimus thread now

[waxen meteor]

Isn't the expire date tomorrow?

[quick minnow]

They will probably release it under the real name

Quasar is a placeholder name

[void umbra]

@waxen meteor on it being OpenAI there are some unique patterns:

• appears to be same tokenizer
• uses the same tool ID call format as OpenAI models (not google / mistral / qwen / etc)
• OpenAI support a "name" field in their API, so does this model (only them and xAI do, but this model exhibits the same character limitations that the OpenAI ones have)

[waxen meteor]

I see

but

when

[void umbra]

https://x.com/jakobdylanc/status/1909228726170108374?s=46

Jakob (@jakobdylanc) on X

Quasar Alpha is definitely (99%) an OpenAI model. Evidence? USER NAME SUPPORT.

In the chat completions API spec, the message object typically only has 2 fields: "content" and "role".

However there's a third lesser known "name" field that VERY FEW providers support. To my

X · Apr 7, 2025, 12:56 PM

wont tag jakob but he's in this chat lol

[waxen meteor]

Probably yea

But

We wait until @empty violet reveal

idk who

[void umbra]

yes

i hope one of them is open source

would be really good for the community

[gaunt igloo]

I have reasons to believe optimus was distilled from 4.5

#1359585862089834566 message

[spice sleet]

https://tenor.com/view/popcorn-guy-relaxing-popcorm-popcorn-gif-26404903

The fun never stops

[empty violet]

Quasar Alpha is now down. but check out https://openrouter.ai/openrouter/optimus-alpha if you haven't yet!

[late vale]

quasar was much worse gem 2.5 for swe.

/closed

[errant pulsar]

Is it a temporary downtime ou permanent? Will Quasar be back?

[compact ginkgo]

err i can't believe i need to go back to a less superior model for today. gonna be waiting for it to come online officially.

i did one more experiment before it went down today and it excelled at writing blog posts

[grizzled orchid]

It’s not that much money fwiw

10bil tokens served by deepseek is only ~$30k

But yeah it’s funny that SILX just bombed their whole reputation

Especially if they’re actually trying to get funding

[compact ginkgo]

It's not about the money, it's about setting up the infra to do it at scale quickly. Also, it's 60 billon tokens for peak day.

[opal breach]

rest in peace quasar

long live optimus

that silx trolling finally hit

[worthy yacht]

Permanent

[dawn python]

QQ - will we ever know which model it actually was in the future?

[waxen meteor]

reveal pleaseeeee

[wicked rivet]

I will always remember you, Quasar Alpha. ❤️ 😢

[empty violet]

Try Optimus!

[empty violet]

Yes we’ll reveal it soon, likely next week

[unkempt grove]

I ain't gonna lie, I thought quasar was a pokemon thing until I found out it's a space thing.

[quick minnow]

Hell, even Fireworks asked openrouter to remove them temporarily, they are struggling hard to scale up

[compact ginkgo]

I thought it's a Mexican dish

[regal tulip]

Right, what I’m saying is that the architecture might have a way of actively storing inference data in a way that lets accessing large context more reliable.

[vague vortex]

No we didn't we already have funding, and our goal is to achieve a SOTA pretraining model under the cost of $1M, so I don't think that's an issue

[grizzled orchid]

the issue is you didn't explicitly disavow the notion that it was your model

you just kept piling on instead

[vague vortex]

and who cares?

[vague vortex]

wahh really????

[vague vortex]

indeeddd so why u guys care?