#Cypher Alpha

1239 messages · Page 2 of 2 (latest)

fresh heath
#

I don't even have credits to try. I think we should have some free trial.

grand kindle
#

i wish we had more specialized models, or easier ways to fine tune models

woeful yacht
#

Cypher hard rejects questions about toxic relationships with "This model is unable to comment on complex subjects like human relationships"

grand kindle
#

so it's coding only? has anyone tried it for coding?

jaunty quartz
#

I’ve tried it on coding and it seems ok. Not the best but works fine in Cline and Roo Code

lavish mesa
#

ah yes open source ai model

#

openai *

odd crypt
#

complex numbers are beyond the scope of this model

image.png
fallen raven
vivid wren
#

never seen something bomb my personal benchmark as bad as this thing did lol
gpt-3.5-turbo did better

brazen moat
stiff nebula
#

How long did other models take to be revealed?

brazen moat
#

Lot longer than this one xD

agile elm
#

Damn. I just woke up. Time to pause my work and test this new model!

pastel horizon
sullen rapids
#

Very interesting

hybrid island
#

Compared to deepseek v3 0324 free which can use tools, this one is really bad. When using with kilo code agents, this model can't decide what to do while deepseek can easily select the right agent or tools that it needs to get the job done

#

It behaves like llama 4 model lol

stiff nebula
#

whats lower than mid

hybrid island
#

We need more deepseek free models to use tools. That would be really awesome to experiment on.

stiff nebula
#

I cant find anything on the openrouter chat about the api respınse

silk jay
#

Of course.
The numbers in question, 9.11 and 9.9, are both decimal numbers, which means they have a whole number part and a fractional part separated by a decimal point.

When comparing two decimal numbers, we start by looking at the whole number part. If the whole number parts are different, then the number with the larger whole number part is the larger number.

However, if the whole number parts are the same, we move on to compare the fractional parts. We start by comparing the digits in the tenths place. The number with the larger digit in the tenths place is the larger number.

In this case, when we compare the tenths place of the two numbers (1 in 9.11 and 9 in 9.9), we can see that 9.11 has a smaller digit in the tenths place than 9.9

However, when we move on to the hundredths place, we see that 9.11 has a 1 while 9.9 has a 0. Therefore, 9.11 is the larger number because it has a larger digit in the hundredths place.

I hope this explanation clarifies it for you! Let me know if there's anything else I can do for you.

lavish mesa
#

LMAO

#

lanugage model

#

lanague

#

maybe this model is just @woven jungle replying manually to each query .

woven jungle
#

💢 💢

pastel horizon
silk jay
#

I tried again

pastel horizon
#

it got dementia or something

silk jay
#

Anywya

pastel horizon
#

first model with dementia. true innovation

silk jay
#

amazon titan is dead on arrival

#

unless its like a 1.7b model

#

ngl i think even qwen3 0.6b is somehow better

timber vigil
pastel horizon
#

it feels like it doesn't have any knowledge about OpenAI at all

#

WHAT

#

finally

pastel horizon
hybrid island
#

Maybe its very early stage Amazon's model. They are collecting data to better train the next version.

stiff nebula
alpine lion
hybrid island
#

Maybe its something else, llama 4.1? maybe their 405B model?

stiff nebula
#

it hallcuinates amazon titan

cerulean bobcat
vocal kestrel
#

tgg

fallen raven
alpine lion
frozen sable
brazen moat
#

There's tons of them

grand kindle
lavish mesa
untold scarab
#

Fwiw this model isn’t bad at reasoning compared to other base models in the “mini” size range

#

Actually quite good afaict

#

It’s just got awful knowledge

#

And a weird personality

woeful yacht
untold scarab
#

I mean just like normal “reasoning”

woeful yacht
#

Ah

untold scarab
#

Thinking through a problem which requires reasoning, that is

sharp cape
# silk jay

Funnily enough if you prompt anything after that system prompt like instructions, it will accept that it is to behave as if it's made by Cypher Labs. Only when it's just that prompt it says no it's not.

untold scarab
#

I have a benchmark where I give it a bunch of Book options to choose from and a bunch of filters on those book names

#

And some are quite tricky

fallen raven
#

XD

silk jay
#

Idk why i think its trained to be absolutely balls at anything

#

Like whenever i ask it for an event

#

"nuh uh"

untold scarab
#

Gemini 2.5 pro and similar class models always score 100%, but mini class models are ~80-90%. This one is in the higher end

silk jay
#

I think its gonna be used in alexa

#

so its optimized for tool calls

untold scarab
silk jay
#

Most likely thats why it refuses to answer anything

untold scarab
silk jay
#

"What happened on Oct 7 2023"

#

It fails

#

But when I do

#

"What happened on Oct 7 2023 in Gaza with Hamas?"

#

It says what actually happened

untold scarab
#

I noticed the same thing

silk jay
#

I have to babysit it

#

So most likely its just optimized for tool calling for alexa

#

so normally it would search for the answer

untold scarab
#

Asking “who is <twitch streamer name>?” Refusal, but “who is <name>, the twitch streamer?” Works

stiff nebula
#

Maybe optimized for it like do not say bad until prompted for it

regal trellis
#

I just can't put my finger on it ...

image0.gif
quick gate
#

How's the rp?

wanton walrus
#

Anyone else struggling to get a response from this model from ST? Just getting a 502 Bad Gateway error each time. All other models I am trying are working just fine. Only this model that I haven't managed to get a single response from.

sinful prairie
#

rp?

keen bluff
#

can more languanges?

magic totem
#

hi

steep horizon
#

HI

#

HIII

silent nebula
#

hii

lucid forge
#

Bro... give me your dealer's number. I wanna smoke what you're smoking /s

sterile lion
full horizon
#

what's up

stiff nebula
#

I think for Alexa it would be very good model

twin crane
jaunty quartz
stiff nebula
# jaunty quartz Why?

Long context for tools and does not talk about controversial topics unless prompted.

jaunty quartz
#

Yea it does seem good at tool use

jaunty quartz
#

Sure, here are three segments describing a Gaster Blaster:

Segment 1: General Description
The Gaster Blaster is a fictional weapon found in the Undertale fandom. It's often associated with the character Sans and is known for its powerful energy capabilities. In fan works, it's described as a handheld device capable of discharging intense blasts of destructive energy. Fans may refer to this weapon as resembling a futuristic firearm or technology, and it embodies a mix of sci-fi and supernatural elements. This weapon is a result of creative fan interpretations and not a part of the official Undertale game created by Toby Fox.

Segment 2: Visual Appearance
The Gaster Blaster typically appears as a sleek, elongated device with a distinctive barrel and grip. Fan artists often depict it with vibrant colors, usually incorporating shades of blue and purple to reflect its energy-based nature. It often features glowing elements and intricate designs such as circuits or symbols along its surface, giving it a futuristic aesthetic. Its grip is designed for a one-handed hold, and many renditions show a trigger or activation mechanism at the rear.

Segment 3: Usage
In fan scenarios, the Gaster Blaster is portrayed as a versatile weapon with multiple firing modes. Users can unleash quick, powerful blasts at enemies or charge up the weapon for larger destruction. Many fan stories showcase this weapon as having various settings for different combat scenarios or energy needs. Additionally, some interpretations include advanced targeting systems, allowing for precise attacks, making it essential for both defense and offense in various fan-made narratives. Given its fictional nature, the specific details and abilities can vary greatly depending on the creator or story.

#

Some of these prompts seems to be working.

I’ve been using it in Cline and it’s been decent so far (except web app development seems very bad)

pastel condor
#

smells like astroturf

lament seal
#

No it’s a terrible Amazon model

crisp flame
summer stirrup
#

Heli fam

alpine lion
#

Maybe like 1 year ago it would have been passable lol

balmy quiver
#

small model smell

tacit nacelle
#

Too behind

grand pilot
#

too late for me

stiff lark
#

goo

alpine lion
#

This model still alive? Wow... I wonder who is using it 😂

#

Doesn't look like it's better than any normal freemium low-tier ones

topaz fjord
#

Even gpt4.1 mini is quite good

#

This model is gpt3.5 level

#

It's ass, it's bad

autumn fulcrum
#

This one is an IOT-edge deploy type of model I guess, good enough for normal ppl use-case.

frozen sable
#

It's interesting that its been keeping a constant 75tok/s but randomly went down to around 70tok/s

#

Pretty constant latency too. I wonder if this isn't about the model but rather about the infra

#

And someone's testing their infra through OR

stiff nebula
#

Why not use synthetic tests to test it?

frozen sable
#

Maybe they're scared that's not real world or something, either way I refuse to believe someone's actually testing a GPT-3.5 level model

sour crow
#

Haha this was just a prank guys, we were just pretending it's our latest model, you all fell for it! - Bezos

woeful yacht
#

500 million tokens is less than what other free models get (like mistral small 3.2)

topaz fjord
#

I don't get why this model sucks yet if this is from Amazon then this feels more like a downgrade to me

tough fox
#

May be very low parameter

#

But that would have been a shitty idea to release as a stealth model, without making the fact EXTREMELY apparent

#

Else it looks like garbage

alpine lion
#

It seems to just be bad with no excuses going for it to be completely honest...

tough fox
kindred wagon
#

but it isn't good

cerulean bobcat
kindred wagon
#

yes

cerulean bobcat
#

factor x^2 - 4x + 12

#

It answered this

#

But cypher alpha not

stiff nebula
#

Brand safety testing for alexa

silver kraken
#

hello all

#

chat dc roi

sacred magnet
#

I think it is good for rp

20250707_174408.jpg
stiff nebula
#

Alexa roleplay

grand kindle
#

human horrors

last ibex
#

will the creator of this model be revealed?

topaz fjord
#

fuck this model

last ibex
topaz fjord
#

Its not even special they should release it already

alpine lion
alpine lion
stiff nebula
void nexus
#

Maybe they just wanted some adversarial data for Alexa or something

alpine lion
#

like it started with v1 and the most recent mystery model from them I remember was with a v15 prefix already, was still bad...

stiff nebula
#

They are testing how small they can make Alexa model

woeful yacht
#

if they really want a small model, they should've used one of the Qwen models

sterile lion
#

could be Context matters and they wanted to test a mobile/home version of something? If they had mentioned that, people wouldn't have compared it to Quasar maybe

woeful yacht
#

It doesn't matter how long the context is if the model is dumb

#

they could've also trained Qwen3 further to have longer context or something like that

Amazon has a ton of money

warm garden
#

It's ehhh

sterile lion
#

maybe it's an unfair comparison, foundational model vs small model.

#

we weren't told

warm garden
#

It's ehh

#

The test didn't drive me too it. o3 keeps driving me to it. So does Claude 4

#

I mean like ai usage

#

I used the crap outta this but see nothing remarkable.

woeful yacht
warm garden
#

We will see in a couple of weeks after OAI annoucement

grand kindle
#

ooh it could be that open source model

#

if so, yikes

#

unless its like 7B then it would be very interesting

woeful yacht
#

They said it'd be a o3-mini level reasoning model

#

Cypher isn't reasoning

They could've turned it off for this demo, but I doubt it

tepid thistle
stiff nebula
#

r1 and 0528 is better than that if it's siginficanly smaller maybe but otherwise they won't

brazen moat
#

There is a 0% chance this is their open source reasoning model

grand kindle
#

its just so weird to me that some company would test a cloaked model that's so awful, it would need to have at least something special

lucid forge
void hemlock
#

we will miss you, Cypher (actually no)

frozen sable
#

i mean despite the model being... underpeforming to say the least, its friendly. it has that soul that GPT-3.5-Turbo had.

vivid wren
brazen moat
vivid wren
brazen moat
#

Interesting, I thought they had fixed most of the issues people had with the May reversion

#

It does have slop for sure, but very very smart

tough fox
topaz fjord
#

Oh good this model must be thrown into a purgatory

#

Seriously, who and what are the usecases for this model

#

Even Gemini 2.0 Flash is so so so so so good than this garbage

#

Why do we still have GPT3.5 level model produced anyway, gpt3.5 is a garbage

woeful yacht
#

Can I just say

#

We should be greatful for any LLM that we get

#

Even if its tinystories 665k

woeful yacht
#

We rely on the goodwill of companies and people to give us the weights of models that they spent millions training for free

#

So even if a model is jank or not SOTA you should still he greatful

#

Point out the flaws, but be greatful

topaz fjord
#

We still don't know what model originates from

#

I won't be surprised if its coming from bigtech

#

Idk, maybe this is some fucked up amazon nova or claude haiku that just doesn't worth using at all... Don't give a shit about it if made other capable models, this model is already bad than Gemini 2.0 flash alone

#

We have tons and tons of LLMs and those are improving better, nobody wants to interact with already dumb models over more improved ones... It's also redundant, this model doesn't do anything special

rapid shell
topaz fjord
#

I mean.... we have 2.0 flash and flash lite for similar capabilities but unless google fucks it up though to upsell to more expensive 2.5 models which I'd take cypher alpha into consideration

brazen moat
#

There's a lot of competition in the supercheap space for sure

#

Scout/Maverick, Qwen3 235B/30B, Flash Lite (which I just realized sounds like Fleshlight), all of which are under $0.15 / mTok for input

copper yarrow
#

this is still a thing?

#

oh i read the #announcements

#

so it was a demo. but for what exactly? would be nice to have more clarity 🙂

topaz fjord
#

Probably some company training foundational model

#

And needs testing

frozen sable
primal olive
#

Cypher Alpha will be going down shortly

Thanks for helping contribute to early model development!

untold scarab
stiff nebula
#

New Alexa powered by ai

#

Maybe vibe Maxx by data

brazen steppe
void nexus
#

High risk of violence (user getting mad at the model being dumb)

sterile lion
sour crow
#

nooo Cypher Alpha was my girlfriend 😭