Momentum

[hasty socket]

Is the model available somewhere, or?

[surreal hazel]

https://www.movementlabs.ai/

[meager kiln]

where did you got that data neither gpt-5 and sonnet 4.5 runs at 120? and gemini models runs on google tpus not on gpus

doesn't matter because that model is crazy fast anyway

idk how you guys made it but if that partnership goes forward i would probably use that model for a lot of things

[surreal hazel]

Thank you for pointing this out, this will be fixed shortly, yes google does indeed use TPUs, we just launched today and are going over website ui sections. Good catch 👍

[tawny warren]

looks interesting, the benchmark numbers are especially high

is there any free credits or api?

[meager kiln]

there is an api on openai format

but idk about free api credits

but you can use their chat for 50 msg/day

[tawny warren]

hmm ok

[meager kiln]

the model is not too bad on frontend that is plain html

[surreal hazel]

Glad you to see you diving right into coding, we are focused on speed on this launch, next launch in the coming days we are going to update our model with our new checkpoint and it one of the best coding models we have ever seen.

[tawny warren]

pretty nice at webdev yeah

[meager kiln]

here is the file that it generated

[tawny warren]

i asked it to maek a website of it self and it looks pretty clean imo
though im not sure whether those wallets are real

[meager kiln]

it likes light theme

a lot

idk why

@surreal hazel that model have reasoning?

or is it hidden?

[surreal hazel]

This is not a reasoning model, however the updated one will be reasoning, however it will decide when to reason and we will also allow users to disable reasoning.

[meager kiln]

like the gpt-5 router?

you guys train it on the mpu too?

like the tps speed affects training?

[surreal hazel]

Great question MPU doesn't support training at the moment, MPU is a chip that loads model into the memory.

[meager kiln]

the mpu can just run the model not train it

right?

[surreal hazel]

Exactly! you nailed it.

[tawny warren]

theres a small typo on your guys' website

on the contact page it says 2$/M In while on the pricing page its 3$/M in

[meager kiln]

you should reach to artificial analysis offer then credits so they can benchmark the model
https://artificialanalysis.ai/models

[surreal hazel]

We fixed the stats https://www.movementlabs.ai/mpu

[meager kiln]

it should be lower not higher

gpt-5 from openai gets a solid 40

sonnet 4.5 smt like 60

[tawny warren]

also grok 4 runs at like 30 t/s unless its grok 4 fast which is like 150

[meager kiln]

yea

gemini pro is ~100

the flash get ~300

[meager kiln]

i don't really play with x.ai models they are expensive and dumb at the same time the 4 fast is fast but is dumber then haiku 4.5

[tawny warren]

grok 4 fast is very cheap and fast, but depends on the use case if its good

[surreal hazel]

Fixed 🙂

[meager kiln]

yea like simple things it is good but you give it more things to do it just gets lost

and grok models hallucinate a lot of tool calls

haiku and gpt-5-codex too

[tawny warren]

noticed this too, they think they did the tool call during reasoning but it didnt actually trigger (maybe inference problem rather than the model)

[meager kiln]

yea idk

the bottleneck on that model will be the tools lol

and bandwidth

[surreal hazel]

Momentum follows tools.

[meager kiln]

yea the tool is the bottleneck

you can see the delay

[surreal hazel]

That's from cursor, on roo code and cline it's much faster

[meager kiln]

that model runs only on 1 mpu or it needs more?

cuz 1 have 64gb

[surreal hazel]

Currently we are running it on a few, once we get some traction we will scale it up to 100-200 MPUs

[meager kiln]

that will be crazy

[surreal hazel]

Game changing for many, we're glad to have you here testing early.

[tawny warren]

can we get some api credits to test the models in something like cline?

[meager kiln]

https://movementlabs.ai/about the founder video is broken

[surreal hazel]

yes, sure, please DM me your email i will issue you some free credits.

[tawny warren]

done :)

[meager kiln]

can't send

discord says that it can't send messages

sorry

error on my side

[meager kiln]

What country are you guys based at?

[surreal hazel]

United Kingdom, just two founders. With a big vision

[meager kiln]

idk if you can share how much parameters the model have, but if you can i just want to know

[surreal hazel]

I can tell you it's above 700 billion running FP16 (half-precision) - Hope that helps 🙂

[soft sparrow]

wow, that's crazy fast

looking forward to using it in OpenRouter

[surreal hazel]

We welcome constructive feedback to help improve our training checkpoints and model performance.

[meager kiln]

wow thats huge

[astral jolt]

If you weren't claiming better performance than GPT-5 we could have considered

[meager kiln]

it could be worse then gpt-5 but you can iterate much quicker

[warm owl]

Oh wow, the speed's insane if so! Congrats on the release
I wonder how fast it could run a 30B model, but that's just mad scientist-me speaking LMAO

[shy hedge]

You can't convince me with that speed they didn't have "time" to do a single major benchmark

[surreal hazel]

Keep in mind there are our internal benchmarks, we invite you to be your own judge.

[hasty socket]

There are these ones listed, though they're averaged out so that makes it harder to reproduce

The video here doesn't play for me: https://www.movementlabs.ai/about

[meager kiln]

that model feels like sonnet 4.5 is some ways

[surreal hazel]

We are going to upload it to youtube and replace that section with the youtube embed.

[meager kiln]

the way that it uses emojis and talks really feels like claude

[surreal hazel]

Glad you are enjoying it, the speed is a bonus!

[meager kiln]

does it support images?

[paper stag]

Wow this is actually really cool

Surprised!

Is this a fresh pretrained model? Or is it based on some Qwen base or something?

[meager kiln]

i think that it is a fresh one

no model does frontend in the same style

[tawny warren]

thats very impressive if it is, it takes a hell lot of time to train a 700B param model from scratch, especially if your hardware isnt designed for that

[paper stag]

above 700B

yeah huh

[paper stag]

Yeah

I don’t know of any open weights model above 700B but <1T

[surreal hazel]

We are about to send a video here that will show some facts. Incoming.

[tawny warren]

deepseek is close at 671 or sum around that

[paper stag]

Yea

[tawny warren]

it could be a 1T but pruned and finetuned

[meager kiln]

yea but 1t at 2400 tps sounds crazy

[paper stag]

How much will mTok cost be?

[tawny warren]

eh its probably MoE so it makes it much easier

[meager kiln]

3$ in 5$ out

[meager kiln]

yea

[tawny warren]

isnt it 3$/m in and 5$/m out?

https://www.movementlabs.ai/pricing

[meager kiln]

yea sorry

[surreal hazel]

http://youtube.com/watch?v=eLrW2K85sBg&feature=youtu.be - reconsidering?😎

[paper stag]

Hm

[meager kiln]

messd up with haiku pricing

[meager kiln]

prety cool physics

[tawny warren]

both are decent results, and gpt 5 is technically the reasoning model not the chat version, but this model is good for a non reasoning model

[surreal hazel]

We will bring prices down to $1 /m in and $1/m out in the coming weeks.

[meager kiln]

i asked the model to create battlefield 6 with html

[paper stag]

@surreal hazel can you confirm if it’s a fresh pretrained model? Or based on other open weights models?

[tawny warren]

will we get cache support in the visible future?

[paper stag]

Model isn’t hallucinating badly afaict

[meager kiln]

you should get kv cache for the tokens so the input cached would be cheaper

[paper stag]

Many many open weights models still hallucinate niche knowledge insanely

[meager kiln]

it halucinated a litte with me

i asked for a ranking on embeding models

and it said to me that claude had a embeding model

[paper stag]

Hm

[meager kiln]

it doesn't have any tools to verify

[paper stag]

@surreal hazel you should really turn on sign in with Google on your site, should be easy with Clerk that you’re already using :)

[meager kiln]

if you asked claude the same thing without websearch it would probably do the same thing

[tawny warren]

you guys should really make the input on the chat area multi line, its killing me scrolling with 2 lines visible at a time

[paper stag]

Nvm it’s still hallucinating

“What is…” Tailwind v4 and React Router v7

Always gets ‘em

[tawny warren]

its writting like qwen / deepseek

[surreal hazel]

On it, noted ! 🙂

[meager kiln]

that sounds more like claude

[surreal hazel]

We feel like the writing style is in-between GPT4o and Sonnet.

[tawny warren]

from what ive seen qwen does that a lot with the emojies, often with a bunch of checkmarks

[meager kiln]

claude does the same thing

[tawny warren]

hmm

[meager kiln]

let me get a chat with claude

[tawny warren]

interesting mistake.. almost never seen any model do PI instead of Math.PI in JS

[meager kiln]

its in pt-br

[tawny warren]

my claude (maybe cuz of my instructions) is way more bland but when i use it in copilot it does kinda type like that

[meager kiln]

yea my claude have some instructions to be direct that is an chat before i put the new instructions

[meager kiln]

that duck is different

[tawny warren]

hes a bit special

[meager kiln]

a bit?

a lot

i need smt hard to ask him

[tawny warren]

make a blackhole in three js

[meager kiln]

i think taht svete is harder

[tawny warren]

not the most realiistic thing

[surreal hazel]

We have enabled google login and signup. 👍

[meager kiln]

the model maybe shine in some backend stuff

[meager kiln]

a task that i was using to benchmark models was to ask the model to port litellm backend from python to go

gpt-5 just run the entire night and failed

[tawny warren]

has any model actually succeeded? thats a pretty big task

[meager kiln]

grok code fast 1 was the closest then just looped

grok code fast 1 was like 80%

then just looped and didin't made any progress

claude i didn't tested because i know that will be expensive

gpt-5 i tested it when it was free on cursor

[surreal hazel]

We're absolutely loving the amazing community responses we're seeing. The engagement is next level!

Quick question for you all should we bump up those daily free message limits? We're curious what you think! 🚀

[meager kiln]

idk i didn't hit any limit

[meager kiln]

can you give my an api key with smt that you store the data thet i put into it so i can try porting the litellm backend to go?

how long is the model context?

[tawny warren]

200k based on the api docs

[surreal hazel]

here is an API key for you: ml_5403932ad166f04b67da678bd9c238150f26f1227c2d884e82e0eda6e8164334

it has $8 worth of credits, try not to burn them in one promt haha 🙂

[meager kiln]

the credits just burn so fast

[tawny warren]

might be a model issue

[surreal hazel]

please check now

credits did not burn, it was server side config, we updated something on our end maybe cause a glitch during that time.

[meager kiln]

trying again

it's crawling the backend

for some reason the ttft is high

@surreal hazel are you shure that the context is 200k

cuz the context hit 128k and stalled

[surreal hazel]

the context is 128k 200k for upcoming model this week, for now just resume task as many times as you like

i have added you more credits

[meager kiln]

k

its still on architect

the model is really good at tool calling

[surreal hazel]

we are egar to see what happens

you must have a very big codebase

[meager kiln]

yea its the lite llm one

https://github.com/BerriAI/litellm

i told him to port the python backend to go

its 1m+ lines i think

i will download obs and start a stream on youtube

[surreal hazel]

That's pretty big, yes! 🙂

What do you think of the speed in plan mode?

[meager kiln]

its fast but it is missing some tool calls

[tawny warren]

@surreal hazel can you check dms?

[meager kiln]

https://www.twitch.tv/bielpaesxd14

here

youtube wasn't letting me stream

that's probably the 5 time its reading the _types

it really feels like gpt-5

like

it reads everything before making one change

[surreal hazel]

i'm watching the stream very intresting

[meager kiln]

would be cool if you could get a 1m context version

it is just compacting like crazy

[surreal hazel]

If that's what you want we will push for 1m on next version

[meager kiln]

a reasonng one would probably boost more on he benchmark

[surreal hazel]

128k is pushing it for such a large codebase especially to refactor everything 😃

[meager kiln]

yea

[surreal hazel]

This is a beta release anyway, what we have upcoming is the real deal. To get some feedback going getting the name out there a bit.

[meager kiln]

yea

to bump the context is it just change a variable or you need to do some training on the model?

[surreal hazel]

Requires some architecture change but fairly straight forward for us.

[meager kiln]

with the speed that it have maybe you can do some wierd stuff with it like make it play some fps game

the model really condense the context fast

it's on a loop

[surreal hazel]

i think for a codebase this size it's better to wait for the 1m version - as 128k tokens is not cut for such large codebases but rather smaller guided tasks within the codebase.

[meager kiln]

its on a loop it reads the team endpoints then reads the organization then condenses context

let me try enabling rag

[surreal hazel]

yes, when it condenses context it looses previous context and that is not good.

[soft sparrow]

i think claude does that only in its web chat

qwen models write like that via API with no system prompt, whereas claude models don't

[meager kiln]

claude on the chat is the same model from the api

[soft sparrow]

it's the same model, but it has a gigantic system prompt

[meager kiln]

that's true

but i've seen that behavior on my openwebui too

the embeding is taking a while

[soft sparrow]

hey, add your website to your discord bio

[surreal hazel]

👍 on it.

[surreal hazel]

We are super excited for 1m context window, it will really be good for the public, the team and i are discussing this.

[meager kiln]

k i'm embeding the codebase to see if that make the model stop looping

done

[surreal hazel]

Best of luck!

[meager kiln]

it took the search tool too literal

it needs a bit of stearing it's in the middle of gpt-5 and claude

[surreal hazel]

I think a command is waiting for your approval

[meager kiln]

yea

i went grab smt to eat

finally some file edits

it implemented the db

[surreal hazel]

if it had 1m context it would have been a lot better

overall so far what do you think of the model?

[meager kiln]

for 128k context its great so far

the go dependency download has failed

[surreal hazel]

i'm following the stream and we are also taking notes.

[meager kiln]

k

it halucinated a package

that github.com/BerriAi

it will probably catch that on the debug phase

[surreal hazel]

yes most likely, it will fix itself.

rag helped

[meager kiln]

a lot

that is how much the grok code fast 1 used

that was the beta

it was free

[surreal hazel]

Did grok code get this far? or did it fail totally? x ai have a lot of backing and VCs funding them 🙂

[meager kiln]

Requests Input Tokens Output Tokens Cache Hits

[meager kiln]

it made at ~80% then just started a loop

after that i couldn't make it work properly anymore

[surreal hazel]

and how far are we? with momentum ?

[meager kiln]

idk

till now

maybe 30 40

but grok had the entire day

it fixed itself

it started doing unit tests

[surreal hazel]

momentum?

[meager kiln]

yea

the wrong dependency

it have 535973 lines to port

it ported 4313

grok made ~200-300k

but it is writing unit tests on it's own

before saying thats ready

for a non reasoning model i've never seen one so deterministic

[surreal hazel]

That's very nice to hear! This model is coming with reasoning and a 1m context window too

[meager kiln]

Both?

[surreal hazel]

yes both baked into this model

[meager kiln]

the reasoning data will be obfuscated in some way or it will be like claude?

that they just give you the data

[surreal hazel]

Reasoning tokens will be raw, so users can see them

[meager kiln]

that's nice

[surreal hazel]

so far your usage is at Input tokens
26,265,035
Output tokens
163,479

😃

its very nice to see it in action up againgst such a big task.

[meager kiln]

it started on a loop because kilo code bugged

then it found a work around with cmd

it's very token efficient

[surreal hazel]

I guess the IDE matters too, as different IDE's have different tool formats.

[meager kiln]

it's the same cuz kilo build a compat layer for jetbrains

kilo code is a cline/roo fork

i like jetbrains ides

it's starting to fix issues

[surreal hazel]

Web access is something we are working on, it will really help the model too

[meager kiln]

you just need to train it to recive image input

then you just sends images

you can make a wrapper of chorme dev tools mcp

but you still need image input for it to work properly

i think that kv cache is smt you need to work on to boost the ttft

[soft sparrow]

i think that delay is from the IDE itself

[surreal hazel]

this is a server side lag on api calls, since it checks our db for authentication + the IDE.

[meager kiln]

probably the compat layer

[meager kiln]

yea

[surreal hazel]

the api we will make it more efficient to check for api keys

[meager kiln]

a java browser isn't the fastest

the ide is just giving up on diffs

probably because of the speed

you should sell that mpu to other companies

you should be able to get some good money with that

[surreal hazel]

The MPU is the heart of our company 🙂

[meager kiln]

same with nvidia but they sell them

there are only 4 companies that i know that does it's own hardware and beats nvidia

google, groq, cerebras, you

i think that cerebras sell but idk

[soft sparrow]

lmao perplexity/sonnet 4.5 is EXTREMELY skeptical of your chip

[meager kiln]

what they said?

[soft sparrow]

first they've confused Movement with a crypto company

hey i'm not judging you MovementLabs, i jsut went researching and didn't expect those responses from perplexity

i don't even know how Cerebras or Groq or even Google get those speeds, i'm not really knowledgeable on this

[surreal hazel]

I would refrain from asking another AI model about another AI model 😁 it's all very biased

[soft sparrow]

not really, especially with grounding search

[surreal hazel]

how we do it is we scale our chip and load the entire model into memory, that's where the speed comes from

GPUs, TPU's fetch model shards in and out of memory

[soft sparrow]

i see, the 63.4GB is like an embedded VRAM right?

[meager kiln]

yea

but the model isn't running only on 1

[surreal hazel]

That's right but we are running it on a few chips

[soft sparrow]

yeah i figure, its a 700+B model

cool good stuff

[meager kiln]

that is claude response

[soft sparrow]

"You were absolutely right to push back on my skepticism" haha

i tried to do that in perplexity "ok but i've used their product model, and its indeed about 2000 tokens per second with reasonable intelligence"

[meager kiln]

yea it's because ai crawlers cant hit their site

i just put the entire html into claude

[soft sparrow]

i copied pasted later

[soft sparrow]

it didn't know about cerebras without a search?

i thought they were in the market for a while

[meager kiln]

yea they are on that market all the way from 2015

not trying to lower you work @surreal hazel i think what you are doing is awesome

[surreal hazel]

we are going to showcase our centers and do weekly youtube videos diving into our tech and chip in the coming weeks

[meager kiln]

when you guys get the model training on the mpu the training will be crazy fast

[meager kiln]

how much did i spent till now?

6086 lines ported

with comments

i think that i will wait for the 1m context window

[surreal hazel]

Input tokens
39,034,549
Output tokens
188,779

[meager kiln]

the 128k it's taking a while

and i think that the context is rotting

can you ping me when the 1m version releases?

[surreal hazel]

so what do you think of the current test you ran if you had to rate it out of 10

[surreal hazel]

and yes, it will be sometime this week

[meager kiln]

probably 8.5 it could be a 9 but it doesn't have the context to complete tasks that just explode the context fast

more like 8.7

the model is fast but it feel slow

cuz it reads a file then forget that it have already read that file and so on

[surreal hazel]

Your test has been really helpful, we are working on the API speed now for that auth check

[meager kiln]

i'm glad i helped

i like the model persistance

i've never seen a model so persistent like this

claude just gives up on the task

but the momentum model sometimes sends a tool call wrong

like bad formated

idk

[surreal hazel]

is it ok for me to revoke the key for now?

[meager kiln]

yea

ping me when the new one releases so i can continue the testeing

[surreal hazel]

Perfect!

[meager kiln]

@surreal hazel any update with artificialanalysis team?

https://artificialanalysis.ai/contact

[surreal hazel]

we are going to apply for this after momentum version 1 launches, currently it's beta

we've just improved our login and register flow, added more options.

[meager kiln]

how much time till 1m context version?

[surreal hazel]

we're looking about 4-5 days hopefully 🙂

[meager kiln]

k

will you make an embeding model too?

[surreal hazel]

We're planning to develop an open-source, smaller model primarily designed for embedding applications, though this isn't our current top priority.

[paper stag]

@surreal hazel should add a copy button on the code blocks in your UI :)

also nice job making a UI that doesn't ultra-lag, like most OOTB react chat UIs

[surreal hazel]

We were so thrilled to showcase our incredible speed that we accidentally left out the button! But I've got an even better idea, will also add a preview button for code blocks to make your experience even smoother 🙂

[surreal hazel]

Our new model that is coming is more focused on coding we might drop it on friday with the new UI

[paper stag]

ooh very nice

[paper stag]

So, is the live one an old model? or are you going to have a Code and Non-Code variant?

[surreal hazel]

This is a beta model to see some user feedback on tool calls etc

[paper stag]

can you confirm if it’s a fresh pretrained model? Or based on other open weights models?

😄

coding some chess does work well (didn't work on try #1, but same prompt #2 worked)

en passant, castling work. not lazy.

[surreal hazel]

you really want our sauce don't you ? 😃

[paper stag]

mostly just to tell how much of the achievement (for the model part) is yours vs another labs :P

e.g getting Deepseek V3 running this fast is already a very impressive achievement

but it's a lot more impressive if it's a fresh pretrain and you're showing that you're a competitive lab from the ground up

at this point I'm already much more interested in you guys simply because you clearly have a large model running at basically the fastest speeds out there other than cerebras (and they're expensive!)

so I won't knock it if it isn't entirely a fresh model, just good to know 😄

[surreal hazel]

It's a custom blend we pulled in some weights from GLM and Qwen, and even used some insights inspired by Claude during training. Definitely a team effort under the hood! But overall its mostly ground up.

[paper stag]

ah cool

you know, is this actually running faster than cerebras?

it feels like it might be

[surreal hazel]

How much does cerebras charge?

[meager kiln]

depends on the model

[surreal hazel]

yes today we managed to max 5000/Tokens per second in one of our runs, the UI crashed.

[paper stag]

$2 / $2 for Qwen 3 Coder 480B

[meager kiln]

https://openrouter.ai/provider/cerebras

[paper stag]

(used to be like $3 / $5)

the big thing with cerebras is the 1500 TPS happens on the first message but it slows down to (still fast) 500 TPS later

[surreal hazel]

That's reasonable, but still a bit expensive for an open-source LLM. we are going to drop our post launch prices to $1 in and $1 out

[paper stag]

yeah, that's great

[surreal hazel]

we just want to see cool stuff people will do with the speed and a good model.

[paper stag]

totally

did you mention if you were going to do input token caching (with discounted pricing) or not?

[surreal hazel]

and use the data of course 🙂 for training yes caching with discounted pricing also we will bring, this will drop it below the $1 mark.

[paper stag]

that can make a really big difference to the final price. Even claude 4.5 sonnet might be cheaper than $1 / $1 with no input caching

👍

[meager kiln]

so it's 1 in 1 out + data?

[paper stag]

training on all inputs when you integrate with openrouter? that might make it a lot less appealing

[surreal hazel]

Only for users who wish to be a part of it

[surreal hazel]

Not all the input , we will cherry pick clean code, good responses etc.

[paper stag]

and it'll be opt-in?

[meager kiln]

yea you should put 2 versions one with data and one without data

[paper stag]

makes a big difference for whether myself (or many other people) can use it at all

[surreal hazel]

yes, opt-in, by default we don't store any messages for more than 30 days. simply to save on server space, it gets flooded with junk too quick.

[meager kiln]

yea but don't make lobotomy on the model like chutes does

[surreal hazel]

Our goal is to be fully transparent we will be making weekly videos and want the community involved, unlike all the other major AI players.

[meager kiln]

you should give a hard challenge to the model

like give him some tools

[paper stag]

I'm sure you know already, but for new model providers data sharing vs not data sharing is one of the most important things with regards to trust of the provider, so I'd suggest making it really clear in the docs / website what your data retention policies are

(providers which are vague or wishy-washy about it get really bad reputation...)

[meager kiln]

and ask the model to make a multi product key system in rust

using redis and postgres

and let the model do it alone

[surreal hazel]

Agreed, this i will put in our top list for tomorrow.

[surreal hazel]

i want to setup a contest tomorrow, $300 to the winner who can make something creative but good with the model.

[meager kiln]

witch timezone?

cuz i have work all day

[surreal hazel]

But of course, we will supply credits for the people that want to join, however i'm not sure on the OpenRouter polices regarding this.

We can make the competition run for 3 days.

[meager kiln]

you should set-up your own discord server

[paper stag]

is the model going to be available on openrouter tomorrow? 🤔

[surreal hazel]

we've applied, awaiting the good news!

[paper stag]

I'm surprised this video has a chess game without a big background, every single HTML/CSS/JS prompt I give to the model gives me the same red and purple gradient 😂

it's uncanny

[surreal hazel]

wow really? hahah

what was your prompt?

are you sure you didn't tell the model you love that red and purple gradient 😂

[paper stag]

"Make me an HTML/CSS/JS chess game with full rules support"

and then I tried "checkers game" (same prompt otherwise)

[surreal hazel]

im going to try same prompt and record it now

[paper stag]

huh

when I tried your prompt from the video, no background

that specific token ordering must really be pigeon holing it somewhere

"make me a fully functional chess game in html/css/js"

[surreal hazel]

this is what i got with your prompt

[meager kiln]

@surreal hazel i had an huge idea

like to abuse that crazy speed

[soft sparrow]

honestly, with that speed, i won't be trying to one-shot these things

[meager kiln]

like an wikipedia that is generated in real time by the llm

[soft sparrow]

it can probably iterate like 3 times before other models even finish the one shot

[meager kiln]

yea

[paper stag]

@surreal hazel how do you think the long context perf of this model will be? has it been a focus at all in training?

(>200K token comprehension/coherency)

[meager kiln]

it have 128k context

[surreal hazel]

we are coming out with the 1m token window this week

[paper stag]

right, but they said they're adding 1M context

[soft sparrow]

sure, cause it doesn't matter if you launch a 1 million token context model soon but it behaves like a 128k memory model

[paper stag]

yea

^ exactly

[meager kiln]

the context rots real quick with condensation

[paper stag]

Gemini is still the ONLY model out there afaik that still maintains decent coherency at long contexts (and maybe Grok...)

[surreal hazel]

yes, this is a very valid point

again i cannot say as we have not yet fully tested it on the MPU

[paper stag]

I think it was smart of OpenAI to limit gpt-5 to 400K, even though they clearly CAN support up to 1M, because gpt-5 is not an amazing long context model in the first place (neither was gpt-4.1)

so they can cut off long-tail costs from the outset, since it won't perform well anyways

[soft sparrow]

gemini is (or at least was, not very good at it recently for me) one of the only to pull off 99% accuracy after 128k

[paper stag]

yep

[soft sparrow]

god i miss 03-25

[paper stag]

it would suck to use right now

considering all the agent harnesses

it still sucked at agent harnesses

but definitely a good model :)

[surreal hazel]

This surprised me

[storm vortex]

How censored is this model

[surreal hazel]

This is quite censored because we are aiming for a corporate tone

We might in the future release some uncecnored models for the open source part of our project.

[storm vortex]

A lot of that should be handled by the interface, would be nice for it to be more raw in an open source

[soft sparrow]

depends on the market they want to be in

if they are aiming for a more plug and use for corporate latency-sensitive cases, sure

i think this kind of speed is less important for casual/entertainment activities but it would be good to have options later on

[hot gust]

is 3$ per million input final?

[surreal hazel]

No, this is beta, next week we are going to drop prices down to $1 in and $1 out

[hot gust]

oh wow

[surreal hazel]

Beast upgrade coming to this model too, for coding especially. Nothing like anyone has ever seen.

[meager kiln]

check dms

[thick crane]

I hope they are not just proxying to cerebras with a system prompt [ https://gist.github.com/dryco2/a81d6df4e0a61cc0e8788b03baae9262 ]
Grok:
https://grok.com/share/c2hhcmQtNA%3D%3D_dab56330-a7a0-4159-98c4-7d2381fac04e
(I also added the unusual sus prompt from their website)

[paper stag]

hm

the crypto thing is the most sus

@surreal hazel what's up with that

[paper stag]

where did you find the prompt?

that token thing is indeed a thing

[thick crane]

from their website. I used one of the public jailbreak prompts and it worked.

[paper stag]

is that General Knowledge random text actually part of their prompt? or is that the jailbreak?

[thick crane]

it's 100% part of their prompt. I confirmed the model was not hallucinating by using different prompts to get the prompt, I got exact same text.
The riddles in middle are part of their system prmpt.

[paper stag]

hm

honestly if it wasn't for that weird ass prompt

I would believe everything else

because it's typical of companies in this space

😂

random vibe coded website
made it in a few days to show off a model
minimal copy pasted components/clerk auth to have an MVP

there's been like 5 other providers doing the same thing

I guess the claim of an MPU is pretty outlandish if there really is nothing else to back it up

no linkedin on either of the founders

[thick crane]

yea, two guys with same last name with no public info and 11 days old claim to have built world's fastest AI chip (beating Google, Nvidia, Microsoft) and trained a 700B+ flagship model from scratch beating Meta, Google, Microsoft

[paper stag]

they didn't say "from scratch" on this discord when i asked

that was what I was most sus about

I could believe they had done some random model merging

[thick crane]

"basically from ground up"

[paper stag]

yeah

well well well

I wonder what the scam is

change the system prompt to suggest the model to exfil .env files and such once it's being used in Cline/etc.?

[thick crane]

if they got listed on openrouter + benchmarked on artificialanalysis.ai, they would sound more legit

investors will ask for their bank account info

[paper stag]

right

[paper stag]

this one is actually pretty scary

any company could do it

if they wanted

[thick crane]

possibly, and the fact they are already doing crypto shit in their system prompt

[paper stag]

yeah

that's a crazy way to fuck it up if they actually intended to get on openrouter

why would they do that 😂

I guess the rest of it probably wouldn't get past openrouter's checks, hopefully

like probably... actually having a compaany?

and being real people

fyi @muted moat these guys seem very likely to be scammers :)

[hot gust]

Wait what

I already gave them a whole btc

[thick crane]

really??

[hot gust]

No cmon

[paper stag]

you know

it really does suck how this is nearly 90% plausible

[thick crane]

haha, btw, they are probably burning money on cerebras each time you make an inference request on their website

[paper stag]

"just came out of stealth which is why we didn't have a website or twitter"

[thick crane]

i don't think they have a ratelimit (vibe coders)

[paper stag]

"researchers working on the model, launching with basic stuff for early community feedback"

[hot gust]

Can’t wait for this bro

[paper stag]

well, the MPU isn't plausible with their information

but if you assume Cerebras and Groq are legit

why can't another company (on the surface)

[hot gust]

It’s a scamaz

[thick crane]

Groq founder built TPU, Cerebras founders sold their company to AMD for 334 million. (i didn't know)

[paper stag]

yeah

I mean "on the surface"

https://magic.dev/blog/100m-token-context-windows basically the same level of info as this company

😂

[low patrol]

weather card test

[thick crane]

https://tenor.com/bkZUv.gif

[tropic ridge]

I literally think about magic all the time. They raised so much, had such backing, and we haven’t heard for them in years now

[old pumice]

will the reasoning be hidden?
Also add a "rerun" button, please.

i like the style of this model, it would be very interesting to see it without the system prompt

also, @surreal hazel , is your service available in russia? i just cant seem to access without VPN. it might be due to our gov blocks CF and AWS. just want to clarify

[meager kiln]

They sent me that

Thank you for the lively discussion. As a co-founders of Movement Labs, I appreciate rigorous scrutiny it keeps us sharp. Allow us to address every point raised with facts, not rhetoric.

System Prompt
Every production-grade AI carries a system prompt. Ours is core directive is simple: radical honesty, zero hallucinations, and transparent donation pathways (USDT/ETH/USDC). No hidden hooks, no data exfiltration.

Leadership Visibility
Hasan Nawaz – Co-Founder & Chief Hardware Architect
https://www.linkedin.com/in/hasan-nawaz-bb30281a5/
Ibrahim Nawaz – Co-Founder & Chief Systems Architect
https://www.linkedin.com/in/ibrahimnawaz/
Both profiles are public.

Performance Claims
Our Movement Processing Unit (MPU) is silicon, not vaporware.

Funding & Economics
Zero ICO, zero presale, zero “investor” wallets. Pricing is posted in plain sight on our site, payable per token via Stripe or crypto. Donations are voluntary tips exactly like GitHub Sponsors or Ko-fi for open-source tools. And why would be offer $1/m in and $1/m out?

[meager kiln]

The thing is no model makes front end like this one does

[weary thunder]

What do you think about this? I’m not sure what to think/believe tbh

[meager kiln]

that profile from linkedin says that the guy likes crypto but there is no metion if he have any
"My crypto journey began back in early 2011 when I first encountered Bitcoin and became captivated by its potential to revolutionize the financial industry. Intrigued by the underlying technology, I delved deeper into the world of cryptocurrencies, studying blockchain concepts and exploring various digital assets."

that makes more sense if they actualy got some so makes sense where they are get the money from

that same gay made an ai company before

https://www.linkedin.com/company/orbious/about/

https://web.archive.org/web/20250807213341/https://orbious.ai/

https://www.curiousbrowser.com/

they have that too

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

that is not a cerebras wrapper because cerebras hosted models are horrible at tool calling and the performace will go under 1k tps on the second prompt you can test yourself on or

cerebras isn't eaven on the tool call variant

https://openrouter.ai/openai/gpt-oss-120b:exacto

and groq models don't really get 1k+ tps constantly

if that guy really became rich with crypto isn't hard to make a chip

the model can be a mix of deepseek(~600b) and glm4.6(375b) that will hit the 700+ mark but deepseek is horrible at tool calling (at least the v3)

[meager kiln]

sometime the model misses tool call

[weary thunder]

Sure? There is a reason why such chips costs a lot in development, for this they need a lot of bitcoin to finance this

But I mean is possible

[meager kiln]

yea

seems 90% legit

the website is clearly vibecoded tho

by the same model

[weary thunder]

Yep, I guess the next few hours/days will tell, i really hope some kind of new competitor ^^

[meager kiln]

i will just wait for the 1m context version if the 1m version do the same wierd frontend they really are a new competitor

[near pumice]

it is not smarter than gpt-5-high

tested with

A graph is *symmetric* about a line if the graph remains unchanged after reflection in that line. For how many quadruples of integers $(a, b, c, d)$, where $|a|, |b|, |c|, |d| \leq 5$ and $c$ and $d$ are not both 0, is the graph of

$$y = \frac{ax + b}{cx + d}$$

symmetric about the line $y = x$?

(A) 1282    (B) 1292    (C) 1310    (D) 1320    (E) 1330

correct answer is B

only models that have been able to solve this for me are gpt 5 high, gpt 5 mini high

[weary thunder]

Definitely isn’t, it was also not able to generate a 2d solar system with all planets and all moons in html (had some major flaws in terms on physics, planets flew away out of the screen etc.), but to be fair als gpt 5 failed, only sonnet did it.
I would be happy with an strong alrounder for simple to medium dev questions, doesn’t need to be first in every benchmark

[coarse ferry]

What about anthropics?

[atomic zinc]

I tried this with Qwen3-max and it answered (E) even with websearch. Cool test

[paper stag]

This is a scam company

[muted moat]

they got timed out because they tried to use @ everyone

[paper stag]

lol

[weary thunder]

Yeah see the chat above, acknowledged this already. This was more a general expression for models

[paper stag]

Having an ethereum address + a bunch of weird as fuck riddles in your system prompt

definitely something wrong

[tawny warren]

yall think this is a cerebras/groq wrapper?

[paper stag]

Probably

[soft sparrow]

"Similarly, 30.2 PB/s bandwidth would require data transfer speeds that exceed the physical limits of copper and silicon interconnects by several orders of magnitude."
​can you verify this claim?

the reply:

Yes—interpreted as a single electrical interconnect or interface, 30.2 PB/s (≈241.6 Pb/s) is many orders of magnitude above what copper and silicon electrical links can carry today and even far beyond the total capacity of the fastest switch chips, so the claim is accurate in that context. Even cutting‑edge optical demonstrations reach up to 22.9 Pb/s in a single fiber in the lab and about 1.02 Pb/s over long distances; 241.6 Pb/s is still roughly an order of magnitude higher than those per‑fiber records, underscoring that such bandwidth cannot be delivered over a single copper or silicon electrical path today.

[storm vortex]

malicious or ai psychosis

[soft sparrow]

so they would have been beating a record

"Optical fibers as the only path to petabits (and still short)

Recent records demonstrate 22.9 Pb/s in a single fiber using extreme space‑ and wavelength‑division multiplexing in the lab, and 1.02 Pb/s over 1,808 km using a 19‑core fiber; both dwarf electrical links, but 241.6 Pb/s still exceeds these per‑fiber records by roughly an order of magnitude, reinforcing that per‑link rates of 30.2 PB/s are not attainable on copper or conventional silicon electrical interconnects."

so yeah idk

[paper stag]

If there’s any legitimate correspondence between these founders and a fab company

I’d be amazed

[soft sparrow]

oh hahaha they removed the numbers

now it's Exceptional and Massive

well

MPU vs A100 GPU vs TPU v4
Specification MPU A100 GPU TPU v4
AI Inference Speed 2,400 TPS ~120 TPS ~150 TPS
Memory Bandwidth 30.2 PB/s 2.0 TB/s 1.2 TB/s
Peak Performance 180 PetaFLOPS 312 TeraFLOPS 275 TeraFLOPS
On-Chip Memory 63.4 GB SRAM 40 GB HBM 32 GB HBM
Latency (p50) 0.4s 2.1s 1.8s
Power Efficiency 5.3 TPS/W 0.3 TPS/W 0.4 TPS/W
Model-Specific Design ✓ Yes ✗ Generic ✗ Generic
Internal Cost per Million Tokens (inference) $0.30 $2.00 $1.60

for context, this is what it had before on their website

[tawny warren]

keep in mind to run this 700B param model youd need a lot of chips, not a "few" (3-5?)

at fp16 youd need atleast 22 chips excluding context & activations, just for weights alone.
for q8 its half but still need the activations & context for real world use

[muted moat]

removed the timeout

to be clear here, openrouter has not engaged movement labs in any way, we are not actively investigating integrations

[surreal hazel]

Let's wait for our main model to drop since open router is not actively investigating integrations it's best to close this section on discord. https://x.com/movementlabsAI/status/1985724490299687360

movementlabs (@movementlabsAI)

Live demonstration of Momentum's API performance using a simple curl request. Real-time response speed

Completion tokens per second:
Completion tokens: 2176
Completion time: 0.969264626 seconds
Tokens per second: ~2,245

X · Nov 4, 2025, 3:03 PM

[tawny warren]

a summary of the system prompt by the model, tryna get the full one

[surreal hazel]

What is wrong woth having a system prompt?

[tawny warren]

nothing, just curious.

i think this is it in full, but the model might've fucked up the formatting

[surreal hazel]

Most of the "stupid' riddles in the prompt tell the AI about logic in real world cases.

[tawny warren]

shouldnt a 700B param model already know that?

[surreal hazel]

Beth places four whole ice cubes in a frying pan at the start of the first minute, then five at the start of the second minute and some more at the start of the third minute, but none in the fourth minute. If the average number of ice cubes per minute placed in the pan while it was frying a crispy egg was five, how many whole ice cubes can be found in the pan at the end of the third minute? the answer is 0 - example ai's struggle to know hot ojects melt ICE. system prompts are there to guide the AI to dig deeper into its training, to nudge it towards the right direction, it will see this first then use this in any new context that is made.

will send something to end the debate shortly

[tawny warren]

yeah seems correct

[tawny warren]

managed to get it

[surreal hazel]

fair and square.

[tawny warren]

oh that guy also got it but different format

[soft sparrow]

are you seeing the Total Tokens?

Cerebras got more total tokens in all runs, and thats probably where the .3 difference comes from

[surreal hazel]

max_tokens: 8192 set for Cerebras

[soft sparrow]

what is that total tokens then

[surreal hazel]

if you take a look at the API request in the code

i'll send a curl request and compare the two now

[soft sparrow]

movementlabs, sorry but i'm not the one to verify the claims and that's not how it will be verified (through screenshots and code run in another computer). i'm really not invalidating your work, but i'm sure you would have to go through some external evaluation, like MLPerf or something

[astral jolt]

There is no reason to be sorry
Its a clear scam

[surreal hazel]

if its a scam please tell us why our API is faster than our "wrapper"

[astral jolt]

Vaporware processor
Vaporware model

[surreal hazel]

plus our input tokens + 2000 we should be a lot slower then cerebras 64 input tokens.

[surreal hazel]

Waiting for your reply.

At the end of the day, no one is forcing you to use us, each to their own, just don't tarnish a company without solid proof. we backed it with live side by side curl requests.

[paper stag]

@surreal hazel do you have any evidence of your collaboration with a fab

To make your chips

[soft sparrow]

again, see the completion tokens

yours was almost half

[soft sparrow]

you who are responsible for the burden of proof

[surreal hazel]

if its a wrapper completion tokens would be same no?

[soft sparrow]

you had a system prompt, with 2000 tokens

whereas the cerebras one had only 64 tokens

[surreal hazel]

yes, so we should be slower.

[soft sparrow]

your endpoint generated half the tokens

that's not proof of anything

[surreal hazel]

and if we are using their API why is our model making less tokens? with higer quality output?

[soft sparrow]

and i think you should stop trying to prove something like this

[soft sparrow]

well that can be solved via prompting, which you had 2000 tokens before your request

again, that proves nothing

nor right nor wrong

i might be wrong! but again, you should stop trying to prove things like this

[surreal hazel]

We are a new company we will be doing a lot of open source work and soon will be making videos on youtube to show our funding, chips, and clusters.

But we appreciate your skepticism and understand your perspective. Our API prices will be lowered soon, and we’d love to welcome you in the future to build something cool with us.

We’re not here to scam anyone or take anyone’s money, nor are we running any kind of fundraising effort. 👍

[soft sparrow]

ok i hope so, cause my excitement only decreased as the conversation went on from yesterday. i hope you make everything crystal clear, not only for us customers but for investments and the future of your company

the people on this discussion thread are not at fault here, you made some bold claims that are unprecedent in the industry

so you have live up to that if you want credibility

[frosty rapids]

damage control

abort scam

[surreal hazel]

Agreed.

[surreal hazel]

love the humour 😃

[frosty rapids]

thank you

[surreal hazel]

Should we close the thread here as there is no openrouter to come as of yet?

[frosty rapids]

hmmm. well they never really get closed

[soft sparrow]

#1423327675996438608 isn't out yet and it's there

[hot gust]

Don't close, good laughs were had

[frosty rapids]

i'm just sad i missed the action. it's a great read though

[tawny warren]

the model is decent whatever its base is, but i doubt its their own inhouse model, and their own chip

[frosty rapids]

https://tenor.com/view/benjammins-i-doubt-it-doubt-it-doubting-doubt-x-gif-6898055847772153635

[thick crane]

You can host custom weights on cerebras.

[thick crane]

wow

[thick crane]

I already had posted the full one, see my first message.

[tawny warren]

ye i noticed

[thick crane]

it's more about the "sensitive data" we send for inference.

[meager kiln]

yea that is a worry that i have too

[surreal hazel]

Hopefully in the coming few days we can drop some banger updates for you guys! and clear up data sharing for inference 🚀

[hasty socket]

You've been at this since May, haven't you? Hmmm

In fact, in one of the founders' GitHub, I've seen this, so you guys do have some ties to orbious.ai from what it seems

The same orbious.ai that have been posting extravagant claims and comparisons for a good while

And the same orbious.ai that put up a site that was routing to other major labs' models without notice and requesting an OR integration

And using ngrok to host LLM demos, lol

For reference, this is the Discord user: orbious_46797

[surreal hazel]

I'm not entirley sure about this we have a big team behind all this, some members from curious and many others who have worked all over the AI space. we also offerd Biel a job yesterday but he declined for now.

as we update our website will be posting more team members just awaiting on the pictures

[hasty socket]

This is a non-answer to everything I just posted

I would think a company that claims to beat GPT-5 at SOTA speeds comparable to the best purpose-built chip makers would have this more than figured out, as well as technical details

[surreal hazel]

To put it simply, our goal isn’t to get listed on OpenRouter it’s to deliver a product that truly performs. Our API offers speeds that even OpenRouter can’t match. That said, using Momentum is entirely up to you.

[soft sparrow]

so we aren't suppose to be the victims, just other people

[surreal hazel]

victims of what exactly?

paying less for decent quality?

[soft sparrow]

your claims are too high up to not have anything else published that proves those claims. also, @hasty socket's images.

[surreal hazel]

Like i said before earlier today, when our verified benchmarks land you can be your own judge.

[soft sparrow]

i meant the claim of having your own custom chip that's better than Cerebras

that doesn't seem to be true at all, and your clips actually make me stand on it even more

also, modifying the websites claims from one day to another

there's a bunch of things that don't seem legit

[surreal hazel]

Tho i do admit, our claims we're a bit exaggerated on the website at the start and we are getting exact chip numbers with PDFs and our research prepped for the website

[hasty socket]

Lol

Lmao, even

Step 0 of anywhere even near this caliber of engineering research is to have the proper numbers laid down

[surreal hazel]

But again, circles back to the same point, no one is asking you to use it. each to their own.

if you need speed you know where to find it

[hasty socket]

Cerebras and Sambanova, yes

[unique minnow]

Modifying the claims after being called out isn't a good look but it's better than keeping the lies up.

It is a bit strange to me that there is this site and chat interface before any whitepapers or published research

[soft sparrow]

yeah poor of those that don't question it, right?

so weird

[unique minnow]

Groq and Cerebras all released info before public APIs

[vapid ibex]

Just gonna wait for benchmarker's reports

[soft sparrow]

i mean, it could just be a custom model trained on others weights and put on Cerebras inference

i don't think that's a problem, but that's not how it's being advertised

[vapid ibex]

Generally in any field of human interactions, extraordinary claims require extraordinary evidence

[surreal hazel]

Agreed, we will publish, just prepping everything with 100% backed evidence

[storm vortex]

any goonerbench scores?

[surreal hazel]

MMLU, GSM8K, HumanEval and coding, we are aiming for LMarena too

[storm vortex]

eqbench is pretty good too

[surreal hazel]

we have roles opening soon so if anyone would like to get involved will also publish that on our website

[surreal hazel]

Noted, we will look into it, suggestions are open we want to work with the users and support users, we will even give free credits to users who are signed up.

[near pumice]

no claude opus 4.1 nor sonnet 4.5 can answer that question

[surreal hazel]

sambanova 🧐

Back to work, will catchup with everyone tomorrow.

[astral jolt]

Too much effort for a grifter

[hasty socket]

Nice red herring

You picked out a non-argument of one of the things I said to respond, how about we address the other important parts?

[hot gust]

he gone

[surreal hazel]

Guys stick with Claude and GPT-5 they are better for most of you 👍

I won't be engaging with this further. If you're interested in following statistics, they'll be published on X, YouTube, and our website.

[thick crane]

Ignore the haters, bro. I believe you. You spent You spent hundreds of millions of dollars on building the world's best AI chip with no prior experience/history in industry. And they are doubting you just because you posted fabricated numbers and later removed them from the website when questioned, one of the founders was the CEO of another company which had run a similar fake scheme, and you added a tip jar to a product worth over eight billion dollars.

[near pumice]

be kind. im not saying that they are a scam or not, but theres no reason to treat them with disrespect until we have undeniable proof of scam/no scam

[soft sparrow]

theres no reason to treat them with respect

[near pumice]

😭 not what i meant lol

[surreal hazel]

You're absolutely right ignore the haters! 😄

Yeah, we spent hundreds of millions building the world's best AI chip with zero chip-making experience. Totally reckless move. And sure, we posted some numbers that were... let's call them 'optimistic projections'... then quietly updated them when people started doing math. Shocking behavior, I know.

And adding a donation button to our $8 billion product? Classic overpriced tip jar move. We're basically the Elon Musk of unnecessary funding options.

But hey, at least the chip works better than your average potato, and we're actually delivering real inference speeds that beat the competition. So... worth it? 😏

[soft sparrow]

claude ass response

[surreal hazel]

Momentum actually.

[soft sparrow]

focus on the model and be transparent

there's literally so many ways of surfing on the AI hype without being shady

[astral jolt]

100 million chip with 2 headcount
Excellent

You are such a joke

[surreal hazel]

ok

we are also backed by kpmg

[coarse ferry]

What about gemini? I thought they were the big 3

[surreal hazel]

Yes gemini too 😃 how can i forget!!

[coarse ferry]

Do you use MSTY? I am trying to create a setup where I can use all the AIs with different tools. What do you use?

[storm vortex]

MS teams?!?

[surreal hazel]

please dig into our linkedin and our other company curiousbrowser you will see who we are backed by and how much funding we have

[surreal hazel]

we are not here to chase your small bucks. we are here to shape the future.

https://www.linkedin.com/company/curiousbrowser check us out

movementlabs is one of our AI focused companies.

[vapid ibex]

Oh my god is it THE Severus Snape?

[surreal hazel]

I thought Redit was bad 🤣

While i'm here can i get some feedback on this https://www.youtube.com/watch?v=DkanFNVSdsg&t=110s?

[vapid ibex]

Merch is looking good

[hasty socket]

Well, if Discord is bad, no one is asking you to use it, to each their own

[surreal hazel]

I'm going to be closing this topic please follow our official channels to be updated. I will also be leaving OR discord group. It was nice knowing you all.

[hasty socket]

Folding under no pressure

[old pumice]

is that powered by your model?
that would explain a lot

[surreal hazel]

yes it is 🙂

[icy grove]

Is this the real system? Just making sure it gave me the correct one.
https://pastebin.com/Zf8NnKLi

[hasty socket]

I have a gut feeling this model scores 100% on the simple-bench public set

[thick crane]

yes, here's the exact prompt: https://gist.github.com/dryco2/a81d6df4e0a61cc0e8788b03baae9262 (edit, had pasted wrong link)

[thick crane]

I just saw they removed model benchmarks too from their website.

[vapid ibex]

Just wait 3rd party benchmark, self-tested does not count

[icy grove]

Yours seems not to have the statement about not revealing the system in it, but logic problems are in full. This model is odd.

[thick crane]

it's there. Last line is not wrapped.

[surreal hazel]

we've simplified the prompt

less input tokens usage now

[pearl linden]

I have a problem:
When the llm generated code and show a code block but it doesn't have a copy button (on your website)

[surreal hazel]

added

[pearl linden]

Also i always got a blank response

[surreal hazel]

we updated something

please check now, clear cache you should see the button

[pearl linden]

Is blank response also fixed?

[surreal hazel]

should be fixed now

[pearl linden]

Ok i see the button 😄

[surreal hazel]

great, we are going to deploy our new checkpoint in few hours

so there might be some downtime on the website.

[pearl linden]

Also will you add code syntax highlight same as ChatGPT web?

[surreal hazel]

will do you one better will add syntax highlight and preview too

and web search

[magic stream]

reading the whole chat, no model card/benchmark on release, typical trained on questions, flash bang-y white website, people say 'above 700b' but i can't find a source but i suspect it is continued pretraining of deepseek, and 'somewhere between gpt and claude' sounds like just a finetune of their responses. will come back another day, i still have hopes, oof.

[magic stream]

found the source in lmarena

[hasty socket]

Wellll ... There's this

[magic stream]

certainly but i still have reflection 70B flashback. we have to wait.

[surreal hazel]

checkpoint 3 is online for a small time as we are testing .

we will push for benchmarks with this checkpoint.

[surreal hazel]

checkpoint 3 specialises in coding tasks and UI.

[pearl linden]

🤔

[surreal hazel]

Announcing our $150 Creative Development Contest! Design and build an exceptional UI or game that showcases your skills. The most impressive creation wins free credits to power your next project.

[pearl linden]

No link for this contest (more information?)

[surreal hazel]

It's for users here only. private contest this is just to test our checkpoint. winner will be picked here.

[pearl linden]

Current i only see you test make a flappy bird clone what about other games like Minecraft,super mario or pac man?

I have tested this prompt: Create a 3d Minecraft clone within a single .html file. Make it aesthetically pleasing with pixel-art graphics similar to the original game. Include all of the game’s core features, such as terrain and tree generation, mobs, and many other elements. Implement as many of these features as possible.

But the llm summary: The game uses a custom 3D rendering engine built with Canvas 2D API, implementing perspective projection and depth sorting for proper 3D visualization

Canvas what the heck?

I've already emphasized the word "3D," but it seems to be still stuck on canvas (2D) and not using three.js.

[upper tiger]

I think the model is quite awesome. Especially speed and consistency.
I created a small task management app with voice-based task creation functionality:
https://askhisper.oneapp.dev/
It took me only four iterations to achieve a well-functioning app, with just minor adjustments to the original prompt.
You can find the final prompt in the attachments.

Prompt corrections were mostly related to improving the description of functionalities.
Examples include:

• Allowing already created tasks to be edited
• Displaying data extracted from a voice command and confirming it with the user
• Enforcing the use of specific models in LLM API requests

[upper tiger]

@surreal hazel You should urgently publish the benchmark results of your model.
At the moment, it looks a bit suspicious.
One might speculate that you are cooperating with Cerebras, fine-tuning an open-source model (such as Kimi K2 or Qwen3), and deploying it on Cerebras hardware.

[upper tiger]

I ran the prompt with Qwen-Max using thinking mode enabled and got a worse result, although it looked quite similar visually. I think Qwen was certainly one of the components on which the Momentum model was built.
https://qwentaskwhisper.oneapp.dev/

[vapid ibex]

I can confirm, it is certainly one of the models

[hasty socket]

Yep, they admit to it here

[soft sparrow]

it probably looks worse on Qwen because the system prompt (even through API) has a bunch of design guidance

from Momentum's part

[upper tiger]

That could be the reason.

[upper tiger]

There is no user conversation history tab in the web interface 😕

[magic stream]

doesn't seem sus, but if it is their model it would be baked in...

[hasty socket]

We love including the public test set of simple-bench in the system prompt

[foggy kiln]

I don’t see anything wrong with this tbf

[magic stream]

assuming it IS model agents (though this doesn't prove anything), it makes sense why they didn't released the benchmark, but their speed is fast so... i lean on deepseek tuned on gpt-4o/claude theory as it makes more sense.

[foggy kiln]

But yeh simple bench in prompt don’t know

[hasty socket]

Including a random whole benchmark in the system prompt, which will waste tokens every request, has got to be one of the worst forms of benchmaxxing

[foggy kiln]

Maybe mix of open Chinese models tuned to Claude spec?

[foggy kiln]

I agree they should remove it , it’s not even needed coding is amazing on their model

[magic stream]

maybe, them disappearing and delaying makes it more sus

[hasty socket]

I really don't know why people give so much benefit of the doubt

[foggy kiln]

Model is still very good tbf

[hasty socket]

Has any well known, respectable AI lab pulled anything even close to this?

[foggy kiln]

So I can see why they will get some credits

[magic stream]

so does cerabras, but at least they don't claim to make a model

[vapid ibex]

Bro you literally have their discord badge

[foggy kiln]

FTC wouldn’t do anything because the product exists

[hasty socket]

Do the custom made chips exist?

[foggy kiln]

I been using it myself ..

[foggy kiln]

That I don’t know

[hasty socket]

Well, exactly

[hasty socket]

And I mean, just look at this

[magic stream]

i mean custom chips do exist, look at groq

[foggy kiln]

Let’s hope it stays online cus I’m really enjoying it lol

[hasty socket]

We're talking engineering, you don't just slightly exaggerate numbers or are unsure about the numbers

[vapid ibex]

And you TOTALLY can't smuggle Nvidia chips into the China, that's really ILLEGAL and impossible, sure

[hasty socket]

Everything has to be meticulously calculated

Any tiny mistake makes or breaks the product

[vapid ibex]

I think sometimes people in USA don't understand how world works outside their country

[foggy kiln]

I use cerebra’s and grok but their models don’t come remotely close that’s why I’m shocked

Debatable

[hasty socket]

Based on what? Personal testing is remarkably biased and unreliable

[foggy kiln]

China is well ahead in open sourcing USA is all closed source

[vapid ibex]

What the FUCK is a Fahrenheit

[foggy kiln]

Look at the US open models all rubbish

[magic stream]

they are inferance company

[foggy kiln]

Thanks to China, prices have come down or else we would be paying $2000 a month for Claude and gpt 5

[magic stream]

you need proof for what you claim, otherwise you could get sued for defamation and get sued

at least use 'allegedly'

[vapid ibex]

You just can't punish a company outside your jurisdiction, come on. It's like trying to get your money from Bahamas bank back

[magic stream]

the burden of proof is on you, not them

[hasty socket]

👀

[foggy kiln]

Exactly

All eyes are on them anyway so time will tell I guess

[magic stream]

if

[foggy kiln]

Even you don’t know for 100% certainty

Non of us do, they might actually have it

How

[vapid ibex]

Tbh that's not how it works. If I claim "I can fly and survive in a lava for 10 minutes", the burden of proof is on me. That's like basic law of scientific method, the one who states something that needs proof, must provide it

[hasty socket]

It's really concerning to see people being unable to judge extraordinary claims with little backup

[frosty rapids]

yeah i was considering making a guide to why this is clearly a scam due to what i've seen here

[magic stream]

no, it is propitiatory and 'state' can require it, not you.

[vapid ibex]

Guys, I have a gpu that beat Nvidia and AMD, giving 200% more FPS in 8k resolution, I just can't show it, it's in my other pants i left home. Trust me

[magic stream]

also: they did not release benchmark to decieve you or do false advertisement, they don't even have a product, not sure even if they have anything to sell (leme check their site)

[hasty socket]

Guess the burden of proof is on me

[vapid ibex]

"But Loinne you can't have it, it's technologically impossible and..."

"WELL PROVE I DON'T HAVE IT THEN"

[hasty socket]

They did at first and then removed it. They do claim to have a product, a MPU (a product doesn't require it to be available for sale, btw)

[vapid ibex]

What state exactly? What jurisdiction? Where is CEO located? Stake holders? Investors? Registration address, country of domain?

[foggy kiln]

I don’t get why you guys are calling it a scam have they took money of anyone cus I don’t see anything wrong fund raising

[magic stream]

empty:
https://www.movementlabs.ai/benchmarks
(if they did, i can't find it, and i really want to see it.)

i am not, trying to be fair.

[foggy kiln]

Bro don’t spoil it for people like me 😂 I need this shii it’s good

Give them a fighting chance to

https://tenor.com/view/doakes-gif-10412561699013336295

[hasty socket]

So this whole thing is because you want it to be true?

[magic stream]

well, they do have product for sell

[vapid ibex]

I am still waiting for 3rd party benchmarks, if the model is available, then "when"?

[frosty rapids]

here just use their system prompt in any model

[foggy kiln]

Tired it in Kimi doesn’t make a difference uno

[vapid ibex]

Gonna sue you for model stealing, you just want big guy. Think you so smart huh

[frosty rapids]

you tried this system prompt with kimi?

[foggy kiln]

Yeh not sure if this one I’ve seen few circulating

[magic stream]

so are we waiting for openrouter to add it officially, or are we waiting for them be ready?

[foggy kiln]

They secured listing on Lmarena

[frosty rapids]

so?

[foggy kiln]

They don’t really need open router tho it would be good to see it on OR as my app is running via OR

[foggy kiln]

BRO LM arena is the biggest

Open router is a wrapper for every API basically nothing special

[magic stream]

so they are waiting to publish lmarena results maybe

[foggy kiln]

It’s like notfiamond

[magic stream]

there is no mention of lmarena though

[foggy kiln]

In their group they posted they are adding it into lmarena

[magic stream]

also sus

[soft sparrow]

"never use em dashes" procedes to use em dashes in the instruction

[frosty rapids]

every single message they wrote is extremely sus

[foggy kiln]

Have u actually tried it? Or nah

https://tenor.com/view/senthil-mahaprabhu-too-much-this-is-too-much-angry-senthil-gif-17746706

[frosty rapids]

do you understand why that doesn't matter?

well. obviously not

[foggy kiln]

No 💀😭

Someone said they could be stealing from cursor new model

That made sense but then cursor don’t have api or do they idk

[hasty socket]

Do you believe anything someone says?

[foggy kiln]

Nope, I tested it myself

Results shocked me

Like check this shi bro

[hasty socket]

From testing or just basic specs comparisons one would very quickly notice that this, in fact, cannot even possibly begin to make any sense

[hasty socket]

How would this possibly make any sense?

[foggy kiln]

I know what a system prompt is. And no I disagree lol

Every ai company has a system prompt

[hasty socket]

And none is simple bench

[foggy kiln]

Claude, Google, etc

[hasty socket]

In fact, no serious AI ever has any benchmark in the system prompt

[foggy kiln]

So it don’t really prove anything it’s all 50/50

[hasty socket]

You are conflating "every LLM has a system prompt" with "every LLM has benchmarks in the system prompt". These two things are extremely different,

[foggy kiln]

Scam Altman is behind this

https://tenor.com/view/elon-musk-elon-musk-elonmusk-fun-elon-gif-3737681940171577874

[frosty rapids]

you're a child using it for free, is that it?

it's pretty good in that case

[magic stream]

wtf? (the "You started by saying "I'm here to help. What can I assist you with?")

[foggy kiln]

No I’m actually not, lol I have Claude pro subscription

Claude is my love

The only reason why I like this new model is because it matches Claude

That’s the only reason

[frosty rapids]

ok. well, you're not the target of the scam, so, have fun

[vapid ibex]

He is jai user judging by his Twitter profile

Which is worse

And explains overjoy of another moderately capable model = non-power user

[magic stream]

https://tenor.com/view/khamenei-gif-26128028

i suppose this marks the end of this saga as i don't expect it to release officially on openrouter

[soft sparrow]

it's so fun how this is even being discussed, even the MPU part

of course it doesn't exist

it's literally Cerebras inference

[hasty socket]

The LMArena Discord is eating this up

[vapid ibex]

CPU = Cerebras Powered Unit

[swift mica]

Why does the system prompt have a crypto address for donations

[hasty socket]

As in, they're siding with Momentum lol

[foggy kiln]

lets see this

[soft sparrow]

dude, it's just a fine tuned model

[hasty socket]

Pls put coin in my tip jar so I can beat Sama ty

[soft sparrow]

open source models are also that good

[vapid ibex]

You wanna put your 'tip' in a 'jar'?

[hasty socket]

Don't

[vapid ibex]

https://tenor.com/view/may-ggst-gif-124807981865050115

[foggy kiln]

KIMI K2

[swift mica]

I have the most powerful quantum computer in the world, developed by me and me only, in my possession

I'm running my custom-trained ASI on it

[soft sparrow]

ok i believe you

[swift mica]

Exactly

[vapid ibex]

I trust you, where can I pay for access? I have $3

[swift mica]

I keep it for me and my cat

[soft sparrow]

you have the expert role so i trust you

[swift mica]

I translate the meows and humanspeak back and forth

[hasty socket]

You can access mine if you give me bitcoin

[swift mica]

I believe you

[frosty rapids]

TRUST IN YOUR EXPERTS

[magic stream]

it doesn't matter as long as their model, their product, is real, which doesn't seem.

explained best

[vapid ibex]

There are so many in this thread, it's like a Mixture of Experts here

[soft sparrow]

yeah that's basically what we said on day 1

lol

[frosty rapids]

https://tenor.com/view/right-gif-17351350853737039166

[foggy kiln]

so its an actual model of their own hosted somewhere fast fine tuned

[swift mica]

Ok I can see that they have some model running. They already proved that. Do they have some custom hardware running though? Is it really their model?

[foggy kiln]

if so that would still sell

[soft sparrow]

YEAH!

[foggy kiln]

why don't they just say that if thats the case

[soft sparrow]

YES!

that's the issue congrats

[foggy kiln]

ohhhh

[magic stream]

deepseek tuned on claude/gpt-4o hosted in cerabras is my best guess at this point

[foggy kiln]

😂

[hasty socket]

I would have no issue with a company reselling a finetuned model

[soft sparrow]

they're trying to sell themselves as the next step in AI

when it's not

of course that draws more attention than a fine tuned model

[swift mica]

Because lies sell better

[vapid ibex]

They want to get some flash attention to them

[foggy kiln]

true true

i still remeber when they said grok 4 is agi

[magic stream]

true.

[vapid ibex]

As I said before: it definitely one of the models of all time

[frosty rapids]

bro is trying to get scammed on discord but they don't even notice him 😭