GPT-OSS 120B Runs at 3000 tokens/sec on Cerebras(www.cerebras.ai)

45 pointsby samspenc2 days ago6 comments

KronisLV2 days ago
The Cerebras GML-4.6 post might also be of (some?/more?) interest to the people here, since it's more useful for programming: https://news.ycombinator.com/item?id=45852751
I don't think that this is a dupe or anything and 3000 t/s is really cool, the other post just has more discussion of Cerebras and people's experiences with using GLM 4.6 for software development.
drewbitt2 days ago
It's a decent general model too - I have it plugged up in llm and raycast since August at great speeds. I wish Cerebras would do MiniMax M2 which should be an upgrade and replacement if it was just faster. It would never be as fast as gpt-oss-120 though
sunpazed2 days ago
This is really impressive. At these speeds, it’s possible to run agents with multi-tool turns within seconds. Consider it a feature rich, “non-deterministic API” for your platform or business.
- 2 days ago
  undefined
freak422 days ago
I absolutely hate it, when a website says "try this" and after you went through the trouble of weiting something comes up with a sign up link first. Makes me leave instantly to never come back.
- traceroute662 days ago
  Headline at the top of the Cerebras page linked to by the OP "Cerebras Raises $1.1B Series G at $8.1B Valuation".
  If you're going after the AI money gravy train then you need to wave the "we have $n registered users" carrot on your PPT slides for the investors because registered user == monetization opportunity.
  I'm not defending it. I hate being forced to register for shit when I just want to try it or use the free tier.
  But it is what it is.
  - Saline95152 days ago
    Well if they give it out for free (aka they pay for it), asking you to register is a reasonable ask. It's not a public service funded by taxpayers.
    freak422 days ago
    Yes they can ask, but do it at the beginning not the end of the process, this is a dark pattern and fucking annoying.
    magackame2 days ago
    Anyone remember those online psychological tests where you spend an hour on one and in the end you need to pay up to get the result?)))
    traceroute662 days ago
    > do it at the beginning not the end
    Exactly this.
    If you present me with a form and a submit button then I expect the input to go through and a result to be presented.
    If you don't want to present me with results before login, then put the form behind the wall too.
    Simple.
    traceroute662 days ago
    > Well if they give it out for free (aka they pay for it), asking you to register is a reasonable ask
    They have other options... rate limiting, serving (more) quantized to non-registered etc. etc.
    Saline95152 days ago
    Those options are still not free. And giving a degraded version of your product to free users is a bad way to acquire clients.
  - cyanydeez2 days ago
    Right, being proud of your money making is not something I consider a consumer focused product unless that customer is other moneyseeking orga, which like cancer, often ends up in a bubble.
- Alifatisk2 days ago
  Same with groq.com, there is a "try this", and after you enter the prompt, it asks you to sign in. Closed the page.
- schappim2 days ago
  I was doing a demo to my colleagues and had the above.
- moralestapia2 days ago
  Off topic but related.
  A week ago I went to a launch party for a product that's supposed to "revolutionize design" (a web app w/ an OAI prompt).
  No demo, only like two pictures of the actual product. Founder spent like half an hour giving a speech about the future, etc...
  "All of you here will get access to it in a couple weeks."
  Couple weeks go by ... I "get access". It's a .dmg, 1) What, I open it, it's not even an app, it's an installer ..., I install it, the app opens up and it's a giant red button that takes you to a website to create an account ...
  These guys are completely lost.
- anonym292 days ago
  This is like declaring that a Ferrari dealership offering you a free test drive in a million dollar art exhibit on wheels is evil for asking for your phone number before handing you the keys.
  If this was some beat-to-hell, high-mileage used economy car, sure, that would be a pain in the ass, and not worth it. But it's a mistake to place Cerebras into that mental bucket.
  You don't even need to use real information to create an account. Just grab a temp-mail disposable address and sign up as fred flintstone or mickey mouse.
  If you're a heavy LLM inference user (i.e. if you've ever paid for a $200/mo sub from any of the big AI labs), I can damn near guarantee you will not regret trying out Cerebras.
  - freak422 days ago
    You didn't get my point at all.
    rpdillon2 days ago
    Would your expectations be more aligned if it's said "free trial"? That might create an expectation of a sign up where "try this" might not.
petesergeant2 days ago
It’s an absolute beast. I run it via OpenRouter, where I have Groq and Cerebras as the providers. Cheap enough as to be almost free, strong performance, and lightning fast.
- jsheard2 days ago
  Cheap enough for now, but of all the companies selling inference at a loss, Cerebras and Groq are probably losing the most per token. Their hardware is ungodly expensive and its reliance on huge amounts of SRAM bottlenecks how much cheaper it can get, since SRAM density is improving at a snails pace at this point.
  - rpdillon2 days ago
    You're pointing out a bunch of high capex costs (hardware, SRAM), but then concluding that their opEx is greater than their revenue on a per unit basis. Are they really losing money on every token? It seems that using hardware acceleration would decrease inference costs and they could make it up on unit economics over time.
    But I'm just reasoning from first principles. I don't have any specific data about them.
    aurareturn2 days ago
    It seems that using hardware acceleration would decrease inference costs and they could make it up on unit economics over time.
    Nvidia GPUs are accelerators too. The reason they can do this so fast is because they're storing entire models in SRAM.
    rpdillon21 hours ago
    There are degrees of acceleration. My understanding, limited as it is, is that groq and cerebras are using highly optimized acceleration to achieve their token generation rates, far beyond that in a regular GPU, and this leads to lower costs per token.
    Is this incorrect?
    aurareturn5 hours ago
    Yes, they're called ASICs on Grog. But Cerebras has more general cores that can do more complex things. Inference is mostly limited by bandwidth though.
  - 7thpower2 days ago
    Switching costs are low, so if that happens we’ll just switch.
  - petesergeant2 days ago
    Not doubting you but anything to back that up? Happy enough to burn VC money until someone shows up who can run it without losing money, either way.
    rajman1872 days ago
    They’ve filed a S1 [1] last year when attempting to go public. It showed something like a $60M+ loss for the first 6 months of 2024. The IPO didn’t happen because the CEO’s past included some financial missteps and the banks didn’t want to deal with this. At the time the majority of their revenue came from a single source in Abu Dhabi, as well
    [1] https://www.sec.gov/Archives/edgar/data/2021728/000162828024...
    petesergeanta day ago
    > the majority of their revenue came from a single source in Abu Dhabi, as well
    I live in UAE, whose continuing enthusiasm in AI investment stretches well beyond short-term profit, so having AD on-board seems like a plus not a minus. I'm sure there are specific exceptions, but generally Emirati money has seemed like smart money.
iFirea day ago
Does anyone know how much one system costs?