GPT-Oss 120b Runs at 3000 Tokens/sec on Cerebras

Posted2 months agoActiveabout 2 months ago

samspenc

48 points

29 comments

cerebras.aiTechstory

excitedpositive

Debate

20/100

AICerebrasGPT-Oss

Key topics

Cerebras

GPT-Oss

The GPT-OSS 120B model runs at 3000 tokens/sec on Cerebras, sparking excitement and discussion about its potential applications and performance.

Snapshot generated from the HN discussion

Discussion Activity

Moderate engagement

First comment

Peak period

6-9h

Avg / period

4.1

Comment distribution29 data points

Loading chart...

Based on 29 loaded comments

Key moments

01Story posted
Nov 7, 2025 at 10:24 PM EST
2 months ago
Step 01
02First comment
Nov 8, 2025 at 5:30 AM EST
7h after posting
Step 02
03Peak activity
10 comments in 6-9h
Hottest window of the conversation
Step 03
04Latest activity
Nov 10, 2025 at 12:04 AM EST
about 2 months ago
Step 04

Generating AI Summary...

Analyzing up to 500 comments to identify key contributors and discussion patterns

Discussion (29 comments)

Showing 29 comments

freak42

2 months ago

5 replies

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.

traceroute66

2 months ago

2 replies

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.

Saline9515

2 months ago

2 replies

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.

freak42

2 months ago

2 replies

Yes they can ask, but do it at the beginning not the end of the process, this is a dark pattern and fucking annoying.

magackame

2 months 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?)))

traceroute66

2 months 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.

traceroute66

2 months ago

1 reply

> 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.

Saline9515

2 months 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.

cyanydeez

2 months 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.

anonym29

2 months ago

1 reply

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.

freak42

2 months ago

1 reply

You didn't get my point at all.

rpdillon

2 months 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.

schappim

2 months ago

I was doing a demo to my colleagues and had the above.

moralestapia

2 months 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.

Alifatisk

2 months 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.

petesergeant

2 months ago

1 reply

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.

jsheard

2 months ago

3 replies

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.

petesergeant

2 months ago

1 reply

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.

rajman187

2 months ago

1 reply

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...

petesergeant

2 months 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.

rpdillon

2 months ago

1 reply

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.

aurareturn

2 months ago

1 reply

  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.

rpdillon

2 months ago

1 reply

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?

aurareturn

about 2 months 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.

7thpower

2 months ago

Switching costs are low, so if that happens we’ll just switch.

iFire

2 months ago

Does anyone know how much one system costs?

sunpazed

2 months 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.

drewbitt

2 months 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

KronisLV

2 months 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.

View full discussion on Hacker News

ID: 45853849Type: storyLast synced: 11/20/2025, 8:00:11 PM

Want the full context?

Jump to the original sources

Read the primary article or dive into the live Hacker News thread when you're ready.

Open link View on HN