/Tech6h ago

Meta Reportedly Trains Mythos-Level AI Model For Q3 Or Q4 Release

505262810987.2K

Original post

Behemoth reborn. When Muse Spark released I said it would probably come in four sizes, Spark was only the smallest version. Mythos changed everything. Now that they know that it's possible, OpenAI, xAI, META, and anyone else with enough compute to pull it off are all attempting to successfully train very large very performant models. xAI has a 10T Grok in training right now. OpenAI apparently also currently has a giant in training as well, and I don't mean GPT-5.6, I mean a new Orion. The future is big, expensive, and compute-intensive. These fat models will all arrive over the rest of this year. I certainly wouldn't be surprised if META has already pulled it off. They have the compute. Zuck saw this all coming back in 2023, and knowing something is possible is enough to change everything.

Dan McAteer@daniel_mac8

I got a big, fat, juicy AI rumor from my friend.

Apparently Meta trained a Mythos-level model.

And Zuck and team are so back.

No timeline yet but probably Q3/Q4 release.

8:07 AM · Jun 23, 2026 · 76.1K Views

Sentiment

Positive users are excited about Meta's reported Mythos-level AI model and potential stock gains, while negative users distrust Meta after past issues and doubt the Q3/Q4 timeline.

Pos

81.8%

Neg

18.2%

11 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Posts from X

Most Activity

VIEWS9.1KBOOKMARKS6LIKES54REPLIES3

Andrew Curran@AndrewCurran_

More evidence from Lisan and the head of the META Superintelligence Lab himself:

Lisan al Gaib@scaling01

I'm now very hyped for the bigger models Meta will release. They are so fucking back.

they literally say "larger modelS"

so we might get 2 additional tiers above Muse Spark: meaning a 1-2T model and a 5-10T model

5h9.1K546

Andrew Curran@AndrewCurran_

I think Orion/Arrakis, aka GPT4.5, scared everyone off. There were many other reasons as well, no one wanted extremely large extremely expensive models until inference dropped enough to make them affordably servable. There were many incentives to hold off on going big. Looking at it this way, someone with a suspicious mind might suspect that Mythos was partially a poison pill for the entire industry on top of everything else.

6h709146

Chris@ChrissGPT

@AndrewCurran_ @daniel_mac8 So there’s actually two 1 was supposed to come around July or August that would actually be able to beat Fabel and most things slightly and then a couple months after that get a monster model and those two were supposed to come after 5.6

6h4.8K343

Dan McAteer@daniel_mac8

@AndrewCurran_ Maybe there was actually a pre-training law but the labs figured out how to break through it...

6h1K32

Andrew Curran@AndrewCurran_

@ChrissGPT @daniel_mac8 Fatboy Winter!

6h79612

Andrew Curran@AndrewCurran_

meowbooks@meowbooksj

let’s assume for a moment that they have successfully trained a model of this scale and capabilities

would that explain the rest of their recent actions

hmmm

2h2K71

Andrew Curran@AndrewCurran_

@daniel_mac8 It's bad optics for the entire industry if the top models are only affordable by the very wealthy.

6h2283

Dan McAteer@daniel_mac8

@AndrewCurran_ I like the way you think.

6h1721

im here so i wont get fined@plumberbutt97

@AndrewCurran_ That’s from April, Andrew…

5h791

Bandit@reliabytes

@AndrewCurran_ @ChrissGPT @daniel_mac8 You might need to break the glass on one of these names early.

6h661

Dan McAteer@daniel_mac8

@AndrewCurran_ True, although that pattern is historically typical for new general purpose technologies.

6h1394

Andrew Curran@AndrewCurran_

@plumberbutt97 Yes.

5h130

cheaty@cheatyyyy

@AndrewCurran_ active params are still pretty low right? 1% of total params? these are super sparse models?

6h3222

Kevin Faircloth@Asset_sMind

Full disclosure, my post extrapolating of what some people say, is not reliable. The Meta Mythos-level rumors are based on homogeneous Nvidia H100s (~16k) then a heterogeneous switch to AWS’s Trainium for cost-optimized training. (So, infer that Meta’s Mythos-level is quite less expensive than AnthropicAI). A very very small team did this!

6h913

moozooh@moozooh

@AndrewCurran_ Onward to intelligence too expensive to meter!

6h2812

𝕱𝖚𝖑𝖑 𝕶𝖊𝖑𝖑𝖞@full_kelly_

@AndrewCurran_ Assuming "Mythos-level" means parameter count, I believe it... Anthropic is dominating on intelligence right now because Claude Code does the best at accelerating self-improvement. It's not like the concept of a 10T model was a surprise.

6h5421

Dan McAteer@daniel_mac8

@AndrewCurran_ *wall

6h3831

Hussain Hashim | Building Sunday Back@itsthedonhashim

@AndrewCurran_ @AndrewCurran_ wild to think how fast this stuff's evolving. wonder if the next step is gonna surprise us even more. haven't kept up with Mythos, what's the big game changer there?

6h822

Pulkit@puhlkit

@AndrewCurran_ I don’t agree with “now that they know it’s possible”. They had to have always known. Hopefully…

6h2211

Haseeb Arshad@HaseebArshaddd

@AndrewCurran_ They are masterpiece.

6h1311