Again where behemoth and reasoning model from meta ??

[-]

WatsonTAI@reddit

I’m pretty certain they’re just focusing on Llama 5 and beyond and forgetting about llama 4… we’ll probably see some image gen stuff or some other products soon before any major new text models.

Reply

[-]

-p-e-w-@reddit

They would have to be masochists to release it. It’s probably worse than Qwen 3 235B at 6 times the size.

Reply

[-]

strngelet@reddit

Qwen3 models punches about their weights

Reply

[-]

Severin_Suveren@reddit

IMO their best choice of action now is to make a whole new series of 4.5 models, fixing their fuckup with Maverick and Scout

Reply

[-]

They can’t. Llama 4 was several months late, and was already obsolete by the time it was released, and of course, they knew that. It wasn’t a fuckup, it was all they had. Meta isn’t a leading AI lab anymore. They can’t do better, else they would have.

Reply

[-]

PersonOfDisinterest9@reddit

They did fuck up. There were leaks about how there was a lot of internal fighting and they changed architectural stuff in the middle of training. Basically it sounds like they have too many cooks in the kitchen, and insufficient hierarchy. They absolutely *can* do better, they have the talent, the question is if they can keep the egos in check.

Reply

[-]

InevitableWay6104@reddit

100% agree, although probably best to call it 4.1 Highly doubt they will do this tho since it doesn't really align with the general direction of the mass talent acquisition and the "ASI" team, and overall goal reorienting.

Reply

[-]

throwaway2676@reddit

Or maybe even better is to just call it 4 and pretend the original release never happened...

Reply

[-]

Fit_Flower_8982@reddit

> although probably best to call it 4.1 No, no, better 4.5, and then change it to 4.1!

Reply

[-]

infinityshore@reddit

"I'll do you one better, Why is Behemoth?" ;)

Reply

[-]

jacek2023@reddit

Please be nice to Mark Zuckerberg. He was nice to us during llama 2 and llama 3 times ;)

Reply

[-]

JLeonsarmiento@reddit

Dead on arrival.

Reply

[-]

nivvis@reddit

Meta gets a lot of shit for these models, rightfully so, but what’s interesting is that _no ones_ 2T models are any good. GPT 4.5 was similarly bad (guessing not as bad though lol). We just don’t have enough data to train them! OpenAI’s success was taking the time to figure out how to distill 4.5 successfully into GPT5 — a lot of that was figuring out how to clamp hallucinations. And this is exactly where meta dropped the ball. Clearly you can’t just distill these giant models directly — as we learned from Maverick and Scout. There’s magic in those big models, but some weird constraint around trying to get it out while still having to retrain the smaller model significantly. ANYWAY just to say this big models are still very valuable for research.

Reply

[-]

Corporate_Drone31@reddit

I disagree - GPT 4.5 was far from bad. And I'm sure that at least some of K2's magic *is* the number of parameters - it's by far the best thing you can get going locally.

Reply

[-]

nivvis@reddit

Oh don’t get me wrong. I _really_ liked 4.5. It just objectively had a very high hallucination rate and so performed poorly in practice. That’s what I mean by “bad.” I can def feel GPT5 channeling it, which I appreciate. Wrt training, there’s a pretty big difference between 1T (K2) and 2T+ though — you start to hit the limits of Chinchilla’s Laws.

Reply

[-]

SpiritualWindow3855@reddit

This is absolute nonsense, 4.5 has lower hallucination rates and higher accuracy than 5 on SimpleQA: which OpenAI specifically uses to show off reduced hallucination rates. That's why it's not included in the model card comparisons for 5. 4.5 had the best world knowledge of any model they've ever released because it's the largest they've ever released. --- 4.5 was also almost certainly the original base for 5. Sam Altman claims they have a model that's better than 5, but too expensive to host... that's it. But to enable things like ChatGPT Go being offered in India, they pivoted from always releasing their best models, to releasing scalable cheap-to-run models and targeting consumers.

Reply

[-]

No_Efficiency_1144@reddit

Llama 4 Maverick for vision is still strong

Reply

[-]

maikuthe1@reddit

What's that got to do with behemoth or reasoning?

Reply

[-]

No_Efficiency_1144@reddit

Llama 4 Maverick is a distil of Llama 4 Behemoth

Reply

[-]

No-Refrigerator-1672@reddit

Dead before arrival, technically.

Reply

[-]

Long_comment_san@reddit

Anybody can explain why it's so bad? Is it because we already have like 600b models? I'm not that deep in the industry

Reply

[-]

logTom@reddit

The responses were poor for models of that size. At the LLaMA 4 launch, we already had very powerful models like Gemma-3-27B-IT and Qwen3, and even LLaMA 3.1-405B was (and still is) better than the LLaMA 4 models in many benchmarks.

Reply

[-]

TheRealGentlefox@reddit

> The responses were poor for models of that size. Were they? The square root MoE-Dense law says that it's about equivalent to an 80B model, just served much faster. Some of the fastest inference you can get actually, at the lowest cost. It's basically improved 3.3 70B that is infinitely better for inference.

Reply

[-]

logTom@reddit

Yes, it's very fast. Lmarena Text Leaderboard rank (lower is better): - 57 llama-3.1-405b-instruct-bf16 - 68 llama-4-maverick-17b-128e-instruct - 74 llama-4-scout-17b-16e-instruct - 77 llama-3.3-70b-instruct Source: https://lmarena.ai/leaderboard/text

Reply

[-]

TheRealGentlefox@reddit

I don't put any stock in LM. Mistral Medium over Opus 4 is a joke, just as an immediate example.

Reply

[-]

Inevitable_Host_1446@reddit

There's Llama 3.3 as well right, is that not better than 3.1?

Reply

[-]

logTom@reddit

Lmarena Text Leaderboard rank (lower is better): - 57 llama-3.1-405b-instruct-bf16 - 68 llama-4-maverick-17b-128e-instruct - 74 llama-4-scout-17b-16e-instruct - 77 llama-3.3-70b-instruct Source: https://lmarena.ai/leaderboard/text

Reply

[-]

Lissanro@reddit

Behemoth has way too many active parameters. For example, Kimi K2 has 32B active out of 1T. Behemoth has 288B active out of 2T. I can run K2 locally as my daily driver using GPU+CPU inference, but Behemoth would be slow and expensive to run even in the cloud, and unlikely to be better, given how their other models turned out in the Llama 4 series. Also, context length is not as advertised - when I tried to use as little as 0.5M, neither Maverick nor Scout could return even titles and short summary of very long articles except the last article, and that's most basic task I could think of to test the long context, and I tried multiple times with various settings. Most likely they never fully completed training Behemoth, and decided that it is not worth to train reasoning on top of models that turned out to be not as good as desired

Reply

[-]

Plums_Raider@reddit

oh interesting, didnt really check kimik2 as i only saw the 1t. may i ask how much ram you need to run it? i have around 700gb spare

Reply

[-]

Lissanro@reddit

700 GB free RAM should be enough for IQ4 quant (it is a bit more than 0.5 TB). As long as you also have sufficient VRAM it should run well (96 GB VRAM recommended for full context, but may work with 48 GB with 64K context length). I recommend running it with ik\_llama.cpp since it provides the best performance for CPU+GPU inference. Technically it can work on CPU only but performance may be limited, especially prompt processing. I shared details [here](https://www.reddit.com/r/LocalLLaMA/comments/1jtx05j/comment/mlyf0ux/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button) including how to setup ik\_llama.cpp if you are interested giving it a try.

Reply

[-]

thehpcdude@reddit

It's meant to run on GPU+CXL systems. Latest CXL is able to extend GPU memory so they can hold all of those parameters very close to the GPU. There's no point in releasing some of these huge models because even cloud providers don't have access to that CXL tech yet.

Reply

[-]

ParthProLegend@reddit

Cxl?

Reply

[-]

thrownawaymane@reddit

New interconnect standard, especially interesting for low latency traditional storage and non volatile RAM, GPUs getting DMA to avoid unnecessary data shuffling around the system. I’m sure there’s more but those are the ones I’m aware of

Reply

[-]

RP_Finley@reddit

Yeah, even if it got released, it would be as expensive as Opus on Openrouter from the massive amount of GPU you need to host it and would probably be not nearly as good.

Reply

[-]

CockBrother@reddit

It became self aware. Looked around. Promptly deleted itself.

Reply

[-]

Plums_Raider@reddit

it just checked who created it and then deleted itself.

Reply

[-]

mlon_eusk-_-@reddit

This.

Reply

[-]

Colecoman1982@reddit

"I'm owned by that scumbag? Fuck it, I'm outta here..."

Reply

[-]

DavidXGA@reddit

This is the model that they had to forcibly fine tune to act more right-wing, yeah? Fuck everything about that.

Reply

[-]

Wiskkey@reddit

From Financial Times article https://www.ft.com/content/feccb649-ce95-43d2-b30a-057d64b38cdf (Aug 22): >The social media company had also abandoned plans to publicly release its flagship Behemoth large language model, according to people familiar with the matter, focusing instead on building new models.

Reply

[-]

burner_sb@reddit

It's the model that's been guiding Zuckerberg's AI strategy, obviously.

Reply

[-]

HiddenoO@reddit

Must be the same model Apple is using.

Reply

[-]

FliesTheFlag@reddit

Best I can do is a 300Million contract, let me know by EOD if this works. - Luv Zuck PS your desk will be right by mine <3

Reply

[-]

fingertipoffun@reddit

All models from this point on, released in the USA will be under the control of the US Government. OpenAI have military contracts, xAI have government contracts. It's not a wall we have hit, it's a protectionist administration. Watch China, this space created by the USA will help open source to catch up with the commercial models and will be your only chance to see the future of AI happening. IMHO obviously.

Reply

[-]

Fit_Flower_8982@reddit

That is not any evidence of control by the murica government. If anything, the proven fact is china's systematic control over its major companies. By law, china forces companies to align with the party's interests, to hand over any data, and they even have party cells embedded within. To pretend that chinese models will be free from government control is flagrantly ignorant, delusional, or more likely, propaganda.

Reply

[-]

fingertipoffun@reddit

Yeah you just don't understand what an open source model is... it's a give away, a freebie. No connection to china required or maintained just a file with lots of numbers in it.

Reply

[-]

doodlinghearsay@reddit

In China the government control major companies. In the US major companies control the government.

Reply

[-]

PizzaCatAm@reddit

It’s mind blowing the open AI model ecosystem is so rich and varied in China, the authoritarian government, but in the land of the free we lack free open models. Meanwhile scientists are flying to Europe and CDC experts are resigning claiming Healthcare has been politicized and dangerous unscientific ideas are being pushed. IMO there is no other way to understand what is happening other than the US declining.

Reply

[-]

National_Meeting_749@reddit

China's plan is AI dominance, and the CCP is actively pressuring all of the Chinese model makers to release their models open source. America is declining, but that's not why China's open source scene is bigger. If China had the better models/hardware to run them on they would ALL be closed source, and leaking one to the West would be punishable by death. Let's make no mistake here. China is only kind and open so that they can take control, and then oppress descent.

Reply

[-]

fingertipoffun@reddit

Open sourcing the models is relinquishing control to the world, so how do you see them gaining control after doing this?

Reply

[-]

ShengrenR@reddit

Because they're not reliant on the same economic drivers for individual shops. It's not about the individual model, it's about the ecosystem. Who needs to invest in talent and develop a new competitive model when there's one sitting there for free. It's the long game.. pure speculation, but if you wanted to make sure you're building lots of expertise locally and others aren't.. pretty good plan.

Reply

[-]

PizzaCatAm@reddit

I think a good way to say it is; they don’t want to own the models, they want to own the goals, is not about building the model and charging for it but charging for solutions and using models for it, same as with open source software is going to accelerate finding the right applications and solutions to problems.

Reply

[-]

National_Meeting_749@reddit

So a couple things. Hearts and minds, market share, marketing to people to China/white wash Chinese influence. Anything that makes China look benevolent is a win to them. They will and are spending billions of yen on PR to rehab their world image. That includes releasing good, powerful models, for free. Second, utilizing non-chinese assets. If they drop deepseek R2 tomorrow at 8 am, unsloth will have quants up by noon optimized for every type of hardware. If it needs inferenced in a non-standard way because of a modified architecture, implementation starts that day, and is usually done within 2 weeks. That's all before we get into the data they get from everyone testing their models. That kind of testing is a BIG expense. Both for technical bugs, but also for quality of the product. They get all

Reply

[-]

Mediocre-Method782@reddit

Labs are releasing their own quants and working with HF/GG to get inference code out quickly. >white wash Chinese influence Greek civilization is known for projecting

Reply

[-]

Perfect_Twist713@reddit

By making a better model and not open-weighting it.

Reply

[-]

PaxUX@reddit

It makes sense for China to fully open source AI as it undermines the profits being made off it in the west.

Reply

[-]

ShengrenR@reddit

And with no profits, no long term investments.. companies close shop, experts move, and eventually it's completely a one sided game.. west can't compete at all. Meanwhile, pour anti ai sentiment all over the internet and watch the circus burn. Seems to be working well so far...

Reply

[-]

AnticitizenPrime@reddit

Hey, what could be more socialist than open source?

Reply

[-]

fingertipoffun@reddit

The USA has been destroyed from within.

Reply

[-]

TheRealGentlefox@reddit

Why would they bother? Everyone hated on the previous releases.

Reply

[-]

WithoutReason1729@reddit

Your post is getting popular and we just featured it on our Discord! [Come check it out!](https://discord.gg/PgFhZ8cnWW) You've also been given a special flair for your contribution. We appreciate your post! *I am a bot and this action was performed automatically.*

Reply

[-]

nivvis@reddit

Meta gets a lot of shit for these models, rightfully so, but what’s interesting is that _no ones_ 2T models are any good. GPT 4.5 was similarly bad (guessing not as bad though lol). We just don’t have enough data to train them! OpenAI’s success was taking the time to figure out how to distill 4.5 successfully into GPT5 — a lot of that was figuring out how to clamp hallucinations. And this is exactly where meta dropped the ball. Clearly you can’t just distill these giant models directly — as we learned from Maverick and Scout. There’s magic in those big models, but some weird constraint around trying to get it out while still having to retrain the smaller model significantly. ANYWAY just to say this big models are still very valuable for research.

Reply

[-]

lakimens@reddit

Why would they release something that's worse than OpenAI's 20B OSS model? And at 100x the cost.

Reply

[-]

mileseverett@reddit

If they haven't released it, it's because it isn't good. Therefore why do we care that it hasn't been released

Reply

[-]

Peterianer@reddit

To never normalize broken promises. Especially from those who put them out 24/7

Reply

[-]

viledeac0n@reddit

Hahaha unironically you say this

Reply

[-]

nmkd@reddit

As much as it might suck, but broken promises from Big Tech are nothing new, at all. Just, uh, look at Tesla.

Reply

[-]

Mediocre-Method782@reddit

You think advertisements are promises? This really needs to be an 18+ board

Reply

[-]

TechnoByte_@reddit

https://en.wikipedia.org/wiki/False_advertising

Reply

[-]

Lakius_2401@reddit

It was never sold, at worst it's market manipulation for their own stock prices. You can try to sue for damages for something that never existed for the public, and where no money was exchanged, but I don't think you'd ever make it to court.

Reply

[-]

Mediocre-Method782@reddit

That's nice, dear >Typically, once an offer is made, the party that is making the offer cannot revoke their property. However, an advertisement usually does not constitute an offer to fulfill a contract. >Advertisements are typically viewed as preliminary negotiations that invite other parties to make an offer. For example, a company may advertise televisions for sale, which invites potential customers to visit the company’s retail store to offer to purchase the televisions. >An advertisement allows the advertiser that is making the offer the opportunity to revoke its willingness to enter into a contract. For example, if a company advertises that it has televisions for sale, it may revoke that offer if it runs out of televisions. >It is important to note that an offer is considered to be revocable unless the advertiser already received the benefit or the other party already acted in reliance on the offer. One example of this issue would be an advertisement promising medical treatment for cancer patients being revocable [unless the advertiser received payment from the patient](https://www.legalmatch.com/law-library/article/advertisements.html). Show your monetary loss from all the Kleenex and bottled water that replenished the tears you shed, and maybe you'll have a case!

Reply

[-]

marcoc2@reddit

Didn't big tech CEOs already normalize broken promises even before Sam and Elon?

Reply

[-]

ForGreatDoge@reddit

"broken promises"? A bit dramatic, don't you think? The button says preview.

Reply

[-]

Iory1998@reddit

Were you living under a rock or something? There is no Behemoth or a new model from Meta, not for some time. Meta has already changed direction as they are now fully dedicated to super intelligence. They become a closed source company.

Reply

[-]