Someone from NVIDIA made a big mistake and uploaded the parent folder of their upcoming model on Hugging Face

[-]

xXG0DLessXx@reddit

I hope someone saved the stuff that might get taken down

Reply

[-]

rerri@reddit

[https://huggingface.co/xeophon/NVIDIA-Nemotron-Nano-3-30B-A3B-BF16/tree/main](https://huggingface.co/xeophon/NVIDIA-Nemotron-Nano-3-30B-A3B-BF16/tree/main)

Reply

[-]

mikael110@reddit

And it's dead. Hopefully somebody managed to get it all and we'll get a magnet link or something like that do download.

Reply

[-]

Forward-Addition5126@reddit

[https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/tree/main](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/tree/main) yep it is

Reply

[-]

YESSSS ITS ON THE WAYBACK MACHINE [https://web.archive.org/web/20251212123038/https://huggingface.co/xeophon/NVIDIA-Nemotron-Nano-3-30B-A3B-BF16/tree/main](https://web.archive.org/web/20251212123038/https://huggingface.co/xeophon/NVIDIA-Nemotron-Nano-3-30B-A3B-BF16/tree/main)

Reply

[-]

cafedude@reddit

If only. Can you imagine how much storage space they'd have to have over at the Internet Archive in order to do that?

Reply

[-]

Straight_Abrocoma321@reddit

"Technology: We preserve 100 million web pages per day! So far, we've saved 45 petabytes (that's 45,000,000,000,000,000 bytes) of data."

Reply

[-]

Yes_but_I_think@reddit

1875 x 24TB drives, ok.

Reply

[-]

mehupmost@reddit

You cannot download the files from there.

Reply

[-]

GenLabsAI@reddit

aw fuck it!

Reply

[-]

Straight_Abrocoma321@reddit

nvm none of the folders are on there

Reply

[-]

xrvz@reddit

duh

Reply

[-]

banyudu@reddit

404 now

Reply

[-]

alongated@reddit

Seems like this is deleted. The fact they just delete stuff like this, makes them quite unreliable.

Reply

[-]

Lydeeh@reddit

The fact that they delete leaked stuff makes them RELIABLE, not unreliable. What are you on about.

Reply

[-]

alongated@reddit

It failed at its prime function, to share models. If a phone stops working due to the whims of the creator, that is incredibly unreliable.

Reply

[-]

Lydeeh@reddit

Your logic is flawed and the comparison doesn't even make sense. Hugging Face didn't stop working. It only worked as intended. By sharing models that are meant to be shared and keeping models that are meant to be kept private, private. The NVIDIA model wasn't meant to be shared yet.

Reply

[-]

alongated@reddit

It makes them unreliable for us, we can not rely on it if we are going to store models, that are for example, leaked.

Reply

[-]

Lydeeh@reddit

Yes bro. One of the most reputable model hosting sites will start hosting leaked models to please alongated's childish entitlement. That would definitely make all the model creators respect Huggingface and post there more often. /s

Reply

[-]

alongated@reddit

Why are you using multiple accounts? Did you leak this model Lydeeh? Or are you working for huggingface? You are very sus with your wordings. But on to your point, we aren't here to fix their problems, just our own, if they can't serve our function we should just start using torrents instead.

Reply

[-]

Lydeeh@reddit

The fact that multiple people are calling you out on this, doesn't mean that I have multiple accounts. My point is: Hosting sites need to abide by certain standards if they want creators to use them. It's not us the end users that make these sites possible, its the creators. If these serious sites didn't exist the model scene would be way different, having multiple smaller hosting sites each with different models and most likely behind paywalls. So be grateful that we're getting these for free all in one place and have some patience and wait for the model to be released when it's ready.

Reply

[-]

alongated@reddit

Maybe clean up the accounts before you comment with them next time https://old.reddit.com/user/Kerbal634

Reply

[-]

Lydeeh@reddit

What?

Reply

[-]

alongated@reddit

Why are you advertising for Huggingface. No one here cares for it. No one would mind if it vanished. They would just move on.

Reply

[-]

Amazing_Athlete_2265@reddit

Factually incorrect.

Reply

[-]

alongated@reddit

You sound like a shill. Do you not want a leaked model?

Reply

[-]

Kerbal634@reddit

Hey man so do you put your pants on your head before or after putting underwear on?

Reply

[-]

randylush@reddit

That sounds like the prime function of BitTorrent.

Reply

[-]

alongated@reddit

We really need to start doing that. That means trying to get people to stop posting huggingface links, and move over to torrent.

Reply

[-]

alongated@reddit

I meant huggingface being unreliable.

Reply

[-]

Joe091@reddit

And he meant that’s what makes Huggingface reliable.

Reply

[-]

alongated@reddit

The fact that they delete the model you want to share? A phone that stops working, or a pen that stops writing, no matter the reason is unreliable.

Reply

[-]

KnifeFed@reddit

Your comparisons make zero sense.

Reply

[-]

alongated@reddit

I was trying to say that we can not rely on them.

Reply

[-]

Joe091@reddit

Real companies have legal and ethical obligations to their partners. Your childish and selfish desires do not and should not matter to them.

Reply

[-]

alongated@reddit

It would always be due to those things. It is why Onedrive or Github are unreliable compared to their local alternatives. Why are you defending them, they legit hurt us by deleting this.

Reply

[-]

Joe091@reddit

You were not harmed. I totally get why you would want to download this, but come on, be an adult about it.

Reply

[-]

alongated@reddit

It did hurt

Reply

[-]

KnifeFed@reddit

You sound whiny and entitled.

Reply

[-]

alongated@reddit

Are you not sad that you possibly lost unfettered access to an amazing model? I am quite whiny atm, because of that, and I'm trying to think of a solution to prevent it from happening again. Regarding me being entitled, I would want everything for everyone if it was possible.

Reply

[-]

Lydeeh@reddit

It's not YOUR model. It's other peoples model that you want to share without their permission.

Reply

[-]

alongated@reddit

I never said it was my model. And yes I would want everyone in this world to have all the models in this world, fucking sue me.

Reply

[-]

rerri@reddit

Who's they?

Reply

[-]

alongated@reddit

I don't know. But I do know that they did delete it.

Reply

[-]

Nunki08@reddit (OP)

Wow Xeophon saved everything. Thank you for the link. We could might in trouble, lol.

Reply

[-]

tiffanytrashcan@reddit

And quickly took it down. ☹️

Reply

[-]

mikael110@reddit

I doubt he took it down himself. HuggingFace often takes down mirrors when they get notified about a leaked model.

Reply

[-]

tiffanytrashcan@reddit

https://preview.redd.it/02vxcr0rjs6g1.png?width=1220&format=png&auto=webp&s=235b12c8f86671c4e7ad38c28ed3869fc2d2d170

Reply

[-]

mikael110@reddit

Thanks for that info. I stand corrected. I've edited my comment to clarify.

Reply

[-]

mr_house7@reddit

Link is gone, is there any other link?

Reply

[-]

MetricZero@reddit

Who? Where?

Reply

[-]

No_Conversation9561@reddit

Anything useful in there?

Reply

[-]

o5mfiHTNsH748KVq@reddit

aaaand it’s gone

Reply

[-]

SteelRevanchist@reddit

\`404\` :(

Reply

[-]

dead-supernova@reddit

he must save them locally and upload them them every place because who know it may get take down from huggingface

Reply

[-]

EternalDivineSpark@reddit

They already deleted it or this is take

Reply

[-]

Lissanro@reddit

I think it is quite real to me, and somebody mirrored it here: [https://huggingface.co/xeophon/NVIDIA-Nemotron-Nano-3-30B-A3B-BF16/tree/main](https://huggingface.co/xeophon/NVIDIA-Nemotron-Nano-3-30B-A3B-BF16/tree/main)

Reply

[-]

quisariouss@reddit

Gone.

Reply

[-]

nmkd@reddit

Someone make a torrent for god's sake.

Reply

[-]

yeah-ok@reddit

All decent functional stuff should be on a torrent tracker for obvious reasons!

Reply

[-]

FirmConsideration717@reddit

People have collectively become stupider these past 15 years.

Reply

[-]

yeah-ok@reddit

Think the distraction level has a lot to account for in that a lot of the developmental basics are being missed in exchange for random media/web cr*p. Sooo yeah.

Reply

[-]

EternalDivineSpark@reddit

Yeah !! Crazy ! Suhara is in big trouble now !

Reply

[-]

LordEschatus@reddit

dear HF, deleting mirrors won't save you now.

Reply

[-]

Lissanro@reddit

It is not HF who deleted it, Xeophon took it down themselves: [https://x.com/xeophon\_/status/1999480999017873802?s=20](https://x.com/xeophon_/status/1999480999017873802?s=20) Xeophon wrote: >To those from Reddit: I’ve taken it down myself, I didn’t expect it to get this much attention and don‘t want to get anyone into (more) trouble To HF users: Always, always set --private (you can set it on org-level as well). Mirroring anything from HF is instant + one-liner

Reply

[-]

One-Employment3759@reddit

And also never rely on hugging face. Download it so you can share it later.

Reply

[-]

LordEschatus@reddit

thats fine. that only *slightly* detracts from my point.

Reply

[-]

Full_Way_868@reddit

yall stop thinking like criminals

Reply

[-]

my_name_isnt_clever@reddit

What's criminal is what nvidia has been doing with the GPU market, fuck 'em.

Reply

[-]

Armchairplum@reddit

I'm sure that's directed at the entity and not the person who made the mistake. After all, its not like whatever internal repercussions are made public for the employee.

Reply

[-]

throwaway_ghast@reddit

Won't someone please think of the poor billion dollar companies?!

Reply

[-]

IjonTichy85@reddit

You can't tell me what to do, you're not even my real dad and the day I turn 18 I'm outta here!!

Reply

[-]

Hearcharted@reddit

LOL WHAT XD

Reply

[-]

IjonTichy85@reddit

Just keepin it real in here.

Reply

[-]

Hearcharted@reddit

A Real G In The Hood!

Reply

[-]

Full_Way_868@reddit

That's probably what goes through the minds of most leakers

Reply

[-]

dydhaw@reddit

I AM THE LAW

Reply

[-]

Hearcharted@reddit

Even Judge Dredd is here :)

Reply

[-]

bioshawna@reddit

Ooof I can’t say I haven’t almost made similar in scale type goofs to this 😂

Reply

[-]

Background_Essay6429@reddit

Did they remove it already or can we still see it?

Reply

[-]

GeLaMi-Speaker@reddit

I get the excitement, but I really hope folks don’t treat this kind of accidental upload as a “drop” to be mirrored/spread. NVIDIA (and others) have been unusually transparent lately (weights + reports + training recipes). If every accident becomes a scramble to redistribute, you’re basically training companies to gate harder.

Reply

[-]

kristaller486@reddit

\>Nano \>30B-A3B

Reply

[-]

bomjj@reddit

3b experts could run on cpu with decent token speed, if you have enough ram

Reply

[-]

ThatCrankyGuy@reddit

Nano.. that's like calling a tiger "fluffy"

Reply

[-]

Amazing_Athlete_2265@reddit

Fucking give it to me

Reply

[-]

vasileer@reddit

A3.5B

Reply

[-]

AcanthaceaeNo5503@reddit

Classic HF cli / sdk. Even though, the orga disabled the public repo, u can still upload it publicly. Thats super stupid in terms of security

Reply

[-]

Suitable-League-4447@reddit

so u can donwload it the "NVIDIA-Nemotron-Nano-3-30B-A3B-BF16"

Reply

[-]

AcanthaceaeNo5503@reddit

I can download sooner or later. But the guy will probably be punished though

Reply

[-]

Suitable-League-4447@reddit

no u can't as the link is expired and doesnt exist anymore

Reply

[-]

qfox337@reddit

These "leaks" are almost all fakes to build hype 🙄

Reply

[-]

ihateAdmins@reddit

[https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16) was it released fully or are things still missing out from the current release?

Reply

[-]

No_Dot1233@reddit

shiittt nemotron is good stuff, will have to try if i get my hands on it

Reply

[-]

haikusbot@reddit

*Shiittt nemotron is good* *Stuff, will have to try if i* *Get my hands on it* \- No\_Dot1233 --- ^(I detect haikus. And sometimes, successfully.) ^[Learn more about me.](https://www.reddit.com/r/haikusbot/) ^(Opt out of replies: "haikusbot opt out" | Delete my comment: "haikusbot delete")

Reply

[-]

TechnicalGeologist99@reddit

I bet it's not under a permissive license

Reply

[-]

seamonn@reddit

It's Intern Season!

Reply

[-]

AustinSpartan@reddit

uh, not in the US.

Reply

[-]

txgsync@reddit

Intern season is just ending in the USA (usually runs late August through early December, depending upon the candidate). So yeah, unfortunately you're getting downvoted not because you didn't contribute to the conversation, but because your information was wrong. Alas. At least at Apple it was that way: we'd review interns starting around Jan-March, finish the last-minute approvals by May or June, and the candidates would have offers with start dates in August or September running usually for 12 weeks. So yeah. It's the very tail end of intern season. And someone may have been trying to commit their final project here. Plausible!

Reply

[-]

Repulsive-Memory-298@reddit

most of these are already open source?

Reply

[-]

mikael110@reddit

Nemotron Nano 3 30B-A3B and Nemotron Nano 3 30B-A3.5B are not public or announced as far as I can see. The latter one especially as it's explicitly marked as an internal checkpoint that is not for public release.

Reply

[-]

Repulsive-Memory-298@reddit

Did you nab either of those? The repo is down and I wanted to save the second one

Reply

[-]

mikael110@reddit

Sadly I did not, I tried but I only got half way through 30B-A3.5B before it got taken down.

Reply

[-]

Repulsive-Memory-298@reddit

appreciate it!!

Reply

[-]

TheArchivist314@reddit

Once more if anyone has a copy I'd love to have a copy please

Reply

[-]

TheArchivist314@reddit

Anyone got a copy I wanna archive it

Reply

[-]

HandfulofSharks@reddit

Did you get a copy?

Reply

[-]

TheArchivist314@reddit

Nope

Reply

[-]

These-Cost-905@reddit

7hhẹbd

Reply

[-]

thefool00@reddit

Grab it before full censoring has been implemented

Reply

[-]

Repulsive-Memory-298@reddit

fuck censoring has been implemented... Might you have a repo? XD

Reply

[-]

LosEagle@reddit

lmao the cloudflare junior who took down most of the internet got fired and was hired by nvidia

Reply

[-]

seamonn@reddit

unpaid intern*

Reply

[-]

raika11182@reddit

The Nemotron lineup is great stuff. Some of these projects look promising.

Reply

[-]

DrummerHead@reddit

[mistralai/mistral-**nemo**-instruct-2407](https://mistral.ai/news/mistral-nemo) was made in collaboration with Nvidia, so I assume this new model is derivative or inspired by it (**NEMO**tron)

Reply

[-]

mpasila@reddit

Nemo is just their framework thing and Nemotron models are models they've trained independently (idk if the quality is as good as Mistral Nemo on any of the other models they've released since then around the same size).

Reply

[-]

raika11182@reddit

I can't speak to the other sizes, but Nemotron 49B is an *excellent* model. All the power of Llama 3.3 70B with a smaller memory footprint, better formatted responses, and better prose. It's getting a little "old" now in LLM terms, but it's still one of my favorites.

Reply

[-]

mpasila@reddit

The 9B and 12B I believe are trained from scratch so very different from those pruned models.

Reply

[-]

SkyFeistyLlama8@reddit

Mistral Nemo 12B is still unbeatable if you want creative text. Not even Mistral has managed to outdo itself.

Reply

[-]

input_a_new_name@reddit

i don't feel like L3.3 70B is gonna get old for a long time. meta struck gold, maybe on accident, who knows, but at least until the big corpos pull their heads out of their asses, it will take a while before we something that's just so... neat, how else to put it? it's at the border where you can run it locally with an \*investment\* but without selling your kidneys, it's really smart, and not butchered by internal \*safety\* filters.

Reply

[-]

Phazex8@reddit

How about this... this was a planned leak to build hype.

Reply

[-]

CanYouEvenKnitBro@reddit

Its wild to me that hardware can be difficult enough to sell that companies are sometimes forced to make their own custom data centers, sell themselves as cloud services and in the worst case even make custom models to be actually use their hardware effectively.

Reply

[-]

Cool-Chemical-5629@reddit

I don't know about the leak, but to me this looks more like a mess than a leak. There are some directories named after older, already released models, but more importantly directories named after several different models that aren't even made by Nvidia - Qwen3-8B, Qwen3-14B, EuroLLM-9B. There's also directories that aren't even named in a way that would indicate a model - "nvidia" directory. That's only what's visible on the screenshot. Apparently there could be more. The actual "leak" is the 30B A3B model and there are only two visible directories related to that out of 21 total (visible on screenshot).

Reply

[-]

Odd-Ordinary-5922@reddit

showing something unintentionally is still a leak bro

Reply

[-]

TheStrongerSamson@reddit

I want that asap

Reply

[-]

emsiem22@reddit

So many leaking mistakes lately

Reply

[-]

T_UMP@reddit

https://preview.redd.it/6z1kinnm7u6g1.png?width=1197&format=png&auto=webp&s=5b9474b0ff50777fca0ff8575acea6f54e9e8a3f Lucky we have this shit...

Reply

[-]

mobinx-@reddit

U know there is no mistake only planning drama

Reply

[-]

_supert_@reddit

EuroLLM?

Reply

[-]

iamMess@reddit

Been out for 6ish months or more. Not very good.

Reply

[-]

GCoderDCoder@reddit

So I guess Qwen being in the list isn't necessarily a marketing opportunity for qwen if there's also a poorly received model in there too lol. I was going to say "ooo look they use qwen too" lol

Reply

[-]

gefahr@reddit

Probably just for benchmarking they ran.

Reply

[-]

Repulsive-Memory-298@reddit

Sounds pretty cool

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

[-]

PrettyDamnSus@reddit

All these disappeared "copies" are why torrents exist, folks...

Reply

[-]

rerri@reddit

Googled it and apparently there was already some info online about this: [https://x.com/benitoz/status/1995755478765252879](https://x.com/benitoz/status/1995755478765252879)

Reply

[-]

Amazing_Athlete_2265@reddit

Any reputable source?

Reply

[-]

MitsotakiShogun@reddit

Photo from Nvidia presentation looks legit.

Reply

[-]

griffinmisc@reddit

Yeah, it definitely looks like they might've leaked more than intended. It’s wild how these things happen, especially with all the hype around new models.

Reply

[-]

rerri@reddit

Apparently it was mentioned by Nvidia in late October: [https://developer.nvidia.com/blog/develop-specialized-ai-agents-with-new-nvidia-nemotron-vision-rag-and-guardrail-models/](https://developer.nvidia.com/blog/develop-specialized-ai-agents-with-new-nvidia-nemotron-vision-rag-and-guardrail-models/)

Reply

[-]

Long_comment_san@reddit

Holy shit, a 30b dense model, really? I thought these went extinct

Reply

[-]

mikael110@reddit

Where are you seeing a dense 30b model? Both of the 30B models listed in the screenshot are MoE with 3 / 3.5 active parameters respectively

Reply

[-]

ASYMT0TIC@reddit

What we really want is native 4-bit models with a single 32b expert and like 20 5b experts, yielding a model with the worldly knowledge of a 132b model and the thinking/reasoning ability better than a 32b dense model, but that fits into a 128 gb system with room for context. So, something like oss-120 but with a some added parameters to expand the size of the thinking expert. One can dream.

Reply

[-]

shiren271@reddit

Do we know if any of these models good at coding? I'd try them out but they've been taken down.

Reply

[-]

Terminator857@reddit

Unlikely, since past models haven't measured up.

Reply

[-]

JuicyLemonMango@reddit

So if those stats from that nvidia slide are real (the x link here in the comments) then this model is on par with Qwen 3 30B (3 active). But what i want to know most is context length and tokens per second. The mamba architecture should be faster so i'm guessing the tokens per second is better too. But context is bad on mamba, did they manage to get that much better? If they did then this might be a nice model! Slight reminder though, it seems to be **on par** with Qwen3, not beating it.Which on it's own is slightly disappointing to me as i expect more from team green.

Reply

[-]

mr_house7@reddit

Link for the files?

Reply

[-]

djtubig-malicex@reddit

shiet that was quick

Reply

[-]

AmphibianFriendly478@reddit

So, uh… mirror?

Reply

[-]

Frequent-District859@reddit

!remindme 3 days

Reply

[-]

TriodeTopologist@reddit

Since when does NVIDIA make models?

Reply

[-]

Inaeipathy@reddit

Upscaling?

Reply

[-]

RobotRobotWhatDoUSee@reddit

Chrck out their hybrid Mamba Nemotron-H series, which I believe is all from scratch. They've been training from scratch for a little while now. My vague impression is that they got familiar with training from all their extensive fine tunes etc, and then got into from-scratch training for large models (I wouldn't be surprised if they have been training small models all along.

Reply

[-]

Acidalekss@reddit

2024 with NVLM, mostly 2025,with hundreds of model on HF and Aerial software stack this December too

Reply

[-]

anomaly256@reddit

aaaaaand it's gone

Reply

[-]

nvmax@reddit

It just got removed at 7:38am cst...

Reply

[-]

ilintar@reddit

Damn, already deleted :(

Reply

[-]

egomarker@reddit

Drip marketing

Reply

[-]

ilintar@reddit

Interesting, so the Nemotron Nano 30B is a 50-50 hybrid model - every second layer is linear.

Reply

[-]

Niwa-kun@reddit

holy leak, this is gonna be an interesting few days.

Reply

[-]

JsThiago5@reddit

So is it a new nemotron based on qwen3 30bA3b? As nemotron is always based on some model like llama3

Reply

[-]

bbbar@reddit

At least there is an evidence that someone actually test the models against Qwen

Reply

[-]

OscarWasBold@reddit

!remindme 3 days

Reply

[-]

RemindMeBot@reddit

I will be messaging you in 3 days on [**2025-12-15 12:37:34 UTC**](http://www.wolframalpha.com/input/?i=2025-12-15%2012:37:34%20UTC%20To%20Local%20Time) to remind you of [**this link**](https://www.reddit.com/r/LocalLLaMA/comments/1pkpxss/someone_from_nvidia_made_a_big_mistake_and/ntmw5kr/?context=3) [**CLICK THIS LINK**](https://www.reddit.com/message/compose/?to=RemindMeBot&subject=Reminder&message=%5Bhttps%3A%2F%2Fwww.reddit.com%2Fr%2FLocalLLaMA%2Fcomments%2F1pkpxss%2Fsomeone_from_nvidia_made_a_big_mistake_and%2Fntmw5kr%2F%5D%0A%0ARemindMe%21%202025-12-15%2012%3A37%3A34%20UTC) to send a PM to also be reminded and to reduce spam. ^(Parent commenter can ) [^(delete this message to hide from others.)](https://www.reddit.com/message/compose/?to=RemindMeBot&subject=Delete%20Comment&message=Delete%21%201pkpxss) ***** |[^(Info)](https://www.reddit.com/r/RemindMeBot/comments/e1bko7/remindmebot_info_v21/)|[^(Custom)](https://www.reddit.com/message/compose/?to=RemindMeBot&subject=Reminder&message=%5BLink%20or%20message%20inside%20square%20brackets%5D%0A%0ARemindMe%21%20Time%20period%20here)|[^(Your Reminders)](https://www.reddit.com/message/compose/?to=RemindMeBot&subject=List%20Of%20Reminders&message=MyReminders%21)|[^(Feedback)](https://www.reddit.com/message/compose/?to=Watchful1&subject=RemindMeBot%20Feedback)| |-|-|-|-|

Reply to Post

165 Comments