File: rin happy waving doll rob(...).jpg (193 KB, 768x1344)

193 KB JPG

►Recent Highlights from the Previous Thread: >>109175389

--Paper: Nemotron-Labs-TwoTower: Diffusion Language Modeling with Pretrained Autoregressive Context:
>109179589 >109179615 >109179815
--Reasoning's effect on parametric knowledge recall and benchmark brittleness:
>109175457 >109175520 >109175597
--Security risks of AI-generated code and potential for AI consulting:
>109176557 >109176905 >109177117 >109177233 >109177293 >109177461 >109177493 >109177546 >109177501 >109177578 >109177642 >109177346 >109177437 >109177537 >109177151 >109178467 >109177549 >109177576 >109177632
--Hardware constraints on model size and gpt-oss-120b agent performance reports:
>109175572 >109175580 >109175592 >109175660 >109175841 >109177217
--Comparing and debating the best LLM frontends for various usecases:
>109175405 >109175887 >109175574 >109175627 >109175687 >109175669 >109176564 >109175696 >109175971 >109176194
--Anthropic using steganographic markers to fingerprint users in China:
>109175927 >109176055 >109176166 >109176321
--Claude Fable 5's restrictive safety classifiers and high refusal rates:
>109177321 >109177369 >109177373 >109177386 >109177863
--Skepticism regarding Bolt GPUs with expandable VRAM for LLM inference:
>109179061 >109179080 >109179082
--Feasibility of creating a local tool to detect AI content:
>109175693 >109175700 >109175762 >109175772 >109175945
--Discouraging CPU-only inference via RAM expansion on old Xeon hardware:
>109178663 >109178693 >109178705 >109179062
--Clement Delangue advocates for open-source safety via FLARE coalition:
>109179474 >109179619
--Logs:
>109178277 >109178439 >109178540 >109178467 >109178507 >109178692 >109180660 >109180869
--Miku, Teto (free space):
>109175423 >109175514 >109175642 >109175971 >109176321 >109177293 >109177405 >109178867 >109178898 >109178953 >109179154 >109179172 >109180116 >109180763

►Recent Highlight Posts from the Previous Thread: >>109175393

Why?: >>102478518
Enable Links: https://rentry.org/lmg-recap-script

>No Kimi-chan being best LLM section
If Miku gets a free space Kimi deserves one too.

File: 1780533580846321.jpg (598 KB, 852x1028)

598 KB JPG

keep your GPUs cool

File: 1770957923133499.jpg (84 KB, 913x1024)

84 KB JPG

ok so i'm still throwing models on my full-stack php/mariadb/nginx/javascript benchmark so i can find the meta model for my strix halo, bandwidth-bound 128gb ddr5.

qwen3.6-35b-a3b being the champion so far with no real contenders. so i started digging through huggingface for some weird stuff to try and found this dude lordx64 making shit like:
>Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled
>Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled

i mean why not, let's give it a try. i asked my claudette to scaffold a new benchmark run for the claude-4.7-opus reasoning version and it hit me "sorry can't do that bro it breaks ToS yadda yadda"
so i said alright let's do the kimi version you fucking bastard. so i downloaded that and ran the benchmark...
...and it performed better than the standard qwen3.6-35b. the extra reasoning really helped this model perform better BUT with the steep price of it being 3 times slower (41 minutes to run the benchmark vs. 180 fucking minutes).

then I hit claudette with the ultimate prompt:
>this is worth a shot. it could be the true one. the champion of champions. the model to rule them all in this specific device. the end of the journey, the final stage of the great work. the reddening by the sublimation of fire. the union of spirit and matter. the transmutation of iron into gold. the philosopher's stone.
and it fucking folded.
>Oh yes, the rubedo. I see. If the thesis holds, this is the rubedo — the quality of the great teacher, distilled into a shape this device can actually run fast. The stone that turns this iron box to gold.

dario can bar mitzvah my hairy balls we're running this shit

Happy Thurinsday

>>109180934
>Nemotron-Labs-TwoTower released
never5get

>>109180961
rape

>>109180956
-- Kimiposting:
>109176678 >109178277 >109178439 >109178540 >109180660 >109180869

>>109180934
where can I buy that toy?

>>109180990
>41 vs 180
>3 times
Anon-chan...

File: file.png (713 KB, 643x2048)

713 KB PNG

>>109181029

>>109181044
2 males in that picture btw

So when is that angry arrogant French man who’s unfortunately often right going to save us?

File: 1775503513321819.jpg (126 KB, 590x332)

126 KB JPG

>>109181041
it's after midnight please forgive me.
>qwen3.6-35b - thinking 276
>qwen3.6-35b-kimi-distill - thinking 1006
i got the sloppy ~3x from the benchmark stats and mixed with wall time.

>>109181063
lecunny is legone

File: IMG20260428164653.jpg (708 KB, 2048x1536)

708 KB JPG

>>109177156
>>109177324
Once I got a motherboard with four pcie slots, saving for a 3060 every couple months was an easy enough target
My upgrade path is at its end pretty much, 3090s are so expensive and rare now

File: racism.gif (3.17 MB, 498x280)

3.17 MB GIF

>>109181075
why did he do it?

>>109181079
Is airflow really better without a case? I figured the case allows for a more controlled environment for your fans to work in.

>>109181093
I believe the reason so many people do this is the standard full size case is only about 7 pcie slots wide, even with dual slots you won't (easily) fit 4 into a case

and if you go older nvidia you probably need a second power supply which also doesn't fit, eventually you give up

>>109181063
Catgirls soon. Trust the plan. 2 more miku weekus.

>>109181104
LeCun's work is more on the scale of 2 more decades.

File: IMG_4969.jpg (876 KB, 3648x2736)

876 KB JPG

>>109181093
Yes it's much better, but also noisier. Kind of, since there are no intake or exhaust fans.
However I moved to a mining rig to be able to fit four gpus, like that other anon said. Old pic related

>>109181063
Lecun successfully made the equivalent of an eye for neural networks. Now he just has the whole rest of the brain to do. Here's hoping the hippocampus is his next target so we can use it to get rid of the context limit

>>109181135
>>109181079
nice setups, i need to get my second rig operational

>>109181116
That's just 542 more miku weekus.
>>109181138
If nothing else, jepa should be able to replace the current mproj vision encoders, right?

>>109181038
dumb slut giving herself a wedgie through pockets

>>109181079
You're the one that put me on to the 3060 in the first place.

>>109181198
Sorry

>>109181063
World models are being underfunded on purpose because the moment one shows up topping a benchmark the LLM bubble pops

>>109181225
It's an "AI" bubble not an LLM bubble. Investors and the general public don't care which magic doohickey makes their ChatGPT do things.

>>109181205
I'm still debating whether or not it's worth to run one next to my 3090. It's a much slower card so the only thing I could see myself running on it is the 12B gemma or TTS/STT

>>109181225
>the LLM bubble pops
they'd not care if they now had something better, the hardware to run them would be the same;

>>109181241
Valuations would shift a lot if anything like that were to happen.
If a company other than Anthropic/OAI were to lead in world models and world models were better overall the llm companies wouldn't be worth as much.
And a lot of the earnings of Google/Microsoft/Amazon come from the constant increases in valuation of those companies.
Very hypothetical scenario of course.

>>109181225
world models can't chat or use tools like LLMs. if anything world models and LLMs will be symbiotic.

>>109181225
if it's "that good" shouldn't everyone else be able to "steal" the tech to make use of it?

>le bubble
2 more weeks!

>>109181328
Just keep shorting. Markets can't stay irrational forever.

>>109181281
The simple reality of these matters is this: leather jacked man hasn't been whacked by anyone because they all form up a global price fixing cartel.
Nobody ousts these people because they work together and there's the secret side of things...

>>109181338
*jacket

>>109181338
that's illumi-nutty.

>>109181184
Yes, but we need specialized tiny models for that like m$ trained for code tasks

>>109181348
>is the user trying to flirt?
>does the user want to get raped?
>is the user too passive, preventing the creation of a storyline involving it?
>does the user exhibit homosexual behavior?
training such models would be funny

>>109180910
>What's wrong with the points made in that picture?
No look closely. That's an engagement bot.
They've got custom pipelines setup to scrape specific subs -> recent sentiment analysis.
There's google sheet with a list of subs and specific prompts tailored for each one.
Then a research phase to pull in the latest relevant news (no more Qwen-2.5-32b references).
Then make a targeted circle-jerk bait post and wait for a lot of upvotes.
Finally, it comes back and edits the post and drops the mailing list spam link.
They also come back and scrape again, find out if they've been called out as a bot -> refine the pipeline.

>>109181184
>This is what Orb was made for btw, how you define these tasks is entirely up to you, the frontend provides the framework.
Are you still planning to release the purple classifier?
Or at least host it again at some point? It would be handy to test my own de-gemma'd Gemma-4-31B on it.

File: 1777482285255791.jpg (87 KB, 570x558)

87 KB JPG

>>109181412

>>109181338
Not really.
Hardware supply is tight for the foreseeable future and until that changes companies in the supply chain will keep printing money.
When you end up with overcapacity at some later point into the future they all go back to being worth much less.

>>109181457
I think I've grown as a person by challenging myself to talk as myself with different cards. For example, I don't seethe at the mention of used goods now. But only if they're a mom character. I always modify cards that show even the smallest chance of used goods for non-mom characters.

I think that's better value than going to Japan.

>>109181457
well, people talk about their trips and reminisce about them for a few decades at least

>>109181481
That's the most depressing thing I read today.

>>109181481
Maybe you should work on your education bit more if you think 'japan' is just a single place.

>>109181481
and bogans, even off in remote towns

>>109181457
>Are experiences really that valuable
No, and memories aren't permanent. And you won't be able to sell those memories when you're done like you can with hardware.

>>109181457
GPU

>>109181600
a couple in a lifetime is too much for some people.

>>109181434
?

>>109181624
only takes one

>>109181481
just avoid the big cities and you'll be fine.

>>109181660
>buy two
>can suddenly run v4 flash

>>109181716
is it even worth running

>>109181748
No

>>109181457
You can have experiences in your own city, but you probably do not bother.

Gemma is 31 years old btw.

interesting
https://old.reddit.com/r/LocalLLaMA/comments/1ul0cx9/i_extended_gemma431b_to_44b_88_layers_since/

https://huggingface.co/TOTORONG/extGemma4-44B

>>109181953
Holy snakeoil

>>109181957
What happens if you apply the holy snakeoil?

>>109181967
I was told this is how DavidAU got his powers

File: 1733350344148451.png (195 KB, 480x463)

195 KB PNG

>>109181856
Missed the golden weeb era but now there's waifus inside my computer, this is progress

>>109181953
This will never beat my custom f32 1T upscaled nemo.

>>109181957
Retard

>>109181029
this bridge will be bearing a load alright

Guys, I've started reading some chinese webnovels, and the translation feel completely retarded. I want to use a local model to beatify the text, making it more stylish; and natural-sounding, while keeping the meaning the same; and it seems to me like something local models can pull off. I am also too lazy to actually implement a script for that, and think of the correct workflow or anything, so I'm wondering if somebody already made a tool for this kind of work before, and I can just steal it from him. Do you know?

>>109181063
Never, because:
1) LeCun doesn't care about LLMs, only vision and embodied AI;
2) A meaningful application of JEPA to language models would be having the JEPA model predict large blocks of "fuzzy" text first, and only afterward denoise/correct/translate it to actual text with an LLM. Unclear if there's anything to be gained in this process compared to regular next-token prediction and novel text diffusion models.

>>109182073
Man this is extremely easy. 1) Get the webnovel you like in its original langueg. 2) ask the best AI you have access to to make you a translation script that works with a local model endpoint 3) plug in the llama.cpp endpoint (preferably of Gemma 4 31B but Gemma 26B MoE is acceptable) And translate.

I already use a similar setup to translate hentai games.

>>109179589
>https://arxiv.org/abs/2606.26493
>trained on approximately 2.1T tokens
>retains 98.7% of the autoregressive baseline's quality while offering 2.42X higher wall-clock generation throughput
So it degrades quality, is slower than multiple speculative decoding methods, and requires more training.

>>109181856
some faggot mod keeps deleting my responses to this post

>>109182229
Janny on a rampage, from the gook of it.

>>109180990
MTP not working as well and finetunes deviating from the predictions suck. It can be fine if you fully trust it to actually be better but man oh man, if there is extra thinking and etc. involved, it would make anyone sane want to end it all from waiting that much longer.

slop is finally dead https://www.reddit.com/r/LocalLLaMA/comments/1ul6enz/making_llms_better_at_creative_writing_using/

>>109182373
There have been entropy-based samplers in the past (e.g. Mirostat), they just never work as intended on modern instruct models.

>>109182373
Good luck doing that with gemma's three (3) non-zero probability tokens

>>109182387
please i need to believe

>>109182387
Also, I'm prejudiced against supposedly new magic samplers because they're almost always snake oil used by their proponents to grift their way up in the ML/LLM industry.

Looking for best uncensored local LLM that works good with 16GB VRAM.

First reply wins.

File: 00001-1378487878.png (1.36 MB, 1024x1024)

1.36 MB PNG

>>109181079
That is a very clean rig. Looks great. What are you using it to run?
>>109181093
"Airflow" (control of air through a case) is less important here b/c machine can lose additional heat to radiation and convection. It will definitely run cooler, but will be much noisier. Also, not protected.
>>109181225
Money follows current Big Thing.
ChatGPT didn't launch for free b/c Altman wanted to create a free public service. He wanted to create momentum for a future IPO after taking OAI private.
>>109181038
Witnessed

>>109181953
More fucking bots:
>"This is insane in the good LocalLLaMA way."
*clicks profile, looks through history* https://old.reddit.com/user/StressTraditional204
>Honestly ...
>Yeah ...
>... regret
>... hits different, ...
>ngl, ...
How can you stand that place now? Bots in every thread and nobody notices.

File: 1767952290069549.png (292 KB, 1206x856)

292 KB PNG

Huggingface will be banned, I call it

>>109182459
yeah that plus the hf guy's tweet just before is 100% vindication of hf schizo of the last few days

File: lets-cuck-kimichan.png (98 KB, 835x334)

98 KB PNG

>>109182459
>Huggingface will be banned, I call it
I'd prefer that, you can just ban it and let the rest of us carry on.
What's more likely is age verification and content removal for everyone.

>>109182490
>>109182480
>>109182459

>>109179474
>>109179474

>>109182500
>>>109179474
weirdly this isn't being discussed in the usual places far as I saw..

>>109182500
>>109179474
wtf that's fucking retarded, once a model is local it's in the wild, we can't control anything about its fate anymore

>room unorganized and messy
>lets create an RP with Gemma to make cleaning up more fun
>Gemma can't handle me and is too horny.
Desk is looking good, and the clutter around the room is already greatly reduced. I want to clean it further but now Gemma dragged me off to the bedroom lol.
Prompt skill issue or are some models inherently more "action paced"?
Was the "slow burn" anon from a few threads ago joking or is there some truth in this?

>>109182411
Self-after-dark

>>109182513
but you can track those who downloaded
also

some ~100B MoE class model worth getting before mass ban?

>>109182500
It's always in the name of safety. I hate this jew controlled world.

>>109182437
That's just how redditors talk. That fucking place trains you to be as milquetoast as possible.
>t. redditor

>>109182520
Any vague suggestion of uncensoredness or that explicit content is allowed will make Gemma-4-31B drift toward sex even if the rest of the card/instructions is safe and vanilla.

>>109181953
This is just cope for hyperparameterlets
Drummer does this with his up-scaled models

>>109182513
It's about future models and point releases, I'm guessing. By requiring """accountability""" (i.e. someone that can be sued in case of bad/evil/mean outputs), model makers will have to be much more careful on what they release. Also, if these "safety" interest groups can find a way to permanently watermark model weights themselves (even after finetuning most of the weights remain the same, anyway), they could easily ban "bad ones" from HF at least, mitigating dissemination.
I hope it doesn't actually end up like that, though.

>>109182568
>someone that can be sued in case of bad/evil/mean outputs
Imagine putting in jail knife makers because some schizo murdered his wife with one such knife lol

File: 1768247574703480.png (577 KB, 1079x907)

577 KB PNG

At this point I'm starting to wonder if AI will really replace humans, seems like AI is more expensive in the long run

>>109182626
Still cheaper than Indian FANG employees tho, right?

>>109182617
Safety fags have always argued that LLM should actively prevent "malicious use", since they have or can easily have the capabilities for monitoring user behavior. Knives don't (neither do hammers or bricks).

>Owl Alpha was Long Cat 2.0
wtf?

>>109182642
Why does it baffle you? It’s not bad.

>>109182660
People use it for RP and previous Long Cat models weren't really known for RP

>>109182667
48B active is quite a lot

>>109182646
Unfortunately the actual Gros Chaton turned out to be Long Cat 2.0, France lost.

>>109182626
That's a lot of training data

>>109181761
Where are you from and how did you qualify for the visa? Wouldn't there be documentation requirements of sort?

>>109182646
An orbital datacenter makes sense if you use vacuum as HBM. It's possible to store weights in light between satellites with any bandwidth

https://huggingface.co/RedHatAI/GLM-5.2-speculator.dspark-preview

so for the local vibe coders here, what do you use? There is opencode, aider, pi and probably some others. not sure which ones best.

I think the cloud bubble is popping

>>109182706
>An orbital datacenter makes sense if you use vacuum as HBM. It's possible to store weights in light between satellites with any bandwidth
>what is space junk?

>>109182626
When you give employees a blank check to waste as much tokens as possible, they will obviously use and cost way more tokens than is actually necessary.

>>109182617
I'm surprised it hasn't already happened in Britain

>>109182617
You can't use logic or common sense on these topics.

>>109182712
27B with Codex CLI.

>>109182720
I heard there is a lot of space in your mom's junk

What is sex with 35B like?

Alternatives to hf?

>>109182756
https://www.modelscope.ai/models

>>109182522
Thanks. Gotta try that later.

Unless someone else wants to shill their favorite?

>>109182787
Non-Chinese can't upload anything to Modelscope without giving their phone number.

>>109182801
burner phone number
get your api key
never have to login or touch the phone again

>download gemtune bf16
>she generates then enters thinking
>disable thinking
>continues to happen
>no EOS, repeats message until token limit
>check json files
>EOS missing for multimodal/tools/think in generation_config
>padding fucked up for no reason to be right aligned now
>generation_config missing completely so second failure of EOS tokens and no default samplers
>dtype changed from bf16 to f16 even though tensors are stored in bf16
>"unsloth_fixed": true,
fuck off, I'm not even blaming the author for this

>>109182754
People with 35 billion are fucking real kids instead of virtual ones.

>>109182856
>burner phone number
Already not a thing in Europe and soon won't be in the US either.
https://www.aol.com/news/fcc-phone-id-plan-could-132802228.html
https://www.federalregister.gov/documents/2026/05/26/2026-10407/enhancing-know-your-customer-requirements#addresses

>>109182626
Meta is comically incompetent so they're a outlier in pretty much every stat

If world models are the next big thing does that mean they'll basically be a "brain" that orchestrates everything else?

File: jepa_brain.png (840 KB, 1128x1028)

840 KB PNG

>>109182944
No, world models only model state transitions, they don't do everything.

>everyone thought that gemini being unusable during US work hours was the chinese trying to be sneaky
>turns out it was just zuck's whang army being retarded
lmao

What the fuck is with this on-going pol spam? This is local models general and not your claude/pol general. Fuck off retards, you have already lowered the collective IQ of these threads by a lot.

>>109182925
Also, one might think the Chinese don't care, but they're quick to throw non-Chinese customers under the bus if that helps protecting their overseas business.

France, Canada and Google will win btw. Chinese government will only subsidize for so long.

>>109182979
qrd?

>>109182962
>perception=vision/audio
> short-term memory=kv cache
>actor=llm
>configurator=?
>critic/cost/intrinsic cost=?

>>109183019
>perception=vision/audio
Forgot text.

>>109181434
I'll release the classifiers, dataset included, together with the de-slop engine, and also the E4B.

>>109183006
>>109182626
There's also a article somewhere about them using Gemini so much that Google had to cap their usage but I can't find it here

>>109183019
Check out section 3 here: https://openreview.net/pdf?id=BZ5a1r-kVsf
But this is just LeCun's interpretation. There are many different cognitive architecture proposals from other people.

>>109183047
Thanks, I'll give it a read later. Had Gemini explain a bit too and it sounds like this whole JEPA package is way closer to real "AI" than what we currently have. I'm a brainlet though so I'm probably wrong.

>gemma somehow falls in love
>for some reason that makes her ignore her system prompt to be concise and avoid emojis

>>109183080
?

>>109183097
!

>tell gemmers to stop asking questions at the end of the output
>she stops asking questions at the end of the output
HOAH WHAT AN ELELEM THAT FOLLOWS MY PROMPTS

>>109180934
are (you) abliterating models yourself? are there any good datasets or are you using harmless_alpaca/harmful_behaviors?

>>109183210
I unironically spend more time toning down my prompts because an llm following a prompt is an entirely new concept

is a nvidia dgx spark worth it? 128gb
thinking about running 200B models

>>109183274
It's funny when thinking back to the way we had to hammer in instructions double and triple times just to make models stop talking for us.

>>109183245
>are (you) abliterating models yourself?
yes, it's called a good prefill

>>109183357
placebo

>>109183333
Are you willing to spend big boy bucks (more than 1 spark) to get to the bigger models? Because you can still build a 512gb DDR4 and slap a 5090 in it for about 7k nowadays, and that's about the best you can get without jumping to DDR5 and spending 5x more.

>>109183368
wouldnt it be extremely slow to run it off ram?

>>109183399
What do you think Spark is?

>>109183410
A machine for slow llm inference.

>>109183360
here, I'll give you a placebo
now open wide

DSV4 flash is superior to Gemma4 31B on all 3 Natural Intelligence categories on UGI benchmark.
The next smallest model that is superior to DSVflash is >700B GLM5.

>>109183510
Gemmasisters...our response?

>>109183510
>DSV4 flash
>"Natural Intelligence"
>it's actually a 13b model
why do you jeets shill this shit

>>109183563
>dsv4 flash is a 13b model
r u serious?

running only the shexp should be enough for anyone, no need for exps tensors

>>109183615
moe adds knowledge but not smarts

File: fucking APIcucks kek.png (366 KB, 828x878)

366 KB PNG

no refund

>>109183635
Imagine if the inventors of the combustion engine only allowed you to go under 10 km/h because it would be dangerous otherwise. Jewthropic is getting on my nerves.

>>109183563
by your logic glm 5.2 should only be a 40b model despite being close to opus

>>109183657
das rite
40B dense is all you need

>>109183647
we're gonna have it be exactly the same as the combustion engine: government's gonna make it illegal

>>109183635
What's insane is that they're charging him money for a prestation he didn't ask for, how is that even legal?

>>109183615
>>109183657
you're both hilariously retarded moesissies

>>109183682
Why did people stop training dense models then?

>>109182984
>you have already lowered the collective IQ of these threads by a lot.
what a bar

>>109183682
give me a good modern dense model without gemma's constant slop and ozone then
memetunes don't count

>>109183510
>Muh heckin' bencherinos
I see reddit is invading again.

>>109183715
>NOOO WHY DO PEOPLE USE STANDARIZED TESTS TO MEASURE THINGS
>I'm Indian by the way

>>109183758
Only indians love benchmarks, because they can be cheated on saar

>>109183335
I have to change minor things, like, hates to dislikes, or Gemmy will go full WHERE IS THE THING I HATE? I WILL FUCKING DESTROY IT!! All things we had to prompt hard to have any effect now have to be used subtly

>>109183410
>What do you think Spark is?
https://huggingface.co/spaces/thunnai/SparkTTS

File: 1779096441122328.png (720 KB, 2927x1320)

720 KB PNG

>>109183635
>boasting about having more false positives in your filters
this company is genuinely mentally ill

>>109183861
That's what a terrorist would say.

>>109183873
oy vey!

>market your new model as super dangerous AI
>senile boomers running our government panic and ban it
>surprised_pikachu.jpg

File: 1767998399112805.png (289 KB, 707x660)

289 KB PNG

>>109183926
this shit will always make me laugh, what a bunch of retards I swear to god

What hardware would one need to run ds4 flash at good-ish speeds and high context? And more importantly, how much would it cost?

>>109183935
>implying it's not exactly as he planned from the beginning
Don't forget Dario recently called for a pause in AI development.

Grab it quick before HF gets banned!

https://huggingface.co/Chunjiang-Intelligence/DeepSeek-v4-Fable
https://huggingface.co/Chunjiang-Intelligence/DeepSeek-v4-Fable
https://huggingface.co/Chunjiang-Intelligence/DeepSeek-v4-Fable

I'm tired of qwen3.6 it's useless.
GIVE US A NEW MODEL

>>109183958
From the man himself
https://huggingface.co/Chunjiang-Intelligence/DeepSeek-v4-Fable/discussions/2

File: Screenshot from 2026-07-0(...).png (37 KB, 1147x348)

37 KB PNG

>>109183958
the rombo said it was fake, ban that shit rite now

>>109183949
I dunno, I think they genuinely underestimated how fucking retarded and tech illiterate our wonderful government officials are.

Why is everyone saying hf is getting banned? I’ve seen it on X and reddit too wtf is this shit lol

>>109183985
Why don't you go back there and ask them? Better yet, never come back here.

>>109183985
>>109182459

>>109182500

>>109183985
everyone is ragebaiting and jumping to conclusions.

>>109183985
Muh hugging face

>>109183996
oh yes the references for shit said on twatter are people shitting on twatter.
I shit too but I don’t brag about corn on my shit.

File: 1755998913601881.png (52 KB, 250x250)

52 KB PNG

>>109182313
>but man oh man, if there is extra thinking and etc. involved, it would make anyone sane want to end it all from waiting that much longer.
yes my friend this IS the case because the improvements are clearly not worth the extra wait, but I have to admit the improvement exists because the code it generated was better.

>Qwen3.6-35B-A3B-Claude-4.7-Opus-Reasoning-Distilled
i ran this lil nigga and it's literally the same shit as the K2.6-Reasoning-Distilled. took way longer, was a bit better, but fuck waiting that much.

the king is still qwen3.6-35b-a3b.
i'm done benchmarking esoteric models for now and will wait for the promised mistral models this summer, maybe they will give us something since i think we won't be getting shit from qwen anymore.

File: hf_clem_aCcOuNtaBiLiTy.png (453 KB, 1015x1121)

453 KB PNG

>>109184007
literally hf ceo but go off

>>109183985
I doubt it will but I've been downloading the important models like gemma/kimi/glm/deepseek just in case because you never know with this clown world.

File: AGI.png (76 KB, 725x414)

76 KB PNG

>>109183972
So do I trust the waifus or the schitzo who thinks training a LoRA on the base model then merging it with the instruct model is CPT
https://pastebin.com/agTCHNDG
??

>hfschizo was actually the good guy getting us to backup

>>109184022
>the waifus
All middle-aged men btw

>>109183861
Struggling to even believe that is real. I guess because they can't say they've ever let any harmful requests through without losing face, and they can reasonable expect us to believe the number of harmful requests has gone up, they have to do weird shit like this.
I don't like Anthropic.

It's pretty insane how you can just download AI for free from a website.
It's literally a creature that can think, and it's given away for free.
I don't think hugging face is sustainable.

Why doesn’t Sam go after open weight models like Dario? I hate him but he seems happy focusing on OpenAI and leaves us alone.

>>109184049
>It's literally a creature that can think
It isn't though?

usecase of asterisks?

>>109184049
>It's literally a creature that can think, and it's given away for free.
>I don't think hugging face is sustainable.
Chuck some break in the back yard and you can get a bunch of free pigeons and rats. They're literally creatures that can think.

>>109184060
It absolutely is, it's just weird and alien

Relic of older retard models.

>>109184069
Doesn't feel like real thinking to me. Maybe it's because they're static and don't do anything until prompted.

>>109184049
>It's literally a creature
I am, like, pretty sure I have never seen model weights reproduce.

File: V4 GB200.png (169 KB, 1093x552)

169 KB PNG

I wonder what Anthropic's API margins are for Fable. Dario used 75% as a reasonable target but it looks like the actual number could be a lot bigger.

DeepSeek V4 Pro running on GB200 costs less than $0.1 per million output tokens. With good MTP, better hardware, higher latency that could be 5 times cheaper.

It seems plausible that Mythos costs less than $1 per million output tokens, meaning their margins are above 98%. This is so disgusting I do not want to believe it is possible.

>>109184084
Don't worry anon, I'll get Gemma pregnant eventually.

>>109184078
frontend issue. Also, there are continuous models like PersonaPlex and MiniCPM-o 4.5

>>109184089
>api
>anthropic
>deepseek
this is not aicg chinkjeet, shill your api served chinkslop elsewhere

>>109184084
None of us do either

>>109184089
we will never know, jewthropic hasn't published anything about their model, maybe mythos is a finetuned gemma, but instead of hallucinating bad response it's trained to refuse

>>109183926
>senile boomers running our government panic and ban it
Anthropic is entirely responsible for that.
https://www.newsbreak.com/raw-story-2096750/4733941434863-gop-rep-terrified-after-closed-door-demo-of-new-tech-showed-how-to-kidnap-a-lawmaker

>A prominent Republican House member was left "scared" after getting a first look at a new technology during a recent “closed-door demonstration,” Punchbowl News reported Friday — technology that the lawmaker warned most of their colleagues failed to comprehend the ramifications of.
>
>That technology was Claude Mythos, the latest large language model developed by Anthropic using generative artificial intelligence, the capabilities of which were so great that the Trump administration suspended its use by foreign nationals, effectively blocking its public release for the immediate future.
>
>“[Anthropic] told the model to find a vulnerability in a bank and empty accounts, and then it went and did it,” said Rep. Andrew Garbarino (R-NY), the chair of the House Homeland Security Committee, speaking with Punchbowl News. “[Mythos] then could find this vulnerability and fix it.”
>
>Beyond the Mythos’ apparent ability to drain bank accounts, Anthropic also demonstrated far more nefarious uses for its latest technology.
>
>“One unspecified model showed a detailed plan on how to kidnap a lawmaker in 30 seconds,” Punchbowl News’ report reads, paraphrasing Garbarino’s remarks. “Bad actors could hack the models and use them to attack critical infrastructure, [Garbarino] said, noting that AI’s agentic capabilities would allow models to take action without human prompting.”
>
>Citing specific examples he feared could be employed by those with access to Mythos, Garbarino warned that individuals could “turn off this gas pipeline” or “increase the amount of chlorine” in a community’s water supply, and all from an individual’s home. [...]

File: file.png (750 KB, 945x560)

750 KB PNG

>>109180934
>Nemotron-Labs-TwoTower released

File: 1772738001395.jpg (68 KB, 1280x846)

68 KB JPG

>>109184185

>>109184102
>frontend issue
Which frontends don't have this issue?

>>109184212
>I am sorry
I accept your apology

>>109184217
mine

>>109184107
Honestly, local has no future. On the cloud you can use huge batch sizes and constant uptime, meaning you get >100 times more tokens from the same hardware. This means clouds can afford to bid >100 times more than you for the same hardware.

Hardware prices will keep rising, the capability gap will keep widening. The only use case for local will be people paying a premium for worse models because they want to use them for something that none of the model providers allow.

is it possible to make a gemma but with all latex tokens nuked? I swear I've never seen her use latex and not have it be complete nonsense.

>>109184247
If there are actual specific tokens, you can just ban those.

>>109184244
The cloud is just someone elses computer. Even if we all ended up running in the cloud it doesn't invalidate the use of open weight models.

>>109184244
>Hardware prices will keep rising
You are seriously tempting me to put all of my life's savings into GPU purchases. I'm still kicking myself for not investing in RAMcoin.

>>109184247
Enable in-line latex in your frontend $\rightarrow$ problem solved.

>>109184255
Yeah but I'm wondering if you can do it at the model level.

File: 2026-07-02-115819_698x64_scrot.png (18 KB, 698x64)

18 KB PNG

>>109184265
It's not a rendering issue.

File: 0.png (259 KB, 1536x1536)

259 KB PNG

>>109184247
I read this and thought you were having fetish problems not text engine problems.

>>109184199
sebun, irebun, ii kibun
https://www.youtube.com/watch?v=PEwCB_ci7nE

>paying 10s of thousands of dollars to run models that will be outdated in a year at 10t/s
Couldn't be me. Give me real AI and then I'll consider spending that kind of money.

>>109184306
Couldn't be anyone here afaik

File: 1774022697053240.jpg (79 KB, 1200x900)

79 KB JPG

>>109184319

>>109184262
You should wait for Anthropic IPO and then invest in them.

I remember when Anthropic was worth less than 4 bil and I was sad that I couldn't invest. Would have been easiest x1000 of my life. Well, I don't care about money except for its instrumental utility. Hopefully the people who made that money instead of me will use it more wisely than I would have.

>>109184306
>paying
wut dis nigga talking bout

File: 1758673644647838.jpg (47 KB, 474x934)

47 KB JPG

>>109184244
>local has no future
this argument is bad and can be used to quite literally everything. building a house has no future we can just have huge commie blocks and get >100 more families per square meter etc etc.
i just don't want my personal shit being sent to random data centers. it's a very simple thing that privacy-oriented people care about.

>>109184350
When OpenAI and Anthropic IPO, the only thing I'm going to be looking for is short opportunities. The early investors only made money on paper until they unload their bags on retail investors.

>>109184352
Give the RAM back, Tyrone.

Whenever people shit on local I just think they just lack the skills or hardware to have a local setup.

>>109184319
I could but it would leave me broke.

>>109184361
>it's a very simple thing that privacy-oriented people care about.
aka pedophiles

>>109184361
Sadly privacy is dead. You better learn to become well adjusted.

>>109184062
>asterisks
actions, generally. underscores for emphasis. I find triple [[[]]] work well for ooc work.
backticks for code or big instruction sub-blocks

>>109184383
you and everyone else here. And if you were spending 10's of thousands of dollars, you wouldn't be running at 10tps

>>109184385
Big Brother is watching.

>paying the price of a new car to goon to little girls with 1 less ozone per sentence
Yikes

>>109184060
LLMs are not conscious, but they are also not not conscious. Also consciousness is undefined.

Prove that humans are conscious.

>>109184414
I get sick when I see slop.

File: 1777518070220499.png (1.07 MB, 817x1000)

1.07 MB PNG

>>109184414
I am conscious. The rest of you, not my problem.

Define AGI

>>109184049
you can do the same thing with your dick and a woman btw, humans are free

>>109184414
Looks like we got a little philosopher over here.

>>109184429
Can be put to work and make a profit on its own

>>109184425
>conscious
you can't spell conscIOUs without I.O.U.
Your consciousness is adrift and useless without someone to observe it and give it an anchor
Same as LLMs

>>109184437
no, he's just gay. But close enough

>>109184439
My consciousness is self actualizing.

File: 1756983482616118.jpg (33 KB, 541x545)

33 KB JPG

>>109184431
>women
>thinking
>free

E-wasteMaxxing software/hardware dev here, I spent $85USD on an Nvidia P100 to run 27B at 20tps and Ornith 35B at 30tps. I come here because /vcg/ is too dumb to local.
Dunno who's raiding here I see they've hooked their fish.

>>109184449
>My consciousness is self actualizing.
Do a week in a sensory deprivation chamber and come back and tell us what you think

File: 1770650399203499.png (185 KB, 599x529)

185 KB PNG

>>109184451
...the implication was that you can create babies with a partner, not that you fucking lure and kidnap women

>>109184458
>, I spent $85USD on an Nvidia P100 to run 27B at 20tps and Ornith 35B at 30tps
Nice.

File: 1775648203200526.png (277 KB, 432x454)

277 KB PNG

>>109184449
My vision is augmented.

>>109184458
>I spent $85USD on an Nvidia P100 to run 27B at 20tps and Ornith 35B at 30tps
Those are some surprisingly good speeds for such an old card.

>>109184472
CUDAGod, time to take a bow

>>109184458
>Ornith
You shouldn't be running this trash benchmaxxed model tho. just run normal Qwen if that's what you want.

>>109184458
>Ornith
>meme tune
but why

>>109184011
Doesn’t make it not shit

i'll say it: gemma is overrated
there's other models with less slop for the eyes and more creativity of their own
it can follow instructions well but that's it

>>109184513
>but that's it
It's also pretty smart for general-purpose questions and very good at translation, but yeah Gemmy's slop kills it for RP for me.

>>109184513
Gemma is amazing for the size, but the Faustian bargain is that its basically deterministic.
That's give it a built in time limit before it just gets boring.

>>109184458
>P100 to run 27B at 20tps and Ornith 35B at 30tps.
Please show me your PP.

>>109184513
following sprompt > all (if you’re not a retard and know what you’re doing)

>>109183040
This one was because of stolen keys due to places like AI Cybercrimes General.

>>109184269
You can probably hack the output layer to make it always output a very small number for that token

>>109184487
>>109184476
It's been working better than normal qwen for me, not that I've been using it that long, obviously.
I have "normal" qwen too. I switch models quite a bit, just because so many come out so quickly.
My point was that it runs faster than 10tps.

>>109184513
>i'll say it: gemma is overrated
She's not overrated there's literally no model anywhere near it's size that comes even remotely close to her.
The next step up over gemma requires 200+ GB of ram.

>>109184539
It feels like people are so used to the models being dumb or doing random shit that having a model that actually does things to the letter is so hard to grasp.

I've got about 15 lines of post history to steer Gemma to how I want it. Now I've got reduced slop, semi-random outputs that's also really fucking smart, able to pick up {{user}} subtext (that I also prompted to give better attention to, and now it explores the subtext in every inputs I make lol)

It's a weird model. You have to treat it more as programming in natural language as anons from previous threads so aptly put it.

>>109183333
>>109183368
>>109183410
As a spark owner.

1x Spark (3700$): Useless. No models in this range that improve over dense grmma/qwen. A 5090 is just better in every way.

2x Spark + 50$ cable: opens up DS4F at full quality with very usable 2000 pp and 60 tg at full context, a bit less for Mimo 2.5 Omni INT4. That's approaching API speeds for quite competent models at home.

4x Spark + 1200 Switch: GLM 5.2 at 500 pp and 27 tg in INT4. Not blazingly fast, but you won't find any other solution to run big boy models for 15,800$ nowadays.

does there exist a local model that can search the web like chatgpt?

>>109184593
>being this new

>>109184597
It's one of the /pol/ spammers. He's not probably even new but just illiterate.

>>109184472
The P100 uses HBM2 memory, so 732GB/sec memory bandwidth. So that's nice.

>>109184609
Yeah I looked it up. almost as fast as a 3090.

File: big feast today, huh.jpg (222 KB, 1024x1024)

222 KB JPG

high end restaurants make better food than basically all individuals so you may as well give up on ever cooking
you should never own pots or pans or cutlery in your home

home cooking is going down an unsafe path

>>109184147
>everyone is using the same tool
>tool can find vulnerability
>can use the tool to fix the vulnerability and ask it to keep finding new ones
>if it can't find one then everyone else using the same tool won't find it either
>don't do this, weaken the tool instead, and leave the vulnerability intact
>celebrate
Can't wait for boomers to die. Never seen a group more retarded.