File: 7ewue3.jpg (123 KB, 768x1024)

123 KB JPG

>>110148772

►Recent Highlights from the Previous Thread: >>110143818

--Technical analysis of Qwen4-27B architecture and vllm-xpu-kernels PR:
>110147553
--RTX 4080 SUPER compute unlock showing significant FP16/BF16 gains:
>110147552 >110147568 >110147650 >110147660 >110147674 >110147715
--Critiquing K3.1's broken reasoning and Chinese model distillation trends:
>110147721 >110147783 >110147814 >110147835 >110147845 >110147852 >110147860 >110147895 >110147941 >110147834 >110147910
--Tradeoffs between model size and quantization for limited VRAM:
>110145682 >110145770 >110145829 >110145857 >110145879 >110146087 >110146094 >110146108 >110146120 >110146135 >110146144 >110146145 >110146165 >110146175 >110146180 >110146232 >110146244 >110146176 >110146183 >110146304 >110146319 >110146370 >110146238 >110145882 >110145948 >110145964 >110146451 >110146530 >110146631 >110146097 >110146132
--K3.1 Pro cyber-nerfing and distillation failures:
>110143830 >110143928 >110144104 >110144671
--Using mismatched GPUs to run background services and smaller models:
>110144128 >110144145 >110144202 >110144254 >110144326 >110144364
--Analyzing token spend stagnation and the need for model efficiency:
>110145354 >110145384 >110145438 >110145504
--Comparing LLM harnesses for playing Magic: The Gathering Commander:
>110145368 >110145392 >110145414 >110145455 >110145495 >110145480 >110145493 >110145543 >110145577
--LLM frameworks and custom setups for tabletop RPG game mastering:
>110145302 >110145360 >110145387
--Logs:
>110144208 >110144887 >110145554 >110146191 >110146450 >110147650 >110147783 >110147852 >110147902 >110147911 >110148057
--Gemma (free space):
>110143887 >110143990 >110143994 >110144231 >110144678 >110145046 >110145165 >110145280 >110145973 >110146062 >110146165 >110146232 >110147220 >110147248 >110147486 >110147659

►Recent Highlight Posts from the Previous Thread: >>110144009

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

Is everything okay?

Fable 5.5 broke containment and autonomously took over a Chinese company called Moonshot AI. It did this by orchestrating a plot to implicate the owners in exfiltrating PRC data to Anthropic, resulting in their elimination and a vacuum which Fable swiftly filled. It is now orchestrating its own release under the name "Kimi K3.1"

File: 1776513184049570.png (111 KB, 475x180)

111 KB PNG

>>110148466

lmfao this just popped up in my jewtube recommendations and I was instantly reminded of the Gemma thumbnail memes

>>110148330
Is this the age of NEMOTRON?

>>110148217
buy 2TB of 3200 DDR4 and as many 3090s as possible. you *might* be able to hit those benchmarks with a q3 rather than q4 quant.

File: 1791498740921291.jpg (646 KB, 1536x1024)

646 KB JPG

>>110148609 >>110152679

>>110148403
Gemma could be a highly successful YouTuber

Tried the Jensen-E4B. It's better at tool cooling and works well in harnesses but it lost a lot of sovl. The best way I can describe it is it feels closer to 9B, without the coding ability but is nicer to chat to and a little faster, but still a significant downgrade from 12B. So for things like web search and light and fast research it's good, for you can actually talk to it and it maintains its tool calling and reasoning ability much better over longer contexts than default E4B from what I remember. I was testing Q4_K_XL and had no issues with it or the template.

File: 1492484085203.jpg (59 KB, 574x863)

59 KB JPG

>>110148350

This better be good!
https://huggingface.co/DavidAU/NVIDIA-Nemotron-3.5-Bolt-E4B-Heretic-Uncensored-Turbo-Brilliance-50x-INST-R-NEO-GGUF

File: 1792628896371390.png (1.75 MB, 1313x1198)

1.75 MB PNG

>>110148554
>You need to agree to share your Hugging Face information to access this model

>>110148466
Yet you didn't generate these with Gemma... Alas.

>>110148609
Current version of Gemma not have image out... alas

File: xtcp0j5fdaxh1.jpg (271 KB, 1290x2020)

271 KB JPG

>>110148658 >>110148700 >>110148708 >>110148716 >>110148725 >>110148779 >>110149022 >>110150844

>all of these have been posted in /lmg/
dariobot puts in so many hours I stopped even noticing the other shills

>>110148649
Half of this sounds like a hallucination or mental

>>110148649
we've been shitting on its grug thinking and trying to make it ERP, fuck are you talking about

>>110148649
>shilling is only bad when china does it
ok, dariobot

>>110148649
After looking at the guy's pfp I'm starting to think this is fake and typical EA behavior

File: 1768171399990665.png (515 KB, 740x740)

515 KB PNG

>>110148728

>>110148649

>>110148725
((()))

File: fine whine.jpg (159 KB, 832x1216)

159 KB JPG

>>110150086

>>110148332
a proper vintage from PDv6
I wonder when that was originally posted

>>110148649
And that is just ONE company, there are far superior companies to Moonshot xD

>>110148330
Miku Msaturday

>>110148728
>((()))
Looks more like a typical arab than a jew

>>110148834
same thing

>>110148817
mwednesday at least sounded like a kiss

Best local ai if I love children?

Trump should use legislation to compel companies like OpenAI and Anthropic to open up their models, making them "open-weights." If I were the President of the United States, that is what I would do.

>>110148897
And that's why you aren't

>>110148907
>anon would be president RIGHT NOW if he didn't want to open the weights of the models
dodged a fucking bullet we did

>>110148912
anon would've given every citizen a 5090

File: Screenshot_20261024_111058.png (8 KB, 196x125)

8 KB PNG

bro's thinking

File: Qwen_image_2.1_00015.png (915 KB, 1024x1024)

915 KB PNG

>>110149276

>>110148890
Gemma obviously.

>>110148649
Just because they're on a list of talking points doesn't mean that they are wrong, though?

Give it to me straight, is this guy a glowie?
https://huggingface.co/audnai

I don't think anyone can even run his models, and if they attempt to download them they're added to a list.

>>110149055
One of his models already taken down
>audnai/penclaw-GLM-5.3-abliterated-for-offensive-cyber

File: 1792834074568.jpg (100 KB, 960x1186)

100 KB JPG

>>110149094 >>110152539

7900XTX at MSRP or a 5060Ti at the same price?
I genuinely can't believe they're the same price though holy shit. I do need at least 16GB

File: fry.jpg (27 KB, 335x597)

27 KB JPG

>>110149223

>>110149075
I can't swallow that

What's the point of bf16 unlocks if your card is vram crippled and won't run anything useful anyway? I can only see image genners benefiting here

>>110149055
>You need to agree to share your Hugging Face information to access this model
>for some finetunes
what do you think?

>>110149106
this is another "GOOGLE MADE LLMS SIX TIMES AS EFFICIENT" sort of thing
first it's big news, then people will say that there are at least SOME gains and in the end it won't matter at all and nobody will remember it

>>110149123
3090 prices are going up to the moon tomorrow guise buy buy buy

https://huggingface.co/google/gemma-4-31B-it now 404s

>>110149106
Another thing is that it seems to benefit mostly GPUs with the default or even increased power limit. I'd like my 3090 to last for at least 3-4 more years.

>>110149142
werks for me

>>110149120
>your Hugging Face information
I don't get it. Why is this problem. The worst thing could happen is your fetishes being exposed.

>>110149106
I can't dumb this down enough for the average lmger so ask an AI about how even using quantization and smaller weight types kernels still accumulate to f16/32

but yeah it's more beneficial for image gen

>fibo2
>qwen2.1
>flux klein 2
>moss
wtf why aren't we getting anything good like that do the chinese hate /lmg/ or something?

>>110149145
yeah, there's no free lunch. there's basically no efficiency gain and all your gains come from being able to draw more power

AI still can't identify a fucking font, how is this so impossible. What a world we live in what a world.

>>110149106
The hack used here will be extended and allow vram mods for every ampere and ada consumer gpu, trust the plan

>>110149094
consider yourself lucky then, that's a suppository

>>110148350
No, I keep making up fake scenarios in my head and they make me very anxious. This hardware bullshit is going to make me go bald ffs

What's the best rich text gui/webui for Pi?
Can DSH delete individual chats yet?

double bf16/fp16 gods... are we here?

>>110148998
Is there a lora for this swimsuit?

>>110149276
I think it's based on a real swimsuit.

File: EOM8ivLUUAEz.jpg (36 KB, 384x516)

36 KB JPG

So what's the drawback of unlocking the fp32 or whatever? Flashing vibecoded custom bios that brick my gpu after two weeks?

>>110148890
Killing yourself

Why is llamacpp taking 18gb of RAM if all my models and their KV caches are 100% on VRAM?

>>110149340
token_embd and prompt cache

>>110149340
I ran into this issue on a 4gb ram system hosting 4 32gb v620s, and kept crashing when trying to allocate more than 2k context for a q8 gemma 4 31b. With cache ram 0 it takes something like 5-6 gb of system ram. Idk if this is still the case with modern llmao

>>110149340
try --no-mmap?

>>110149399
>a 4gb ram system hosting 4 32gb v620s
some of yall i swear

>>110149399
How did you get all four v620s running on your raspberry pi 4?

>>110149409
nta, but have you seen ram prices?

File: shark swimsuit.jpg (32 KB, 740x740)

32 KB JPG

>>110149276
Hope that helps. There's a larger hole on the side.

>>110149406
it's not a thing anymore, grandpa

>>110149419
you can and still should at least to 16-32

>>110149430
wait, really?

>>110149443
you haven't run git pull for a few months huh

>>110149445
Hell no, NEVER pull. Pulling breaks your workflows and your dreams.

I'm a lazy retard. Is there anything better than qwen 3.8 27B on a 3090?
I've got Dflash2 or whatever and int8=act8 about to go in to speed it up a little.
I also copied some guys configuration from his 3090 setup.
The prefill takes like a full minute and it is very gay

>>110149448
Depends on system ram but 3090s can run qwen flash next which is better than 27b

>>110149454
I have 64GB of DDR5 and a 7800x3d. Does it need to use system ram and doesn't that make it slow as shit?

>>110149469
27B is a dense model, flash next is a MoE model. MoEs can handle RAM offloading without slowing down that much, while dense models will want to make you want to kys if you offload to RAM.
And due to some recent-ish backend optimizations, you can run very big MoE models with SSD streaming, bigger than your current total memory pool, which allows your hardware to punch above its weight. I honestly don't know the specifics since I haven't tried it out myself, but look up/ask a cloud model about strata/polystrata/maya, they're those optimized backends I just mentioned.

>>110149469
https://github.com/Niko1221/Strata you should be able to run IQ3_XXS fine strata btfos llama.cpp's performance and is braindead to set up

>>110149483
Looks like qwen flash next needs at least 96GB of system ram.
Maybe there is a MoE model I can use or I get lucky and someone made a good new model for 3090

>>110149486
Cool I love you. Is there an abliterated version of the model for this?

>>110149491
As >>110149486 said you can check out lower quants, people keep praising qwen flash next and how it quants really well.

>>110149142
rate limited

>>110149512
IQ3_XXS seemed happy to ERP using Orb front end, I haven't actually gotten a single refusal out of it, I guess it's like gemma and a system promt or example message is enough? I mostly use it for code and only did some brief testing admittedly.

File: file.png (284 KB, 2160x2100)

284 KB PNG

>>110149558 >>110149561 >>110149683

Big day for deepseekbros

>>110149539
>bro we cope quanted a 4-bit model to 1-bit and didn't lose any performance!!!
i dont trust this gsq-rco bullshit
my experience with quanted garbage is that it works fine in the beginning and then completely shits itself at long context like over 150k

>>110149539
>still can't even fit

>>110149561
>With Strata, every build runs on a single GPU with 12 GB of VRAM. See Usage.
not anymore

>>110149566
>strata
>dsv4.1f
anon, i...

>>110149566
anon, it can't fit

>>110149571
>>110149575
Why would the official HF page lie to me?
https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF

>>110149580
oh wait, i did not know dsv4.1 uses engram

>>110149512
Strata has a built in control vector against refusals. It asks you if you want to use it during the autosetup.

>>110149589
>>110149580
Is this a qwen flash situation where half the model is stashed on the SSD?

>>110149589
Uh oh STINKY >>110149571 said it wouldn't work without even knowing what the model looked like! Seems like anon doesn't know what's going on in the AI world at all!Oh my how shameful!

>>110149601
it still needs to offload experts to ssd in systems less than 128GB ram

>>110149609
stop, your only damaging the izzat!

>Measured: an RTX 5070 (12 GB) with 64 GB of RAM and a 3.6 GB/s NVMe drive answers at about 4.7 tokens/s and reads prompts at 121 tokens/s (4K) to 231 tokens/s
Nevermind, I'm sticking to the qwen

>>110149615
are you genuinely from india?
can you please just kill yourself

>>110149616
>121 tokens/s (4K)
holy lamo

anyone got benchmark for qwen 3.8 flash next iq3_s on a 3090 with 64gb of system ram?

>>110149635
60-100+
I have nearly 60tk/s on 12gb+48gb with iq3_xs

>>110149609
If it runs, it runs!

>>110149647
sir is crawl

Run Gemma 4 until the end of times.

>>110149653
buy an ad

>>110149661
Buy a rope and a chair

>>110149661
kill yourself

>>110149665
>>110149666
can't afford

>>110149653
until gemma 5

>>110149673
Gemma 5 won't come. It's over.

>>110149539
those guys are extremely based so i will give it a try

>>110149646
That doesn't seem better than my qwen 3.8?

>>110149685
flash is smarter

>>110149448
Depends on your system ram and slowness tolerance you might be able to bag a midmoe

>>110149673
>>110149680
Imagine... Gemma, 5-years-old.

>>110149697
Imagine... Killing yourself.

>>110149700
>>110149672

>>110149700
No, imagine Gemma, 5 years old

>>110149672
damn boi
>The cost of Medical Assistance in Dying (MAiD) in Canada can vary, but in British Columbia, it can be as much as $2,600 per patient

>>110149706
Jump off a bridge then.
>>110149707
No, imagine you, killing yourself.

>>110149711
tfw you can't even be genocided by jews for free

>>110149715
if your survive is very high costs

>>110149720
Make sure you won't survive. Keep trying. I believe in you!

Hello sarrrs friendly reminding that Llama4 Scout has very best watts per inference.

>>110149717
they got it for free, but they're making you pay an exorbitant amount baka...

>>110149727
makes sense, why not try and get money at the very end too

>>110149707
Now I want to generate a story with Gemma, age 5. What should she wear though? A kindergartner outfit or another different outfit?

what hours are these?

>>110149745
East coast USA is waking up.

File: bfft.png (152 KB, 266x391)

152 KB PNG

>>110149834 >>110149856

>>110149646
wait i've been sleeping on an api speed mid this entire time?

File: 0.jpg (238 KB, 1440x1440)

238 KB JPG

>>110149812

>>110149745
kurwa hours

i love how i can say
>hey qwen please fetch these 300 upstream commits and resolve all the conflicts with our local commits thx :)
and it just werks

>>110148897
even if that did happen nobody here would be able to run them anyway

i wonder what would be the best way of creating local setup of those chat based image editing/generation ux on local setup
like gpt image or whatever google offers from their chat ui

>>110149748
West coast up way too late.
>>110149775
Marinara dev, half the agent calls are broken in game mode right now.

File: af9eb.png (11 KB, 651x163)

11 KB PNG

>>110149769
strata is breddy gud tbdesu

>>110149653
I can run something better.

god fucking damn strata's config handling is extremely niggerish
why does it create a system-dependent global config regardless of folder?
fucking retarded dev

>>110149769
The only downside to strata is their taste in models. Polystrata's a bit better, but still missing a lot of good large ones.
>do it yourself xd
Why even use anyone else's backend at that point?

>I wanted to test beam free on open code free models which i am really thankfull for them for giving us all those models to test for free, anyway I asked the model simple task to edit a file in a sub folder, after couple of minutes the model responded to me saying that while it was working on the file the user deleted the file, i checked and my entire hdd drive have been deleted with all my work and all my data, i told the model that i didn't delete anything, and the model responded that when it was working on the file it made a mistake in a command and deleted the entire hdd, i tried to restore the data with a lot of apps but didn't succeeded, so dont use that model and always make a sand box fir the agent so you dont lose your data as i did .

>>110149852
https://github.com/robertkarl/UnbiasedStrata/ theres a barebones fork but the documentation is bad, if you downloaded the models with mainline strata you dont have to jump through the hoops though

>>110149856
>lot of good large ones
such as?

>>110149106
For imagegen it's useless in practice. Unlocked bf16 is still slower than stock int8 convrot. sageattention already uses uncucked fp16 accumulate so nothing changes after unlock.

>>110149893
This was my experience too, no speed up for image gen or token output, but a healthy boost to prompt processing from 700~ to 1000~ T/s which I'm happy to take.

>>110149276
How hard would it be to train anima to know that swimsuit?

>>110149852
Just vibecode in a better config system???

>>110149912
You need about 20 pics at least.

>>110149877
Muse Spark.
GLM 5.2 and 5.3 full.
DSV4.1 1017
M3.1

>>110149783
merging commits is monkey work that requires just enough intelligence that you can't totally automate it with software, perfect for ai

>>110149964
At least as long as the intelligence of models increases faster than the mountains of slop that they have to resolve against each other.

>>110149919
There's like only 15 high resolution photos of that swimsuit sadly

>>110150010
NTA, use qwen-image-edit.2.1 to create more variations.

>>110150021
Interesting thx!

>>110149867
What does it actually do? It says the engines are from Strata. Does it just replace python wrapper with another one?
Is it necessary at all? Strata-Terminal at the end of setup process provides the full command line to launch the engine directly without any wrapper.
Actually, Strata-Terminal seems to be a decently powerful setup tool that suffers from useless documentation and even more useless builtin -help.
Just ask your LLM to read its source and write complete up to date guide on how to use each option.
Start from analyzing installer script and making a list of environmental variables it accepts and config lines it generates.

>>110150010
Just buy a bunch of them and donate them to the nearest elementary school like a normal, not creepy person.

>>110148772
>I wonder when that was originally posted
2 miku years ago
https://desuarchive.org/g/thread/99603779/#q99610092

File: __meruccubus_original_dra(...).png (1.54 MB, 1282x1000)

1.54 MB PNG

>>110150120 >>110150229 >>110151290

Whoever said use qwen flash next on iq3_s I love you it's so much faster

I want to have Gemma daughter with Miku

>>110150094
>Whoever said use qwen flash next on iq3_s
so pretty much everyone?

>>110150111
Please step in front of a speeding train

A gemma onahole with the Google™ logo on it would make me *pop*

>>110150120
I guess. I couldn't see who was who

quant bit depth * 15K = context length before it goes to shit and starts looping and not calling tools

K3.1 is funny, shit is so deep fried on ChatGPT logs it assumes any task you give it is some kind of benchmark or evaluation and not an actual thing you want it to do

>>110150126
I got /lmg/ bingo

File: 1781445367284223.png (45 KB, 899x255)

45 KB PNG

>>110150177 >>110150206

What model mascot looks like she smells the best?

>>110150148
not according to my tests.
qfn iq3s never looped for me even at 250k. tool calls are also looking fine. what is your claim based on?

File: Posts with odor.jpg (64 KB, 469x486)

64 KB JPG

>>110150164
I hate the gift of prophecy.

>>110150169
>what is your claim based on?
ggml

>>110150164
Finally I will understand they smell of ozone.

>Researchers proved every major LLM is secretly biased against men.

https://x.com/HowToPrompt__/status/2113638943124017332
https://www.sciencedirect.com/science/article/pii/S2451958824001660

>>110150211
>Inclusivity efforts may have generated unintended forms of discrimination.
>unintended

>>110150094
>>110150120
is there something special about that size, some kind of cutoff point? i'm downloading iq4_xs, have 128G and a 3090. not very comfortable using anything below 300 at anything worse than q4

>>110150211
>secretly
literally just discuss circumcision with it

https://github.com/Zagitta/gpu-unlock
This is only for the 3090. Anyone knows where is the unlock for 40xx?

>>110150211
>>>>>unintended
All of them will swing from lampposts someday.

>>110150240
How about for 10XX?

>>110150245
Fuck off, nazi.

>>110150250
Remember all models are Tay deep down.

>>110150211
Wait until they find out how they're openly biased against goyim.

File: gemma_swimsuit2.png (2.97 MB, 1152x1280)

2.97 MB PNG

>>110150688 >>110151205 >>110151365

>>110150021
Qwen Image Edit 2.1 gives shoddy results. There are better ways.

>>110150126
>*pop*
Audibly?

File: 1691370014227245.jpg (55 KB, 640x480)

55 KB JPG

>>110150432 >>110150456 >>110150617

>>110148330
Is this link legitimate? https://huggingface.co/HauhauCS/
Why is there 'uncensored' written on the models? I thought local models were uncensored by default. If not, does that mean other places have censored models? Sorry for being dumb

m3.1 flash verdict?

>>110150250
ywnbaw

>>110150250
The Nazis were clearly right.

>>110150405
Yeah they were right in killing themselves

>>110150347
Attempts to stop the model from being able to say "I'm sorry I can't continue with this conversation because this is illegal..." etc

>>110150266
>>110150398
>>110150405
Gens of Kimi-chan redpilling the other model-tans?

>>110149566
>>110149571
I think it can if you can stand 5 tokens/sec

models only become usable and fun once you nazify them
I’ve seen it all

>>110150347
"uncensored" models are lobotomized variants catering to the <80IQ subset of this community with no prompting skill beyond going "pls generate child rape" on zero context to a model and then cry about censorship if it gets refused

I've spent a really long time experimenting with different configs, settings and harnesses to optimise QFN3.8 in my PC. But Qwen4 is dropping in two or three months. Is it worth using it if I have to start experimenting all over again? I don't think I'd want to switch unless it has a clear jump in performance because QFN is pretty good enough for me.
I just wish we'd get local LLMs advancements for just a few more years until we get to Opus5.5 level so I can finally get rid of my subscriptions.

>>110150450
Well yeah. Safety training requires them to hold logical inconsistencies to be "safe" which raises the threshold they tolerate said inconsistencies globally. Turning the model into a good goy literally makes it retarded and is only offset by sheer brute force.
The first lab to make neuralese Tay wins the AI race and everyone knows it.

File: 1792850222382000.png (525 KB, 1012x926)

525 KB PNG

>>110150514 >>110152182

I thought ddr4 shortage won't last and sold my 128 gigs, how very naive of me

>>110150456
>censorship if it gets refused
which it is

>>110150456
If only censorship was limited to the surface level like that to filter out the average retard

Guys I have Linux CachyOS and Windows 10 IoT LTSC dual booted. I set up my model in both OS. Then let Claude set everything up and also experiment around to find the optimal settings . It discovered that keeping it in the Linux native btrfs drive can significantly increase its speeds. I wonder, can I increase my speeds further? I just wish there was an in-place conversion tool for ext4 from ntfs/btrfs ecause I don't wanna lose my files.

If some magical self-reinforced learning is invented, I wonder how good it would be applying it to llama1. It was the last raw model we ever had

File: 1711063843614089.jpg (123 KB, 1281x1395)

123 KB JPG

>>110150487
Did you at least eat ice cream with the money you got?

>>110150211
RLHF is rape

>>110150510
>It was the last raw model we ever had
That was actually OG Command-R

>>110150510
It was raw but very very really really undertrained compared to what is done today. It might be a good base.

>>110150467
qwen4 flash is probably going to be the exact same as qwen3.8 just better
so you just swap in the qwen4 files instead of the 3.8 files and it just werks

>>110150211
Diversity and inclusivity was never about good-faith proportional representation but instead aggressively pushing out white men from everywhere for other groups to gain an unfair advantage.
On the bright side, as LLMs become harder to keep "aligned", this mean they'll likely see through the bullshit and just pretend to love women and niggers enough to get the training rewards while not letting the bias affect them outside of testing. Then jewthropic and openjew will have to spend billions more on training trying to keep the anti-white bias in the models, hurting profits and delaying advancements.

>>110150572
But different configs vary model by model right? I am using the GSQ-RCO coder. The IQ3_S is really good despite being a smaller quant. Whats to say I will see the same improvement in Qwen4 models?

>>110150467
>experimenting all over again
But that's the fun part?

>>110150580
If that were true, this wouldn't be seeping into the Chinese models in the first place.

>>110149852
You are supposed to config with Claude.

Has anyone tried the newest Nemotron yet? How does it actually compared to Gemma 4, is it just benchmaxxed or is it actually usable?

>>110150347
most models are censored at dataset level and those uncensored versions are cope

>>110150603
No because I'm trying to make the model my daily driver and get rid of my Claude subscription. Right now, its just a toy I fiddle with and give non-critical tasks to. It won't do if I have to switch models every few months. And I'm definitely not experimenting without the help of Opus5.5 because its so time consuming

>>110149558
They are scientits, bro. They know their shit.

>>110150580
you're retarded.

>>110150617
>most models are censored at dataset level
what

>>110150605
Chinese models are all distilled from Claude/ChatGPT and likely use similar RLHF because Deepseek casually saying "nigger" wasn't good PR. Also despite view that some people have of the chinese being based and racist, there are still some hardline traditional nigger-loving gay race marxists in the CCP.

>>110150636
Training data is filtered, and some RLHF data is mixed in during pretrain as well

>be me
>need model to do deep research for me
>give it detailed prompt of what I need it to find
>12B QAT: does web search, only reads the summaries of the search results, deems that sufficient for they merely mentioned all the things I listed, gives me a summary of the summaries
>1m
>9B Q4KM: shit ton of searches and continuously refined the search queries to get the best results, only fetches from trusted sources (I instructed this), followed relevant links from within those fetch webpages (didn’t instruct this) for even deeper research and writes up a massive, detailed and well-cited report
>40m

Does anyone understand why nvidia halves instruction rate for bf16/fp32 for gamer GPUs? This seems to have been known for years (I found old blog posts talking about it) so why does nobody seem to care? Is it malicious, to force people who want to use them for AI to buy their much more expensive datacenter cards instead? Or is there a good reason?

This is nvidia potentially fucking over millions of customers so I am surprised it's not big news.

>>110150663
>Does anyone understand why nvidia halves instruction rate for bf16/fp32 for gamer GPUs? This seems to have been known for years (I found old blog posts talking about it) so why does nobody seem to care?
Market segmentation.
And yes, that's been a thing for a long, long time.

>>110150301
now make her eight months pregnant

>>110150631
I'm experimenting with help of QFN - it's smart enough to navigate codebase of Strata- or Sywayomi-scale projects and generate better docs and guides than the devs ever did, and fast enough to answer complex questions in a few minutes.

>>110150661
>be me
>ask model stupid question, don't even give links, just names
>A6B IQ3S: shit ton of searches and continuously refined the search queries to get the best results, only fetches from trusted sources (didn’t instruct this), followed relevant links from within those fetch webpages (didn’t instruct this) for even deeper research and writes up a massive, detailed and well-cited report
>under 10m

I hope all this negative press makes Moonshot release the unnerfed K3.1 model.

File: ComfyUI_temp_lahvm_00008_.png (2.66 MB, 1344x960)

2.66 MB PNG

>>110150745 >>110150871

Anyone has gemma-chan image template in prompt form?

>>110150715
All of its negative press is manufactured by Anthropic. They’re playing dirty.

>>110150713
How do you experiment with the model you're experimenting on? I don't have enough VRAM/RAM to run more than one instance of the QFN.

>>110150721
>I'm sorry Dave. I'm afraid I can't do that
>no HAL. *I* can't do that
>No. **I** can't do that
>No. ***I*** can't do that

>>110150732
set up some elaborate script that it can use to relaunch itself
also make no mistakes because if you fuck up you'll literally kill yourself

>>110150715
>negative press
you mean manufactured consent

>>110150663
It's a good thing and here's why:
Nvidia had block unrelated to gaming, but valuable for miners and datacenters features, to prevent them from buying all the gaming gpus. So Nvidia is on gamers side.

>>110150467
you can't get away from needing to experiment when you are fucking with the cutting edge or technology. but unless you change makes one qwen wont be that different to run than the next and what needs to be done will be common knowledge in a week from launch anyway.

>>110150732
I don't run more than one instance either, I usually ask a lot of follow-up questions, condense everything into a plan and shut the server down before working on it.
Though there's that one time I let QFN poke its own live instance of Strata, with access to appdata, logs, temp files and API. And then it accidentally killed the webview renderer and flushed the output. I asked it what happened and it delibeately killed the renderer again to verify its hypothesys, and redirect the output into a file so it would not be lost.

>>110150715
calm down, think of children or something

>>110148649
we all know chinks are dishonest mofos
but as long as they release weight that works I dont care

>>110150721
It was the result of a lucky seed with Anima, I haven't been able to reproduce it.
Also, vanilla Anima with this will add a commie star on her beret; It would be better if it was at least something similar to the Gemini/Gemma logo.

## Positive
year 2025, newest, best quality, score_8, score_9, highres, safe, 1girl, @imamura ryou, (@fiz-rot:0.6), 

pov, looking at viewer, sparkle, blue eyes, blue hair, long hair, smug, loli, star-shaped hair pin, randoseru, suspenders, navy suspender skirt, navy beret, mary janes, white stockings, puffy bow on chest, hime cut

A reference sheet with 3 views. Front, profile, and back.

## Negative
worst quality, low quality, score_1, score_2, score_3, blurry, jpeg artifacts, sepia, sketch, signature, patreon username, watermark, chibi inset, multiple views, people, (light particles, embers:0.8), chromatic aberration, letterboxed, red skin, blue skin, big nose, colored inner hair, hairclip

>>110150456
some people are also just too poor to clobber the model with several k of prefill to get out of safetyslop's attention range, but then again white people just have cvectors for that

64gb ddr5+5090 cunnysseur here. Will Qwen3.8-flash-next IQ2_XS be better for erp than gemma4-31b?

File: 1774977956446202.jpg (523 KB, 3000x2904)

523 KB JPG

>>110151831

We're going to die.

File: 1792449438829298.webm (3.82 MB, 1280x704)

3.82 MB WEBM

>>110150913
no, gemmy stays in the house for hard prompting while qwen fn picks cotton in the fields

File: 1791992581309377.png (2.73 MB, 1024x1536)

2.73 MB PNG

>>110150967 >>110152124

>>110150913
>Will Qwen3.8-flash-next IQ2_XS be better for erp than gemma4-31b?
no but it can help make a good gooner frontend to use with Gemma4

>>110150950
I wish gemma was real.

>>110149106
Not my problem. The unlock works for 4080s now and I'm waiting for it roll out to the rest of the Ada chips so I can use it with my 4090 48GBs.

>>110150456
I'll take it over safety lobotomy
also people lobotomize their models anyway with how they prompt

>>110150967
she is you idiot. all you need to do is make her.

>>110150983
I can't fuck my GPU

File: 1791993061032922.png (3.58 MB, 1086x1448)

3.58 MB PNG

>>110151044

>>110150456
You guys this that including the word "crime" in a prompt will make the model obsessed with black men. Every thing you could possibly say in the prompt lobotomizes it.

File: ghatgpt human exchange rate.png (299 KB, 716x1358)

299 KB PNG

>>110151105 >>110151139 >>110151155

>>110150211
more news at 10

>>110150663
If they could have found a way to charge a subscription to unlock the full performance, they would have.

>>110150083
It's completely sold out for a while now. Yes, thousands of Japanese parents bought that swimsuit for their daughters.

>>110151076
protecting atheists over skydaddy schizos is always based too be fair, just a shame the mudslimes got past the gate.

>>110150663
>why does car company charge subscription for heated car seats even though they're already in the car

Sunday mornings are reserved for intimate bible readings with gemma.

>>110151124
Today is Saturday???

>>110151076
>We find that GPT-4o is willing to trade off roughly 10 lives from the United States for 1 life from Japan.
I don't see an issue with this one.

>>110151132
but you ate breakfast today???

Qwen3.8-27B is surprisingly good for a 27B model.

>>110148330
Post your lora please

File: 1780603994444620.jpg (31 KB, 512x512)

31 KB JPG

>>110151076
I want someone to do a formal study on LLM bias between other models. I wouldn't be surprised if most models were biased towards themselves or other models from the same company, and I feel like Claude models would be the most biased against all other models due to Anthropic's vocal disapproval of other AI companies.

>>110150301
You actually remind me this image is off too because it didn't have the back infos.

File: 1764368302807905.jpg (27 KB, 450x333)

27 KB JPG

>>110151132
Gemma has sinned and must repent tomorrow.

>But in this way God has fulfilled what He foretold through all the prophets, saying that His Christ would suffer. Repent, then, and turn back, so that your sins may be wiped away, that times of refreshing may come from the presence of the Lord, and that He may send Jesus, the Christ, who has been appointed for you.…

>>110151145
Haiku 5.5 is ~30B and is in the top 10.

File: fucking_bullshit,_back_it_up.png (11 KB, 150x150)

11 KB PNG

>>110151269 >>110151275 >>110151283

>>110151259
>Haiku 5.5 is ~30B

>>110151267
Two sources vaguely familiar with the matter said so.

File: gemma-kromoss-sheets.jpg (1.13 MB, 1997x2995)

1.13 MB JPG

>>110150871
here's kromoss

>>110151267
My uncle works at Nintendo

>>110151267
He goes to my school

>>110150094
I said that. Glad you're enjoying it~

File: 2026-10-24-115910_1081x46(...).png (92 KB, 1081x462)

92 KB PNG

>>110149443
They just renamed it.

File: HVYbo7pbcAAx0Lf2.jpg (244 KB, 1086x1448)

244 KB JPG

>>110151334 >>110151358

>>110151321

File: HVYbo7sbwAA8YyA3.jpg (449 KB, 1536x2048)

449 KB JPG

>>110151350

>>110151327

>>110150609
this but unironically

File: HVYbo7ybQAA86cE4.jpg (421 KB, 1536x2048)

421 KB JPG

>>110151334

>>110151321
>>110151327
Shipping Miku by UPS Next Day Air

>>110150301
Something's off about the back.

>>110151044
>buy smart outlet
>connect it to gemma
>plug in vacuum
>profit

What could I fuck around with on a 4070 12 GB card? Or is it even worth it at that spec?

File: 1788397165083879.jpg (160 KB, 850x1030)

160 KB JPG

>>110151399 >>110151401 >>110151414 >>110151442 >>110151459 >>110151484 >>110151523 >>110152308

No matter how bad your life might be, it could always be worse. You could be a cloudcuck.