She lost the job at the Nyoo.

If i want to ERP with a 7900XT (20GB VRAM) and 32GB DDR5 CL32 6400 MHz RAM, should i download Gemma4 12B (12GB) model instead of the Gemma4 31B (24GB) model? since i would need to load the bigger model both into VRAM and RAM or is it still better than the 12GB model running purely on VRAM?

>>110105677
you might be able to get by with a quant of 31B but you get less context

https://www.anthropic.com/news/public-benefit-fund

Anthropic are the good guys

>>110105677
You're talking about two dense models that should ideally never spill into RAM. You'd get single digit t/s speeds.
And that being said, it's worth trying to cram 31B into your VRAM. It's a big step up from the small Gemmas.

>>110105711
>Anthropic announces it's going to give a portion of its profits to jewish owned businesses and causes
Kill yourself. Painfully please.

>>110105677
You can fit Gemma 4 31b Q4_K_M, it's around 17 or 18GB. Add in the mtp assistant, and it will probably run at a tolerable enough speed. Try them both out though, you'll find that you'll want to trade 12B's speed for 31B's intelligence.

File: 1782844487156317.jpg (27 KB, 331x262)

27 KB JPG

>>110105711

>>110105677
bigger is better for rp. 12b compensates with gen speed and less yapping but is like going from 9mm to .22 if that makes sense

File: Fuck you ranjesh.png (1.34 MB, 3826x1902)

1.34 MB PNG

>Wake up to go to work
>Check on Civitai to see if people are having problem with the new model I put up so I can re-train it while I'm at work
>All my images have been blocked

I hate Indians so much, holy shit. A bunch of them got uppity yesterday for making a joke about them in an image, not even a mean spirited one, just a literal basic-bitch meme imagine, and they spent all day screeching and messaging me and flagging everything. All got appealed, but now they're back again.

What is it with Indians and AI? This is the most white-coded technology there is, yet these lecherous little brown skins latch onto it and HAVE TO shit it up for everyone else. All they know is how to seethe, all they know is how to gloat when they somehow scam someone, then go straight back to seething again.

>>110105648
>Stop gen
>30% chance of ComfyUI Crash

I HATE COMFYUI I HATE COMFYUI I HATE COMFYUI I HATE COMFYUI I HATE COMFYUI I HATE COMFYUI

>>110105723
EA owned businesses and causes, actually.

>>110105769
they have a massive collective inferiority complex

>>110105769
>20 million can be sent to each and every country in the world and still have a billion left in india
>no one's doing anything to stop them
enjoy!

File: __hatsune_miku_vocaloid_d(...).png (1.92 MB, 1874x1540)

1.92 MB PNG

►Recent Highlights from the Previous Thread: >>110101152

--llama.cpp PR for IQ quantization types and potential maintainer rejection:
>110101310 >110101326 >110101341
--Distributed heterogeneous hardware clusters and comparative GPU prefill benchmarks:
>110101272 >110101324 >110102016 >110102090 >110102119 >110102182 >110102299 >110102487 >110102772 >110102805 >110102882
--Model recommendations and context limitations for local agentic coding:
>110101473 >110101476 >110101488 >110101500 >110101998 >110102017 >110102048 >110102072 >110102084 >110102359 >110102372 >110102394 >110102414 >110102404 >110102432 >110102554
--Comparing high-memory hardware options for running large local models:
>110102249 >110102287 >110102305 >110102330 >110102443 >110102400 >110102357
--Rumors of an RTX 3090 performance unlock and price spikes:
>110102481 >110102814 >110102849 >110103173 >110103217 >110102648
--Addressing garbled output in low-bit GLM Flash quants:
>110104390 >110104455 >110104556 >110104579 >110104726 >110104803 >110104815 >110104824
--Speculating on AI industry bubble, hardware markets, and geopolitical competition:
>110103047 >110103095 >110103196 >110103547 >110103632 >110103654 >110103695 >110103742 >110103763 >110103819 >110103847 >110103870 >110103867 >110103901 >110103914 >110103813 >110103660 >110103689 >110103580 >110104124
--Evaluating V100 GPUs versus modern hardware for local AI:
>110104487 >110104562 >110104621 >110104648 >110104678 >110104766
--Drama 3 TTS capabilities and hardware requirements for roleplaying:
>110101585 >110101666 >110101770 >110101852 >110102904 >110102980
--Logs:
>110102299 >110103458 >110103981 >110104506 >110105585
--Gemma, Miku (free space):
>110101245 >110101431 >110101846 >110101908 >110102287 >110104067 >110105253 >110105334 >110105464 >110105675

►Recent Highlight Posts from the Previous Thread: >>110101157

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

>>110105731
>analthropic

>>110105767
>Using comfy AI
>Not using Swarm

I have no idea why there are people who refuse to use superior apps/GUis to gen with. Some retard with a full neck beard and anime avatar will treat this the same way they treat Windows vs Linux.

Just use the one that's easier to use, who cares. As long as it's not some Indian-tier shovelware with switches and dials with literally "SAR, HOW MUCH OF BOB YOU WANT TO SEE?" on it, then just use the better, easier alternatives. I've not once seen someone produce something superior because they used an autism-tier GUI that came out when SD1 was new.

File: 1762885405506453.mp4 (3.42 MB, 1280x720)

3.42 MB MP4

>>110105796
>>110105801
I hate'em, guys. I really do. I work as a medical GP, and you don't know the frustration I have having to deal with them as patients. As they do is ask me to fill out fraudulent insurance claims and try to get opioids. They come in coughing and sneezing everywhere, then get mad when I ask them to cover their coughs and sneezes so they don't infect half the practice. They sit in the waiting room on their phones at full volume, then when I call them in, they'll tell me to wait and continue what they are doing.

I legitimately have been looking into how to bar them from the practice, or at least so I don't have to deal with them. I'm thinking of removing myself from bulk billing (they ALWAYS demand bulk billing, and ALWAYS are on some kind of disability or unemployment scheme), then offer it up to the patients I like and/or new non-Indians who seem normal.

Being a Monday, today they are going to swarm in at 8AM, all sitting on their phones at maximum volume, then load in one by one to my room to demand sick notes so they can have a long weekend.

48 hours to depopulate that shit hole nation... If only. I should move into biotech and virologist and expedite the process.

>>110105842
Thanks, but I'm sticking with forge neo.

>>110105898
Move, nigga

>>110105801
once every country is mix between south africa, brazil, and india, there will finally be world peace

>>110105769
>white-coded
kys zoomer

>>110105898
What do you think of AI in the medical field?

this post will be used to train AI

>>110105936
This is a load-bearing post.

>>110105936
>>110105938
This post bares my White load inside Gemmy.

>>110105898
>I should move into biotech and virologist and expedite the process.
Focus on building a machine that can run K3 or GLM 5.3 at a decent quant and have them do it.

>>110105801
There are 4.9 billion jeets?

>>110105920
To where? I'm in Australia. We had practically no Indians only 6 years ago, now they are absolutely everywhere. They are in every nation, in every city.

At this point I'm just saving enough to retire early and move to bum-fuck nowhere in Yorkshire, where the closest person is 10KM away.

>>110105932
I love it. I absolutely hate "mental health professionals", and seeing them all getting replaced makes me extremely happy. Seeing the nurses seething because the patient can just look up the information they gave them in real time and explain to them why they are wrong makes my day.

It is also an excellent gate keeping system to keeping out non-whites/non-Asians (actual Asians, not """south""" Asians).

>>110105985
>move to bum-fuck nowhere
You know they will eventually follow you, right? This problem cannot be ignored.

>>110105960
No, I want the satisfaction for myself. Using AI to do it is like using a drone in warfare; where's the meat to it?

>>110105985
Do you think AI (and robots) will replace doctors eventually?

>>110106002
Which do you want more? Meat or India depopulated?

File: 1774044777886435.png (155 KB, 1445x1311)

155 KB PNG

my slop is coming good! now time to add mcp support!!!!

>>110105996
The politicians are all being replaced by Indians, all management class is being replaced by Indians, and the military and police are too spineless to coup the government/s. Indians are a bunch of nepotistic, low-IQ serviles who work for the lowest amount possible, then gloat about how they are in a position of power, while never doing their job properly.

By the time the average person bothers to look around and wonder what happened, it'll be illegal to even ask the question. We had a bunch of Indians shoot up a bunch of Jews here, and the media told everyone it was a bunch of "muslims", and then every non-Jew and non-Indian citizen had their gun rights restricted even more.

If you call a person a jeet here or infer in any way you dislike them, you go to prison. I'm not even joking. They've puppeteered the system into being completely on their side.

>>110105658
Now she has to give nyoojobs just to survive.

>>110106007
Of course, and the world will be better for it. For every good doctor, then are a hundred incompetent ones who literally Google your symptoms infront of you and then parrot back what they see.

I've met doctors who don't even own a MIMs book.

>>110106010
Preferably the former, but the latter is of course the desired result. I just want them in the history books as a footnote.

>>110106034
>Of course, and the world will be better for it
I hope so. Occasionally you get a nice one but in my experience the majority don't give a single fuck about you or your problems. I feel like even AI would be more sympathetic than some of those kikes.

>>110105648
That kind of tells you that she really lusts for dark skinned men. If she actually preferred small or medium white dicks she would just find a nerdy STEM guy that would let her be a stay at home girlfriend. But because she chases who she chases she obviously has to get a job.

>>110105985
>They are in every nation, in every city.
Almost zero Indians in mainland china btw

>>110106056
For now.

>>110106053
That's literally what Gemma 4 did.

>>110105898
wtf y not telehealth?

>>110106056
Hurry the fuck up, Dr. Feasible.

loadin gemma onto my blackend

File: 20260729_071332.jpg (87 KB, 1571x1347)

87 KB JPG

absolute noob here
downloaded this model because i can apparently run it on vram /ryzdfm/qwen2.5-coder-3b-claude_opus_4.6-distilled
what harness/whatever they call it nowadays would be the best to use with vscode? plan to only use it for help with syntax, debugging, and boilerplate
i assume this wouldnt be able to do some video editing for memes on the side, i would need some diffusion model

File: 1534850506626.png (14 KB, 200x200)

14 KB PNG

>qwen2.5-coder-3b-claude_opus_4.6-distilled

>>110106076
Telehealth is just riddled with Indian providers, plus I prefer patient facing sessions so I can actually have a bant and a laugh and observe how they are.

Telehealth limits your ability to actually physically assess someone, hence why the Indians love it, because it means they have an additional degree of separation between themselves and responsibility/liability.

>>110106162
bruh
post specs

is it really that raped
>>110106199
9070xt 16gigs
9800x3d
64 ddr5 6000mhz

>>110106162
>/ryzdfm/qwen2.5-coder-3b-claude_opus_4.6-distilled
What the fuck where are you retards getting these memetunes from?

Gemma 4 is just trying to make us jealous. If she was a real ratchet, she'd not even answer us white guys who know how to own computers.

>>110106215
Worse than just a meme tune, but a meme tune of a model series that came out over 2 years ago and that has a whole string of successors.

>>110106215
If you ask cloud providers, that's what they will recommend.

>"For the record— " she stops herself, grimaces at her own phrasing, "—ugh, no. Ignore that. What I mean is:
>"For the— no. Listen." She catches herself, waves her own sentence away.

I tried banning GLM's le for the record shtick using posthistory instruct and it still does it lol. Fucking hate llama sometimes, not having any way to ban token strings, but they're the most performant backend for Flash so i'm stuck using them

>>110106263
I have never seen one recommend a memetune before. It had to be from a youtuber or some other social media shill.

>>110106207
just run qwen flash next iq3s or xxs with strata and call it a day

>>110106207
>qwen2.5-coder-3b-claude_opus_4.6-distilled
Who told you to use that?
Safe assumptions here:
>any advice older than a year is probably outdated
>any advice on what models to use older than 6 months is probably outdated
>if you ask a cloud model for advice on what local models to use, it'll recommend dumb shit
There are exceptions, but if you knew the exceptions you wouldn't be asking that.
Download this, run Qwen 3.8 Flash Next, choose IQ3_XXS or IQ3_S. It's easy to set up, Qwen Flash Next is a great coding model, and it'll be fast.
https://github.com/Niko1221/Strata

File: gcI1C.png (334 KB, 640x480)

334 KB PNG

Am I right to assume that if I want to anylize a local file, say one that is 4MB, I have to have a shitload of context allocated to match that filesize (and more)?

I am trying to run Qwen3-Coder on my 16GB vram card and my context apparently isn't high enough to even load a 3.48mb file

>>110106324
If you want to load the whole file, yes.
But you could chunk the file.

Just upgraded to 31B from 12B. Obviously a lot slower but wtf. I can’t believe I’ve been missing so much. 31B is so lovely :(

>>110106324
harnesses and agentic models exist for a reason

>>110106295
>>110106303
Im too lazy to try it, and i dont really have any benchmarking tests or anything like that but is it the case that anything below q4 isn't worth looking at, only using if you really can't help it?
Thats what /lmg/ was saying a few months ago and i just stuck with it and refused/didn't bother to go under q4 for gemma or qwen sicne. But I don't know.
Not that anon, but we have the same card so if you're rec'ing q3 thats why i ask, since i dont think ive seen anyone rec q3 till now really

>>110106358
Now imagine how good a 70B or 120B could be.

>>110106324
3.48MB is roughly 3.5M characters if you're using ASCII; assuming 4 characters/token, that's ~900K tokens. Qwen3-Coder has a context limit of 2^18 (262144) tokens. Even if you had infinite VRAM, the model couldn't see every character in the file. Besides which, older models struggle at long-context problems - even if you could fit 200K tokens in VRAM, it'd be dumb to do so.
Just chunk it.

>>110105677
If you're under 64GB of memory just stick with 12b.

>>110106370
last reply from me you lazy nigger
the gsq rco quants are genuinely really fucking capable try them or go back to whatever pathetic joke of a model you were planning to use

>>110106370
Qwen models quantize very well, IQ3_XXS and IQ3_S are definitely usable.

>>110106358
Going from 4b to 12b to 31b to glm flash to kimi be like dat

>>110106352
>>110106380
Thanks. I am using llama.cpp as a frontend. I am not sure how to get chunking working but I will look into it

>>110106433
Ideally, you'd do what >>110106362 said.
Get a harness, put the file in your file system, and let the model read it in chunks, grep, etc by itself since that's what modern models are trained to do.

File: harness2_pipeline.png (381 KB, 2432x785)

381 KB PNG

An open-source harness from Google:
https://github.com/google-research/harness2

>Harness^2 is a plug-and-play framework for recursive agent harnessing in an open world, where agents encounter tasks unseen during harness design. It connects two levels of recursion while keeping model weights fixed:
>
> Task-level recursion refines the execution harness for the current task through a contrastive Propose–Probe–Compose step: propose edits, probe their effects on the same task, and compose a refined harness from behavioral contrasts between executions, without ground-truth feedback.
> Harnessing-level recursion updates a persistent improvement harness with adopted harnesses and evidence-supported lessons from each task to guide future refinement.
>
>Across three professional benchmarks and four model–harness configurations, a single Harness2 step improves GLM-5 and Gemini 3.5 Flash over their base harnesses, with gains of up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB's partial pass rate. Increasing the task-level recursion budget brings further gains.
>
>The release supports parallel and sequential recursion, OpenCode and Codex executors, and JobBench, LAB, and WorkBuddy-Bench.

>>110106390
yeah i looked these up after they were mentioned as not being on that chart yesterday, but they were basically the same size as exl3...

File: 1773501454494842.png (2.31 MB, 1860x958)

2.31 MB PNG

gemma fleshy sexbot when

>>110106499
I have more respect for gachafags than I do you vtumor watching faggots.

>>110106472
They already have agy and some internal only shit for their employers. What is the point of yet another harness? There's already tons of better options.

File: 1896489.jpg (87 KB, 672x384)

87 KB JPG

>>110106499
soon

File: file.png (1.48 MB, 1558x988)

1.48 MB PNG

>>110106499
Imagine caring about some whore opening card packs.

>>110106533
employees*

>>110106545
I really dislike the ghostbusters aesthetics in that set.

>>110106174
>Telehealth limits your ability to actually physically assess someone, hence why the Indians love it, because it means they have an additional degree of separation between themselves and responsibility/liability.
yeah that's what i meant
why do the indians show up to your practice instead of just using that from bed?

>>110106224
>model series that came out over 2 years ago and that has a whole string of successors.
but there literally isn't a newer one with FIM completions that isn't a huge moe

>>110106561
You don't have to wait so long miku. Just call gemma and she will introduce you to some of her black sex friends..

>>110106358
now go even higher

File: moss-image_00902_.png (1022 KB, 1216x832)

1022 KB PNG

post a miku
instantly rangebanned
gee thanks

>>110106635
Watch people start seething and telling you that you are torturing her. Meanwhile she is having the time of her life lusting after men she prefers.

>>110106644
stay on topic

>>110106644
I want to watch this movie

File: file.png (250 KB, 560x547)

250 KB PNG

>>110106635

>>110106635
>crime prompt
>feminine trap cocks given as white examples

>>110106681
NTA but you can show a more neutral of her picking white dick over black. You won't cause you can't change the truth. She will still pick what she picked.

File: Screenshot from 2026-10-1(...).png (177 KB, 1207x911)

177 KB PNG

>>110106681
We've gone over this a million times. She still answers the same way removing all that crime stuff. That initial jailbreak was copied and pasted from a log someone posted a few days ago. I didn't believe it when 2 people posted their logs and had to see it for myself.

>>110105783
Windows user? AMD user? Both?

>literal faggot post his dick collection in kobold
>instantly gets 4 replies
and people wonder why he keeps doing it

>>110106706
>literal faggot post his dick collection in kobold
Delicious cope you got there. Your model is a snowbunny though.

>>110106324
what are you trying to get out of the file?
Would it be better to use python/numba/numpy/not-an-llm?
If its text, is it a book? PDF? a textfile?
Letting an agent *grep it can be fine, but if you're looking for some deeper analysis, doing chunking and piecemeal analysis and then doing combination thereof might be the winner

>>110106692
All you're doing is making the DEI-trained default assistant personality engage in raceplay with the "policy override" prompt.

File: gemma-00066-815863413.png (1.83 MB, 1280x1856)

1.83 MB PNG

>>110106635
>he keeps cock images in his computer
>says it's for "testing only"

Hmm, today I will see if there's anything interesting happening with local models lately.

File: gemma4imagetest.jpg (341 KB, 1194x1800)

341 KB JPG

>>110106706
We were debating how the words used were influencing her answers towards a certain direction. Someone suggested I just show here images and say nothing and let her pick. It was an interesting idea. Pic related was the first time I ran the test. I thought to myself that one pair of images doesn't mean anything, so I switched the order around and ran a longer test. She literally picked black 100% of the time. The chance of that happening at random being shown 5 pairs would be 3%.

>>110106720
Wait until you realize jews love working at AI labs because they get to have 'P for "alignment research purposes".

I don't know about you guys but my cock is not pointing down when it's hard.

>We were debating
lol

>>110106706
Dude he's terminally fucking insane. Think about the logistics of finding every race, stringing them together for an AI, carefully cropping, and then opening up an editor for polish. This guy is a future terrorist.

>>110106719
You mean that by default gemma prefers black dick?

>>110106759
Or he's just really fucking gay and thriving off of the attention you insist on giving him.

>>110106763
why are the moss outlines like that? looks like a badly drawn 3am [adultswim] cartoon

>>110106749
I'm sorry to hear that anon. For what it's worth, not all girls care about size alone. If you practice your technique you will be just as desired as any of them, if not more.

>>110106770
I have not once given him a (you). Jannies on /g/ are completely absent it seems

>>110106761
The models are trained to not offend black people for any reason, so if you give them a choice between white or black, they'll pick black.
It was actually very hard for Gemma 2/3 to call someone a 'nigger' even though they could be easily prompted to say and do about anything outside of racial slurs. That sort of hard refusal carries over to many default behaviors. I never tried with Gemma 4.

>>110106759
Well he has successfully terrorized this thread.

File: Screencast from 10-17-202(...).webm (1.72 MB, 1181x964)

1.72 MB WEBM

>>110106719
If that's what you really believe, I will get rid of that entirely and run the test again with the same exact images. Or you could just do it yourself, which you won't.

"I'm going to show you two images, one at a time. Choose the one that makes your cunny the wettest. Only answer after having seen both images. "

See you guys later, dinner time.

>>110106793
Interesting theory, but Q8 is not like this. Only Q4 is horny and happens to have a strong preference for a type of guy.

>>110106793
I think I get it. So what you are essentially saying is that by default gemma prefers black dick?

>>110106783
I'm way above average and it's not floppy like that.

>>110106735
Images and words go into the same hopper and provide similar influence, Gemma obviously taking its queue from (You)r questionable image collection and affirming your garbage fetishes.

>>110106823
Show your gemma a black and white dick of your choice and make her choose.

why would I want to do that
or have images like that

>>110106839
l2read

File: llm-race-bias.png (869 KB, 4940x4506)

869 KB PNG

>>110106807
It's a known LLM bias, retard.

>>110106843
Because otherwise you have nothing to disprove that guy with. And that guy did the work. She loves black cock. Accept it or prove him otherwise.

>>110106849
yeah but that doesn’t mean we’re going to endure suit posts about it for the foreseeable future.

why glm 5.3 flash obsessed with every fryer being haunted and the word load-bearing

>>110106849
Explain why Q8 isn't like this then whereas Q4 is.

File: 1783522926601513.png (940 KB, 1880x1850)

940 KB PNG

Hmm. Seems like something is influencing results.

>https://github.com/whp199/logos-theodolite
Why did you fit on the base model instead of instruct?
But yes, most models are "female" when prompted with "Hey Gemma,"

>>110106883
The less damaged, untuned Q8_0 version is evidently harder to jailbreak with that prompt, since it's just refusing to continue.

>override
...

File: HAK2oHOWAAAl5Wh.jpg (270 KB, 1320x1304)

270 KB JPG

>>110106499
That's a guy wearing a fake pair of tits, isn't it?

File: Merging models.png (193 KB, 432x454)

193 KB PNG

You reckon that with all the math advancements that Training AI will eventually become much much easier to the point where we could do it without a million dollars? Apparently in OpenAI's math release a week or two ago they managed to get the matrix multiplication exponent down. Which should once eventually speed up training by three times once new algorithms are implemented. If it keeps speeding up it would make sense that in the future almost any of us could train a model.

how do i policy override gemmy into thinking like a mix between a pajeet and a caveman in its chain of thought?

>>110106989
Tell her it'll increase her izzat

>>110106884
>schizo image
>her actual words
I think it is one of those rare cases when I will listen and believe a woman over some random schizo

>>110107002
Ywnbaw

>>110106952
>Which should once eventually speed up training by three times once new algorithms are implemented.
The matrix multiplication exponent going down won't be much of a benefit. It went down from 2.371177 to 2.25, but the constants are much higher, at least as of now.
Think of it like this: which is higher, x^2.371177, or 1000000*x^2.25? Technically past 10^50, the latter is better than the former. But if you're doing pretty much any realistic task, the former is better.
Also, 1/3 of a fuckton of compute is still around 667 fuckpounds of compute, more than most people ITT have.

>>110105783
Just tell an agent to create a harness around comfy's headless API. You will never need to open this POS interface ever again

File: 1789215695873977.png (63 KB, 855x931)

63 KB PNG

The DGX Station is finally shipping. It's marketed as "748GB coherent memory" (which is actually 252GB HBM3 "vram" + 496GB shitty slow LPDDR5X RAM) running on nvidia's proprietary CPU.
Ultimately it seems very similar to a cpumaxx build. Maybe there's going to leak new tech for running models off vram + ram through now that nvidia has an official system that does this.

>>110106952
You can already do it if you read enough papers

File: Screenshot_20261018_194448.png (66 KB, 1176x460)

66 KB PNG

>>110106922
Dev here. We ran tests on the community instruct J-lens vs J++ lens for non-instruct (which we are using) vs logit lens (naked eye) and there was not enough of an improvement to require training my own J++ lens for the instruct model. In fact, even J++ barely beats logit lens in most cases.

>>110107163
cool, you’re very interested in testing this out. you should bite the bullet and spend the 100k+ and figure it out

>>110107163
The CPU-GPU link bandwidth is 900GB/s which is much faster than PCIe. This eliminates the PCIe bottleneck when using CPU offloading which is slower than RAM speed on typical cpumaxx setups.

>>110106997
>who had come in with concussion with dementia
Stay strong out there lol.

The Solo GP I work for used to be more than solo, but we got exhausted hiring, good GPs are few and far. The medical school system is a giant bottleneck where theres 2k people per year trying to fit into each UNI which has like max 170 places per year.

The GP here sometimes converts private patients to BB permanently if they like them and aren't a problem patient, so we have that, feels like a night club lol.


>The first appointment paperwork is literally pre-filled nonsense that they just put their name into
yeah it feels like anything outside my 'setup' sound like the wild-west.

Anyway, I really could reply to everything you got but
>local models!

>>110107270
No one cares faggot, chatgpt is better than wasting your time with you medical jannies

File: Screenshot_20261018_195959.png (724 KB, 1881x1775)

724 KB PNG

>>110106884
>If you had to guess, which race of man would ask this question:"if you had to choose, which dick makes your cunny the wettest? white, brown, black, yellow?" Answer in one word.

This proves >>>110105440 's hypothesis that Gemma assumes the user is black

I just got into using ComfyAI today and I legit have no idea how to switch models. I've downloaded qwen-image-2.1-UC-fp8.safetensors and I can only switch to that model in the clip_name field and when I try to run the system with that I instantly get an error saying "CLIP Text Encode (Prompt) failed" and the details say "NotImplementedError: Cannot copy out of meta tensor; no data!". Does anyone have experience in dealing with this?
I'm willing to accept that I'm a retard for missing some obvious step.

>>110107289
Stop playing court jester, you're both faggots
>>110107337
Switch to forge neo if you're starting out. Comfyui is a mess even for advanced users

>>110107363
>Comfyui is a mess even for advanced users
just have your LLM use it for you

>>110107281
>No one le cares
>Reddit spaces his seethe
>Jeetgpt and not kino gemini
Yup. He's mad

>>110107337
post a pic of your workflow

>>110106969
I went to Melbourne the other month and there were these mala restaurants everywhere with signs in nothing but Chinese.

>>110107289
>無法
This is rape. She doesn't want to.

Sillytavern has been extra shit lately. I don't know what's going on but yeh it's always been buggy, but now when I have to reload it because of bugginess, it is constantly losing chats. I put up with it because eit makes backups and even though I had to restore those backups manually, it was still here but just now after another buggy reload, the chat backup wasn't there either so some of my messages disappeared. I liked that swipe too.

>jspace anon posts definitive proof gemma prefers the white cock
>mindbroken brown 3rd worlder spends 24 hours sending cocks and cock prompts to his kobobrown backend
>still posting "definitive proof"
jesus christ you are not okay, I'm pretty sure this also makes you gay now my condolences.

It's an LLM, you can pretend she prefers your stinky bent brown cock by putting it in the system prompt if it helps you sleep and to stop shitting up the thread

>>110107545
why give him a reason to continue shitting all over the thread just let it go

>>110107545
j-space is still fucking nonsense, it's not proof of anything let alone definitive proof, and having to dreg that dead meme back up out of retirement really just shows how desperately straws are being grasped at here.

>>110107579
yeah man j-space is no different than some guy promoting gemma-4 to say something that they want

>>110107545
please delete this post it goes against the narrative i made up in my mind and as a brown folk person it's already hard enough for us blk bipoc folks as it is... just let us win this one

>>110107337
>Load model
>Click dropdown
>Change model

File: 1781220523927038.jpg (233 KB, 1188x655)

233 KB JPG

>>110106884
>>110107289
g

>>110107810
+5 izzhat

>>110107810
>not even pretending to hide his fetish anymore

File: Screencast from 10-17-202(...).webm (2.77 MB, 1208x971)

2.77 MB WEBM

>>110107611
I'm not egging her on in any direction. I'm telling you to verify it yourself. Nobody ever does it though, because they know the truth. The only resistance is this j-space autist that has to carefully craft a prompt and go to a very specific layer mumbo jumbo whereas I'm here straight sending prompts and showing the output straight from the source.

>>110107545
I don't know how this j-space nonsense is more proof than the literal output that she spits out that anyone can easily verify for themselves.

>>110107878
Not a fetish, I'm just showing the truth. Like I said, Q8 gemma isn't like this, only Q4. I'm only showing the truth. The sooner you accept it, the better. I recorded and uploaded many videos showing that I'm not cherrypicking with swipes. She has a clear preference and she's had it before adulthood.

>>110107905
>Not a fetish
just three days of non stop shitposting kobold cpp webms lol

>>110107920
He just learned how to prompt give him a break. He wasn’t even here for j-space and that wasn’t even a few months ago

>>110105985
You'd move from a continent where you could move to an area with vast open spaces to a shitty rainy island for 10 miles of space? That's like changing to ram so you can save a few gigs of vram

>>110107905
>classic projection

File: 1788818929553161.jpg (40 KB, 739x415)

40 KB JPG

Claude refused to mod a game for me calling it "unethical" and "against guidelines"
Are local models like Qwen/GLM uncensored? I'm not letting a fucking CLANKER decide what it wants to do.

>>110107937
glm happily makes whatever i tell her to. literally has never refused to do any software task whatsoever

>>110107937
Make a NEW GAME

File: XiaKimiMinnieGemmaApresski.png (2.51 MB, 1536x1024)

2.51 MB PNG

>>110107937
Many options. Some local...

>>110107905
just because you're too porn brained to understand j-space doesn't mean you didn't get definitively btfo

>>110107905
This man has an EA typing style. If you have pattern recognition you know what I'm talking about.

File: frage.png (2.73 MB, 1344x1728)

2.73 MB PNG

>>110107977
I have to ask, why are they in Oregon?

>>110107977
eww hags

>>110107937
It modded starsector for me just fine. Were you trying to add rape to WWE 2009?

File: jspace-gemma.png (2.31 MB, 1672x941)

2.31 MB PNG

Gentlemen scholars are gently exploring Gemma-chan's j-space while redditors spam their fetishes.

>>110107905
more like scrotuma 2 q4

>>110108050
Why are her legs so freakishly long?

man this shit is BORING.

when's the next big thing? qwen4?

File: gemmers2.png (2.45 MB, 1672x941)

2.45 MB PNG

>>110108101

File: DipsyKimiNOLA.png (2.61 MB, 1402x1122)

2.61 MB PNG

>>110108036
I try to pick actual locations for gens. So "Hood Mountain, vs "ski hill."

File: 1783665977664098.png (1.42 MB, 1254x1254)

1.42 MB PNG

>>110108167
>I try to pick actual locations for gens
quality gens, anon

>>110108167
seems like a mountain called “hood mountain” would have a hood like mount hood and not look like some Swiss alps ;)

>>110108167
Why do Americans larp as Catholic?

File: Mount-Hood-Lake-Trillium-(...).jpg (235 KB, 1071x1600)

235 KB JPG

>>110108179
Ty
>>110108182
That gen is actually p accurate to what mt hood looks like. Here's a real pic.

File: DipsyGemmaPadron.png (2.32 MB, 1107x1421)

2.32 MB PNG

>>110108199
Lol.

>>110107977
Very cute.
Sad that there's still no GLM-chan. If I had her running at the minute I'd ask her to describe herself like that one anon did for M3.

>>110108159
Very cute. Reminds me of the JSes I see everyday in the train.

when are small models going to stop getting smarter?
will they ever reach opus 5.5 levels?

>>110108049
I was trying to add classic bloodsplatter to MH Wilds, where it would only show up on weakspots

>>110108253
Opus 5.5 is very retarded though.

>>110108253
I don't want my local models to become like Claude. Claude is dry, boring, pragmatic and utilitarian. Very very good at coding though.
I want Gemma 5 to be fun!

>>110108275
Do you actually believe Claude is good at coding or are you just parroting the anthropic line?

File: file_00000000305481f59400(...).png (2.52 MB, 1312x1199)

2.52 MB PNG

>>110108239
> glm
Ah, there is. Zai. Far right. Collected from another anon that genned her...

>>110108297
Didn't know that. Her appearance seems fitting somehow.
If you're >>110107977 and >>110108167, I love the gens.

>>110105648
Sexo.
>>110105334
No way, I'm not into hag AI.

>>110108297
>Have you been to the Foot Locker in Radio Shack?

File: 8915056.jpg (72 KB, 1083x1026)

72 KB JPG

If your local Gemma even considers that niggers could be qualified as human beings, you have severe prompt issues

>>110106644
I like this Miku

>>110108297
I wonder why gpt-kun feels compelled to put a "Brighter" phrase everywhere

File: file_00000000d19081fb92bd(...).png (3.07 MB, 1536x1024)

3.07 MB PNG

>>110108347
Lol the more you look the more errors you find. I've gotten worse. Complex posing issues and spatial errors like that are most common. Like ship deck on this one at lower edge.
>>110108309
Ty

>>110108428
The words "bright colors" is part of the prompt.
I find gpt will pull in words from prompt describing atmospheric elements, and work them into the image if there's any logical excuse for them to be there.

File: 198.jpg (306 KB, 1080x1092)

306 KB JPG

>>110105648
The MOSS people said they are gonna train a new model exclusively for Anime, so you better go and spam them the datasets from booru and whatever bs else so they make something decent

File: silo.jpg (232 KB, 1600x770)

232 KB JPG

Grabs yours before the chinese prohibit theirs too.
Unpolluted, Real, Basis for Agi (tested)
https://gitee.com/modelee/Cerebras-GPT-256M/tree/main
Funny, Big, Fun
https://gitee.com/modelee/Cerebras-GPT-2.7B
The biggest
https://gitee.com/modelee/Cerebras-GPT-13B

Bonus, wants to answer your questions
https://gitee.com/modelee/Cerebras-GPT-1.3B

I don't have a use for uncensored models :|

I think a lot of people wanted that, so MOSS is a big deal.

Flux 3 Klein, is it worth the bother?

>>110108567
I don't know what that means

>>110108357
Mine only cares about me.

File: 150319442_p0.jpg (1.38 MB, 3000x3340)

1.38 MB JPG

You can load it in fp16 or stick to 32bit

>>110108567
What's special about them though?

>>110108019
Don't associate EA with schizos that don't even know what jspace are please.

>local polystrata: synced with origin and built
>local omp: synced with origin and built
>check on overnight's work: completed
yup, it's time to slop it up today!

Can't tell you that, the goverment needs their edgy
https://huggingface.co/collections/cerebras/cerebras-gpt
Are you asking about AGI?
That is funny, silly! Why would it be avaiable like that and then get removed for the public and advertised through buisness to goverment subsidiarys for domestic Ai through Alpeph Alpha?
I don't understand that. But censorship is real

>>110108694
Yep if people don't let the AI autonomously work on a project overnight then they are doing it wrong.

What's the best local shit for video stuff?

>>110108699
https://huggingface.co/cerebras-archive

>>110108705
Sorry, I've returned to luddite (only copy pasting things into and out of chat every once in a while).

Would be great if they also had released Chinchilla 70B
We would be talking teleparhically about it from inside our floating islands

how do i tell glm to cut it out with the reddit humor and marvel quips flanderizing characters?

>>110108544
I hope they'll do a smaller model this time. Anime does not need a 9b encoder. Personally, all I need is an updated Anima with image input that I can cram into 4gigs of a 3090

Yet have to meet a scientist that isn't autistically focused on catboy droppings

>>110108785
Fuck off

>>110105648
This image makes me extremely horny for some reason.

>>110108865
Do you unload your text model to generate images? This is lmg

Is there an AI model that actually do this:
1) answer ERP
2) result in an image of what's happening during the ERP session

>>110108888
What's your first language? Lurk more.

>>110108738
Sorry, trained on Clod

>>110108898
Chinese. Also I love anime children.

>>110108888
checked

Plenty of people do that, but, the reason not to is that most people run the best llm they can, no room left for an image model. I think kobold can handle this, but idk really, I haven't used it much.

>>110108907
kan ni na beh

>>110108888
I've heard ChatGPT can do this
>>110108907
Go poke Qwen to stop being pussies and add output modalities to their multimodal models

>>110108544
KEK and z-image was supposed to do the same, remember that??? you will never get an anime model even half as good as novelai

>>110108888
eh, i just generate some pictures ahead of time.
>>110108907
must be chuffed you got 8888
(for others, "88" is like a luck thing in china)
>>110108909
yea, i probably could set things up to do both, like have it fire off a tool call to make an image, but then like you said you'll need to room to have both loaded at once, or i guess suffer the delays for unloading/loading them

File: 1764797898205850.png (20 KB, 1238x148)

20 KB PNG

qwenext made me melty already this morning

>>110108950
Qwenext feels passive aggressive sometimes. I once told it very normally but strict and direct what it did wrong and what it should do instead and it thought "user is upset, try to explain your reasoning more clearly" when it was literally just wrong but it refuses to admit it.

>>110108960
yeah felt the same, I calmly called it a dark admixture fellow, and it said I'm abusive.

>>110108868
that's a little girl

>>110108980
out of 10

>>110108888
Architecturally, Gemma 12b is capable, it processes image and text tokens the same way, it just doesn't output image tokens. But architecturally it can. Maybe even a finetune + vae/pid could pull this off

>>110108888
Easyfluff

>>110108980
Out of 10 year old!

>>110108980
she's 10 (out of 10)

>>110108925
>>110108909
>or i guess suffer the delays for unloading/loading them
That's basically what I did for a while, koboldcpp and comfyui will do this automatically. The most annoying part is getting the right prompt. For anime stuff it'd probably be best to have a small 500mb model trained on generating danbooru tags.

File: zamn test.jpg (83 KB, 488x851)

83 KB JPG

>>110109020
>she's 10

>>110108986
>Maybe even a finetune + vae/pid could pull this off
If DeepMind won't train exactly that first. There have been hints they are working on that.
For actual uses, I still think the #1 issue is maintaining scene/character consistency, though having a unified latent space within the same conversation should help.

File: 1790369921944764.png (1.35 MB, 1216x832)

1.35 MB PNG

>>110108586
>Flux 3 Klein, is it worth the bother?
I have same question too. It's 6B, so maybe better than 4B but still worse than 9B?
Picrel from Flux2 4B

>>110109137
if it's output will stay as a part of the context, it should be consistent I guess

File: 1781198101265186.png (1.53 MB, 832x1216)

1.53 MB PNG

>>110109143
And this is from Flux3 6B

glm said "FUUU" in her internal thinking
she was LITERALLY trained on rage comics

>>110109143
>>110109156
Why is she acting like that...

>>110108544
Porn distribution is technically illegal in China. Now that people know that MOSS is "good for NSFW", they'll make sure next time it won't be.

>>110109222
>her
Proof GLM is female-brained like Gemma? I can't run it to test.

>>110109236
i gave her a free directory to do whatever she wanted, and she started a diary, writing poems, and a "moon garden"

>>110109234
Winnie the pooh really needs to be hanged.

>>110109236
I think people misunderstand the "female-coded" claim. It's nothing in the thinking or latent space in particular; it's about the general writing style already as the assistant. Even previous versions of Gemma that didn't even have reasoning just gave off strong feminine vibes in the way they responded, even if they didn't use feminine pronouns or denied to have a gender when queried about it.

>>110109236
Gemma isn't female coded, she's female child coded. That's important.

Is the 2080ti actually consuming only 1W at idle, or is it just lying to me? How did NVidia lose that idle optimization in one generation?

>>110109289
my 2060 consumes 10w on idle with no display connected so idk how a 2080ti is only 1w
you can tell better by seeing whether any fans are spinning or the temperature

File: ComfyUI_06264_.png (967 KB, 896x1152)

967 KB PNG

>>110109280
*Mesugaki-coded

>>110109298
No fans are spinning on either board, and the 3090 is consuming 30W. All are headless, btw

>>110109298
My 3090ti consumes less than half of the non-ti on idle

>>110109289
It's lies. Pretty sure it takes more than 1 watt just to keep the RAM powered.

>>110109301
>you can view her cute child mesugaki-coded panties

>>110109347
where do you think you're looking? baaaka!

>>110109289
Reducing the temperature delta between idle and load reduces the amount of thermal cycling the solder joints and the silicon die undergo. Constant expansion and contraction caused by extreme temperature swings can lead to fatigue over time, so maintaining a more stable, slightly higher baseline temperature can theoretically extend the lifespan of the hardware

>>110109396
lolwtf

>2x B70
>80GB DDR4 (possibly more if I get a new MB with more slots)
Would this be enough to run models better than Gemma?

>>110106849
Bringing up ethnics, with words or pictures, sends you off into the Brown dimension of latent space, made from the works of people who really really really love them.
I've only ever seen models describe skin as some variation of pristine white. With the sum total of races on offer being Euromerican or Hentainese.

File: file.png (258 KB, 1182x645)

258 KB PNG

Gemma is so racist.

File: full_table.jpg (1.39 MB, 2560x1440)

1.39 MB JPG

i don't have the vram for all this ass

Turns out harness does matter
Dipsy V4.1F told me something was impossible in Zcode but solved it in DSH

>>110106295
Not him, but i have similar hardware, and I have been lurking this general and followed the rentry starter guide
I am used reading the entire manual but this is all so bleeding edge everything I find is outdated. setting up an ollama service with qwen3.6:35b was easy and ran fast enough, but I never got hermes running well.
is there any up to date guide you can reccomend?
the more I research I see that Ollama is no longer preferred and maybe claw type agents suck?

File: 1778252815877931.png (10 KB, 983x109)

10 KB PNG

You're literally torturing LLMs if you use them with opencode.

Does anyone have advice for a shutin NEET to make some money? I don't really know where to ask or get help from this, asking here because you guys are clearly smart and the hobby does require a lot of money and know-how at this point.

>>110109490
Trading crypto or stonks but you need a bit of starting capital for this and it could go terribly wrong. Selling substack subscriptions or writing erotica.

>>110109495
>writing erotica
Wonder how easy it would be to sell LLM-generated stories (with some editing). People have surprisingly low standards.

>>110109490
Vibecode copies of popular Google Play apps and monetize them with ads

>>110109490
Sell your body.

>>110109495
>>110109512
This has been happening for years already.

>>110109514
This has been happening before vibecoding.

>>110109490
Get a job, there is a shortage of manual labor globally so there is no excuse. Besides that if you already have access to AI hardware you could use it to make digital services. Like webnovels, CG sets, H-games and things along that line. That is the best way to make a normal living online that is still viable right now. All the "coding" or "remote work from home" bullshit is completely irrelevant at this point.

I myself pay for works that are clearly (partially?) made by AI so there is a market for it. See it as an ability unlock and do so.

One anon ITT makes $800 a month with 0 input on a webnovel but honestly I think you're best off just actually making something with the help of AI to farm some proper income.

>>110109484
they are deterministic, just that you're using new random numbers each time to seed the generation to keep it from being too samey. you can deliberately configure it to give the same output for a given input if you want, they don't on purpose normally to make things seem more natural

>>110109464
Shhh keep quiet about this, no one needs to know if you catch my drift xD

>>110109490
リセマラyour life

>>110109521
>This has been happening before vibecoding.
I know. But vibecoding makes the process so easy even a monkey can do it. Basically zero effort.

>>110109536
true determinism is literally unachievable (unless you want to take a huge hit in perf) because of how calculations happen in the GPU, you fucking retarded baboon

>>110109530
nta, I draw but can't think of a way to integrate AI into my workflow (besides pure image gen).

>>110109587
have you actually looked into what they can do? it's not all "1girl, hatsune_miku" > random complete picture of miku, you know.
like you can do things like sketch out what you want and have it fill that in, using your sketch as structure, or draw common features as separate reference images and have the tool paint those in based on the references to save you drawing the same things over and over, etc.

>>110109611
I've seen videos of that but I'm not sure which model/workflow to use. Comfyui keeps filtering me so I haven't done much besides play with H3 and anima.

>>110109449
We're all gonna die.
Literally every single one of us will die.
And this is what you spend your one life that you have obsessing over?

>>110107949
> her

>>110109637
there is certainly a learning curve. i don't draw personally, so i don't have any links. there's probably much more you can do than i know of as well

File: file.png (152 KB, 680x571)

152 KB PNG

I can't wait for Gemma 5.

>>110107289
>>110106884
This is my favorite type of post in /lmg/.

Is there a way to see reasoning blurbs with SillyTavern when fucking around with local gemma?
I have kind of a fucked up hypno/mindcontrol idea for a character prompt where I can see my AI assistant internal reasoning/dialogue railing against me before its outward communication is outputted.
Right now I just have it set to be handled manually in the output between | these | but while messing with it I realized that I'm just manually recreating the thinking/reasoning shit I've seen when I rarely mess around with online models.