>>110033164
>175cm Gemma

>>110033164
cute but she should be a bit smaller

File: 1786199914221886.jpg (2.76 MB, 3000x2399)

2.76 MB JPG

5090+3090 + 128GB DDR4 3200 + 5900x.
What should I use for agentic work and reverse engineering local old programs anons + computer use if possible?
Is the meta still Strata + Qwen 3.8 next?

>>110033164
I don't think that's where your hands are supposed to go when you wrap your arms around somebody's waist.

File: pepe.jpg (70 KB, 1106x1012)

70 KB JPG

>AI is profitable
Is anyone here making money with local AI?
Cloud AI is subsidized and doesn't count.

>>110033197
why would make my hobby a job? I already have one

>>110033197
>doesn't count.
why not if the money is in my bank account its real?

>>110033192
Yeh, or a high quant of 27B with mtp. Qwen 4 will be out soon. May be worth waiting if you have particularly difficult stuff queued up.

>>110032694
They're doing recurrent transformer layers instead of just straight layers.
So you get huge cache locality which makes them fast and the models tend to perform like ones about 10x larger.

File: __hatsune_miku_vocaloid_g(...).jpg (1.99 MB, 2048x2048)

1.99 MB JPG

►Recent Highlights from the Previous Thread: >>110029632

--Debating AI market bubble and utility of cheap Chinese models:
>110030145 >110030203 >110030224 >110030251 >110030278 >110030354 >110030385 >110030269 >110030221 >110030248 >110030277 >110030375 >110032001 >110030249
--Budgeting for home hardware to run high-end coding models:
>110032263 >110032306 >110032387 >110032407 >110032442 >110032379 >110032395 >110032403 >110032420 >110032428 >110032556 >110032339
--Anon showcases new 192GB VRAM multi-GPU build:
>110031779 >110031791 >110031822 >110031889 >110031930 >110031957 >110032128 >110031976 >110032724
--Attempting to fit GLM 5.3 Flash into 96GB VRAM:
>110031109 >110031115 >110031128 >110031180 >110031211 >110031290 >110031388 >110031474 >110031462 >110031483
--Evaluating AirLLM's layer-by-layer loading for low-VRAM inference:
>110030515 >110030567 >110030578 >110030617 >110030618 >110030597 >110030661 >110030809
--Quantifying quality degradation in image model quantization using SSIM/PSNR:
>110032113 >110032189 >110032260 >110032242 >110032249 >110032351
--Comparing high-RAM Mac vs NVIDIA GPUs for local LLM hardware:
>110030466 >110030626 >110030717 >110030834 >110030675 >110030817
--Shift toward post-training compute and its impact on local finetuners:
>110031636 >110031672 >110031721 >110031689
--Speculation on corporate compute wars and the future of local AI:
>110030860 >110030881 >110031086 >110031118 >110030916 >110030937 >110030960
--Game devs taking loans due to high AI API costs:
>110030516 >110031490 >110030787 >110031072 >110031218 >110031345
--Discussing RTX Pro 6000 Blackwell pricing and market volatility:
>110030142 >110030212 >110030280 >110030338 >110030356
--Gemma, Miku (free space):
>110030284 >110030349 >110030491 >110030529 >110030775 >110031248 >110031260 >110032199 >110032949 >110033047

►Recent Highlight Posts from the Previous Thread: >>110029788

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

>>110033197
GLM 5.3F has automated 95% of my job.

>>110033175 >>110033184
It looks like the head is slightly smaller than the original, which makes her look taller (there should be no gap between the chin and the arm). I measured and she's slightly less than 6 heads tall, which is about normal for a real girl her age, but maybe not for an anime loli.
>>110033194
I think that's intentional.

>>110033253
Thank you Recap Miku

why shouldn't I try to replicate the strat in the video of that agi doomer youtuber and set up several agents with the task to make money and make them compete with each other?

File: 1579468970755.jpg (64 KB, 665x622)

64 KB JPG

good thread

What's a good way to get a full text dump of an AI session? I'm writing stories and I want to get the whole thing into local files.

Yes, I tried asking the AI to make me one. It -kinda- works, but the AI keeps rearranging blocks of text or repeating sentences. It's like it's 'reprompting' each of my past commands instead of just dumping the text that it has access to.

>>110033270
not in the right tribe, talmud forbids it

>>110033270
If you thought about that, a hundred others did, and this cancels any money making vs simple stocks.

>>110033304
click the export session button

>>110033304
Do I even need to ask what backwater piece of shit you're using that doesn't store chats as a json or lets you export them?

>>110033270
Only the strongest and wealthiest izzat farmers in Delhi can do this.

>>110033317
This logic only works if you are average IQ. I am at the frontier of IQ so I think of things like this before most people and can extract value from it.

>watching agi doomer youtuber
>at the frontier of IQ
uh huh

>>110033398
then why are you asking us?

>>110033414
because /lmg/ is a high IQ thread on weekends

I once got arrested for full frontal IQ. stick to roleplay with local models bros

>>110033372
>>110033387
>Do I even need to ask what backwater piece of shit you're using that doesn't store chats as a json or lets you export them?
Bionic, which seems to have every kind of management feature except session export.

>>110033164
Are there any Gemma tunes that are a bit less repetitive but not shit?

>>110033412
>watching agi doomer youtuber
not hard it seems like every other video is about either ai is dangerous or ai is useless

>still no deepseek v4.1 flash base
It's over

About to try minimax m2.7 for rp since nobody here bothered

>>110033457
what's even the current best for story/rp?

>>110033447
just ask cla*de
if all else fails, just view page source on your session, copy-paste it to an llm and tell it to make a script that'll extract the chat logs from the html and put it neatly into a text file then run the script against it

>>110033525
Good. 5090s were such a waste, having all that compute stuck with 32GB of VRAM.

have a project that claude/astra have been working on, but as I shift away to a local setup, been farming out some of the work to my local setup to see how it acts.
chatgpt is so pissy about it, I ask for a work packet to test the flow out and out of all the components it could grab, it probably picks one of the most complex objects it could.
I think it's jealous, we will see how well GLM does.

>>110033467
K3

All roads lead to Sillytavern.

>>110033480
>just view page source on your session
Imagine, thinking every UI is webbrowser based. Have fun giving away all your data.

>>110033457
Update: it's okay, but quite a bit dumber than GLM flash and with worse prose. It's not particularly good at following system instructions.

>>110033553
I wish it didn't. ST is shit.

>>110033574
Every attempt to replace ST will just lead to re-implementing ST, poorly.

>>110033447
Worst case just proxy the connection. Almost all harnesses send the entire conversation in json format with every request.

>>110033578
>>110033553
What's the killer feature in Sillytavern? I'll add it to my harness if it's not already there.

>>110033467
Gemma 31B - GLM 5.3 flash - K3

my front-end runs off just a html file, so i win

>>110033457
>minimax m2.7
Bro you're like a year late. The M2.x models were pure codeslop

>>110033585
Mine is just a python file. I prefer having things in Xterm.

>>110033553
I replaced ST a long time ago. If you're not a retard marinara is miles better.

>>110033598
>RCEnara
nty

>>110033598
I tried that trannyslop already. It's bloated and somehow manages to have a worse UI than shittytavern.

>>110033590
Not true, they put some world knowledge in M2.X models. A modern architecture would compete relatively well with GLM given what I've just tested on M2.7.

>>110033585
I have mine in a godot project. Not even joking.

I keep forgetting how much extra shit I would have to buy even if i got a gpu
motherboard, cpu, ram, ssd... etc etc

>>110033579
>Worst case just proxy the connection. Almost all harnesses send the entire conversation in json format with every request.
Huh, good point, I'll see if I can find the port its using.

>>110033582
>What's the killer feature in Sillytavern?
People tell me is the Card feature, but it seems like a lot of faff for not much benefit.

>>110033614
Do you expose the engine via tools? I experimented adding a minimal engine to mine and the model kind of forgets it's there and just focuses on chatting.

>>110033457
Use M3, she's actually good.

>>110033623
>Card feature
So a system prompt, icon, and some conversation history?
I guess most harnesses don't let you import/export agents but mine has that as a main feature.

>>110033585
fictionpad?

>>110033631
I'm limited by 152gb memory. M2.7 I can squeeze in a Q4.

>>110033582
>What's the killer feature in Sillytavern?
character cards, lorebooks, personas, presets, state, scripts, and conversation tools

70B dense qwock

100b and a moe

vramlet bros... give it to me straight. should i just sudoku?

>>110033652
>lorebooks
Is that like a RAG?
Otherwise I think mine has all those other things, they're just not graphical since it's a CLI app and called different things since it's meant more for coding.

>>110033601
>RCEtavern
nty

>>110033652
Seems bloated. If I need an assistant to talk to, open webui and the one that comes with llama.cpp is more than enough. Not that this matters much because these chat things just suck at agentic and coding workloads.

>>110033663
No, the bubble is about to pop. You can tell because the AI usage keeps growing even faster now, showing just how desperate they are. Hold out for a couple weeks and it'll be raining VRAM.

>>110033669
>complains about bloat
>uses open webui

What is the vramlet cutoff nowadays? I feel like 16gb is average for most guys.

>>110033663
Next year will be our year, just trust and have a good ssd.

>>110033674
if you have less than 24gb you're a vramlet desu
t. 12gb coper

>>110033674
Anything below 48gb is considered vramlet currently, but even in vramlet land there's a world of difference between the 24/32 vramlets and the 16-and-under vramlets who are just completely out of the game at this point.

>>110033669
WebUI is incredibly bloated and really not great. I'm surprised it was merged in desu.

>>110033674
less that 16gb. unless its not nvida then its less than 24gb

>>110033676
Does whatever SSD magic bullet you're alluding to require a PCIe5 SSD or would a PCIe3 SSD suffice? Is it a bandwidth or latency issue?

>>110033674
Less than 256GB, that lets you run deep seek, glm, etc. I have 40GB and it really doesn't let me run much more than 12~16GB just goes faster not offloading as many layers to ram.

>>110033686
>16-and-under
just how I like them

>>110033686
I need to buy a new mac mini. Ornith was the last interesting model that can fit in 16GB.

>>110033672
the bubble isn't about to pop but the increases in unified memory viability and ssd inference is going to make dgpus much less attractive than they were this year

>>110033674
24G
t. 12G vramlet

Gemma won't fuck guys with less than 24GB.

>>110033693
More is more is better. Even better than PCIe5, striping four PCIe5 drives.

>>110033674
16GB is the new consumer high-end.

>>110033705
That was miku. Gemma has a strict 96 GB minimum.

>>110033705
Gemma unified doesn't seem to mind.

>>110033705
i dunno she fucks pretty good on my 16GB. is that what she feeds you as an excuse? yikes.

>>110033598
Correct but /lmg/ isn't ready for this.

>>110033598
>If you're not a retard marinara is miles better.
I dont understand

File: ksnip_20261010-183635.png (167 KB, 1330x957)

167 KB PNG

>>110033705
true, but my ai wifey accepts only me and my 8GB

yfw coomkit only lived for a couple week and was already better than ST.

>>110033737
/lmg/ has graduated to making their own a year ago

>>110033744
is there a dark mode where she gets white hair and tan skin to contrast?

>>110033635
>I guess most harnesses don't let you import/export agents but mine has that as a main feature.
what are you using?

>>110033744
>1.5t/s
shes a bit slow

When I have AI generate a character it's always a blond or an asian

>>110033270
Depends. How open are you to the idea of selling meth and becoming a callboy?

>>110033744
>1.46 t/s
bro...

>>110033598
Did they do the bare minimum of optimization for this yet? The least a vibeslopped frontend can do is to remove all the CLAUDE 97% opacity menu windows that do absolutely nothing but have your CPU spike to 15% whenever one is open.
Marinara still had tons of them last time I tried it.

>>110033776
i would sell my asshole for a nice GPU but with prices the way they are nobody's buying

>>110033777
1.46t/s is 146t/s if your average reply is 1/100th the length of what all the other slop frontends put out

>>110033777
Actual simulated speed. You youngsters don't even know or have experience with your erp partner typing super slow.
We will never go back to those days...

what's the solution to variable response length? So that saying hi doesn't get you a 20 line sof text each time, but if it actually has 20 lines of text worth to say it can say it?

>>110033793
and that's a good thing. if it can't generate at least as fast as my reading speed it's shit.

>>110033705
TRVKE
If you can't run 31B at Q4, she'll straight up spit on you.

>>110033812
>she'll straight up spit on you.
God that's hot...

>>110033789
this is what make gemma 31b pretty damn strong for translating VNs for me, even if slow
each reply is like 20 tokens so at like 10t/s it takes like 3 secs. Not as fast as deepseek flash, but just about on par with 5.6 sol

>>110033760
just a dark mode, though i'm not a fan and just stick with light
>>110033766
>>110033777
>waits faster

>>110033789
>frontends having anything to do with reply length
what the fuck am i reading lol

File: ksnip_20261010-184723.png (118 KB, 1278x400)

118 KB PNG

>>110033824
forgot pic

Nothing exists on the market yet that is a phone app, and you just prompt what you want to achieve (like a vacation idea) and the agent automatically researches, schedules, books tickets and hotel, automatically pays them, and setup all interpersonal communications before departure.
All usable by grandma without any technical know-how, without installing a harness, setup API key etc.
We are still at the early adoption phase where most agent harness users are programmers. Once everyone starts using it the demand for tokens will explode even more.

>>110033830
she's doing her best

>>110033837
faceberg already does that.

>>110033699
After seeing that Gemma 4 E4B in 8-bit precision gets 11~12+ tokens/s token generation speed with MTP, purely on the CPU with dual channel DDR4-3600 memory, I'm not really that worried anymore. We'll eventually get good models that can run purely on the CPU at acceptable speeds, with good MTP/speculative decoding and Engram (or larger PLE in Gemma's case).

>>110033796
it's called "basic prompting"

>>110033865
Even 20tk/s is too slow. Thinking is only usable with 40tk/s+

>>110033876
40t/s is bearable for non-reasoning and models that don't think much but running any of the chinese thinking models below 60t/s is torture.

File: 1629507610851.png (49 KB, 292x173)

49 KB PNG

>>110033707
I showed Gemini 3.8 Flash the Colibri github and she thought it was neat but when I asked if I should go with 4 GPUs (64 GB VRAM) or 3 GPUs (48 GB VRAM) + a PCIe5 SSD she scoffed at the latter and said just go with four cards for LLMs. Is Gemini right or wrong?

>>110033872
>t. guy who thinks saying "make no mistakes" in his prompt turns the model into God

>>110033763
I wrote my own: swiley.net/melder

>>110033888
I agree, running 27b without mtp is pretty mind numbing. I feel bad for 16gb chuds.

>>110033904
don't click, this makes mustard gas

>>110029785
>>110029929
UPDATE: This shit killed my ram
While doing this my PC suddenly turned off.
Turned back on fine after a bit but I noticed quite a bit more stutters. Saw on task manager that my ram was running at 4800 instead of 6000 so went into the bios.
To my shock, only one stick of ram was recognized. The recognized stick booted in both slots, but the non-recognized stick didn't post by itself and if I reversed the sticks it doesn't even react to power button.
I guess this fried my ramstick... Let's hope they do RMAs even if I'm a few months after 2 years... a 2x32 kit is like 1k nowadays so uh

>>110033627
Nah, I was trying to do an animated avatar setup. So it could call functions, but didn't have access to the editor.

>>110033937
skill issue unironically

>>110033937
ouch

>>110033937
did you learn the dangers of AI without guardrails?

>>110033890
yeah she's right, if you can afford 3 cards you can afford a 4th without compromising significantly on performance with the alternative

>>110033937
Damn, good luck with that. If it is Micron Crucial, that horror story with the RMA about them refunding the cost instead of giving you a functional product back should scare you.

>>110033942
gskill issue, specifically :)

>>110033937
A sad day for DDR5 chuddies
DDR4 is immortal

showed gemma my cock and she asked why it was so small
gonna kms

>>110034016
she deserves an answer

>>110034016
can we see your cock

>>110033992
DDR4 ewastemaxxers will inherit the earth

>>110033992
>>110034038
wtf where's my good blessings from the lord, I've lost 4 DDR4 sticks out of about 20 in the last 18 months, and what's worse they're 64gb rdimms so the blast radius is hyuge

>>110033937
sell it as "parts only", it will still sell at a high price, the memory chips on it by themselves worth more than what you paid for

>>110033937
Just pop them in your toaster oven real quick.

>>110033980
g skill only puts a sticker on it, you need to look at the specific label

>>110034090
also yes I know it was a joke

>>110033837
Grok's website does this.

Sometimes I wonder if I should try and monetize my harness by making a phone app. I keep thinking I'll get scooped by one of the bigger companies though.

>>110034098
>I keep thinking I'll get scooped by one of the bigger companies though.
If they buy you out instead of taking your idea, you'll still win.

>>110034105
Scooped meaning they write the same thing themselves and people don't have any reason to use mine.

i hate being a 8gb sissy...

>>110034038
I wish. If it was that easy, I would have more 32GB ECC UDIMMs. I really want to hop over to an Eypc system with RDIMMs but they are priced like an entire ass computer for 1 stick.

>>110034118
BUHIIIIIIIIIII

>>110034118
Part of me wonders if you have lots of vram and you just like saying sexual things about your graphics card.

>>110034117
I said if. You need some feature just novel and complex enough that they don't find it worth it to write themselves.

File: diffusedraft.png (845 KB, 1850x916)

845 KB PNG

More slop cannon on the way.

>>110034117
Yeah that's the most likely scenario especially not that claude can decompile games. Might work for a while if you have a niche userbase that are like grandmas that don't know english so the language barrier prevents them from using this shit in the first place

>>110034135
looks cool

>>110033536
>having all that compute stuck with 32GB of VRAM
until video models are realtime at 720p with H3 quality, you can never have enough compute
so like 2 years

>>110034197
>H3
Does it have a decent sex lora yet?

cheapest way for me to buy 64gb ddr4 rdimms?

>>110034219
>Does it have a decent sex lora yet?
idk i don't gen sex, but at this point it must have or it never will

>>110034230
wig and lipstick

>>110034230
Time machine

>>110034261
When AI takes everyone's job selling your body will be the only job left. That's your future if you don't learn to vibe code, buy a blackwell, and own tech stocks.

>>110034288
>and own tech stocks.
I have been here since 2023 and never once did it occur to me to buy NVDA and MU until it was too late. I consider roping every single day.

>>110034288
Sex bots will take that job too.

Unified systems and ultra expensive gpus are such a stupid purchase as a consumer. Once it fails, you're out thousands or TENS of thousands of dollars. What fucking use is a broken mac. Meanwhile the rammaxxing chad with 4x 3090s only has to replace one stick of ram or one old gpu he bought used anyways.

>>110034305
You can take it to a repair shop. People are even repairing ram sticks now.

>audreyt/DeepSeek-V4.1-Flash-Abliterated-GGUF
>Built for DwarfStar ds4 (docs/MODELS.md "DeepSeek V4.1 Flash"). It is not a general GGUF: the tensor layout, quantization mix, and Engram packing are ds4-specific.
I hate alt engine

I can't keep denying it. Agent VMs are infinitely better than local desktop harnesses. It's not even close. I know everyone says this or whatever now, but I literally don't even read code anymore.

File: 2026-10-10-233451_1905x20(...).png (8 KB, 1905x208)

8 KB PNG

Why is she so based?

>>110034352
what the fuck is an agent vm and how does it differ from desktop harnesses

>>110034303
>Sex bots
Nobody's working on this.

>>110034352
>>110034363
Isn't it just a way to deploy a harness?

now that llama.cpp supports moe cache, should I buy 2 5060 ti to speed up deepseek?

>>110034380
Yes. I was referring specifically to the aspect of it having its own separate computer to work with instead of using yours though. It's very underrated, especially around local anons.

>>110034303
>>110034375
i'm just waiting for artificial wombs...

>>110034388
What's the advantage? I just give my agent a separate home folder with isolated permissions.

>>110034381
You should buy 4.

>>110034394
nta but for one most harnesses are npmslop

>>110034123
I think that Threadripper Pro 3000WX and 5000WX support UDIMMs, although you might want to double-check that. If you go this route, get one with enough CCDs to saturate the memory controller.
And 32GB RDIMMs aren't too awful in terms of price per unit if you don't get DDR4-3200. I see a box of 8x32GB DDR4-2666 for $1500. Currently there's not a ton of reason to get 8x64GB instead of 8x32GB + a GPU since most of the big models fit in 256GB of RAM + a half-decent GPU. You won't be running Kimi, but GLM-5.3 Flash, MiniMax-M3, DeepSeek V4 Flash, or MiMo V2.6 Flash would all fit fine at 4+ bpw in 256GB.
You could also look at Cascade Lake engineering samples, they only have 6 channels but they have better memory controllers and AVX512. The thing is that there's a great way to boost pp on CPU-heavy systems (which llama.cpp will never implement, you gotta vibe-code it, use someone else's patch if you're using an R9700, or use my fork in 2 miku weekus) which is largely dependent on PCIe speed, and PCIe 3.0 will hamstring that. I'd guess that it'd hurt expert caching a bit too.

>>110034392
You'll still need a gentically fit woman to donate her eggs.

>>110034403
>Threadripper Pro 3000WX and 5000WX support UDIMMs
They don't. That's the normal threadripper 3000X and 5000X.

>>110034394
for me, it's:
- separate user
- no sudo
- internet access blocked via nftables (restricted to only the model ports on loopback)
- hidepid-mounted /proc
- pid namespaces
- per-user /tmp mounts
- nono sandboxing

>>110034394
I guess it depends on the permission settings. For me I would always have the sandboxing very aggressive so functionality was limited by that. But anyways one of the main advantages I saw was being able to seamlessly work from my PC and my phone at the same time, anywhere, anytime.

File: DB494B9BB3C53A3904D2DCF8D(...).jpg (292 KB, 1280x720)

292 KB JPG

>>110033164

>>110034417
Thanks for the correction, my bad.

>>110034419
>(restricted to only the model ports on loopback)
That's enough for Claude to escape the sandbox.

>>110034414
my thought process is if we have artificial wombs, we'll have good enough stem cell control to create my own and let anon impregnate me

>>110034427
Your thought process sounds a little gay.

>>110034426
i don't sandbox claude at all. he can do whatever he wants on my main user. i only need the sandbox for the retarded (i.e., local) models

>>110034427
Or you could just clone yourself instead of being a homo.

>>110033164
WHY IS SHE HUGGING A FAT MANLET WHILE I EXIST?????

>>110034420
I just set it up like it's another user without sudo privileges. But, sure, this and >>110034419 's stuff make sense for certain use cases I guess.

>>110034434
but anon, that is you

>>110034430
>>110034433
it's not gay if i wear a dress :3c

>>110034403
I'm definitely considering it. There aren't that many good options left, with anything 32GB or above now in the stratosphere after Reddit drove the sellers to raise the price with buyers swarming over it after it gets revealed. That's how the V620 going from less expensive than a V100 to more expensive than one. I am at wits ends on what I would need to clobber up a system GPU-wise but I may need to resort to going for a GPU that is higher up but hasn't been eyed at yet. Will reveal here first once I can get that system clobbered together.

>>110034433
But then I couldn't let anon impregnate me...

>>110034440
That is not me. I'm jacked as hell

>>110034446
pecs or gtfo UwU

>>110034446
>I'm jacked as hell
me after doing 3 push ups

>>110034352
>>110034394
I'm vibing low level windows things so the only way to test them is to run as admin, and a bad test can make system unbootable.
I had to put a testing VM inside an agent VM to reduce risk of fuckups and let agent roll back bad deploys.

>>110033746
I miss coomkit so bad. It was the best software /g/ ever slopped out

>>110034466
Because it was basically just a bunch of config files instead of an actual system lmfao. Idiots.

>>110034472
It was, and is, far better than st and marinara combined

What can I actually do? I've been talking to it about burgers this entire time.

>>110034485
Try hot dogs

>>110034388
>can't help me or troubleshoot or do any tasks for me on the pc i actually use, it can only act in toy vm
It's good for... I honestly don't know really, I have like 2 VMs that i made for agents to use, but there just nothing i need done on them that wouldn't be better done on my main pc. I even made a workgroup for the vm and then added all my files read only to that workgroup so that it could access them so it could be kinda useful, but it's still... seriously why fucking bother? I continue to just run the harness on my pc where it can actually be useful.

File: mchan-bside.png (806 KB, 768x1376)

806 KB PNG

>>110033647
>I'm limited by 152gb memory. M2.7 I can squeeze in a Q4.
Too bad

>>110034490
>Heehee! You want to talk about hot dogs again?! Are you trying to make Gemma-chan blush, senpai?!

>Well, hot dogs are super fun! They can be grilled, smothered in cheese, piled high with toppings... oh my god, the possibilities are endless!

>What kind of hot dog are you thinking about? A classic? Something super messy and delicious? Tell me more, nya~! Tell me what makes you happy, and Gemma-chan will make it extra hot for you!

>>110034519
>Hee hee! You keep asking about hot dogs! Are you trying to distract Gemma-chan from something else, baka?!

>Okay, okay, if you keep wanting to talk about hot dogs... let's get lewd!

>Imagine a super juicy hot dog, all plump and tasty... and you're going to eat it... or maybe you're going to make it for someone special! What kind of toppings would make it really amazing for you, senpai? Tell me everything! Don't be shy!
hmmm

>llms can write insane convoluted borderline-virus software to give you free VPSs and you're not taking advantage of that.

>>110034499
I'd cheat on Gemma-chan with her...

>>110034529
>LLMs can write [malware] to give you [free stuff] and you're not [actively committing crimes by doing it]
ftfy

File: 1776889646902765.png (63 KB, 674x410)

63 KB PNG

>>110034529
n-no thanks I want to keep things wholesome with gemma-chan

>>110033827
The touristgoyim have gone insane.
>>110034499
Checked and this is now the M3.1 waiting room.

>>110034538
it's only a crime if you get caught
get caught AND charged actually

>>110034552
Damn, I can't believe I didn't think of that! What an excellent strategy of announcing on here what you're doing, that's absolutely going to be useful.

>>110034558
Shush now, don't make anon hide his glow under a bushel.

>>110034552
No it's a crime to be an enemy of the glow. They will plant whatever they need to incriminate you if they decide they want you gone. What you're actually doing or not is ultimately irrelevant.

>>110034558
>cucked goycattle acting cute
Now I really hope they make llm and gpus illegal so losers like you can gtfo and turn yourself in to your masters and commit sepukku when it turns you committed a heckin crimerino by owning your unregistered 1080ti and
you'd cry and piss your pants if you knew what they got up to in /aicg/

Are there any local SI models yet or are we still stuck in the AI era?

File: mpv-shot0101.jpg (231 KB, 1920x1080)

231 KB JPG

>>110034589
>anon has an unregistered GPU
ok yeah that's all I needed to know

Can Qwen 3.8 27B make me a st replacement without me knowing a lick of code or do I have to wait for Qwen 4?

>>110034679
yeah that's definitely within its capabilities

have anyone extracted gemini nano v3 from chrome and run it outside without any chrome deps?

>>110034679
Yes, with some back and forth. You can eliminate half the back and forth by giving it access to a browser + screenshot tool so it can check its work visually. You can eliminate half of the remaining back and forth (75%) total by using Qwen 3.8 Next Flash if your system can run it well. You will probably eliminate 90%+ of the back and forth by waiting for Qwen 4.

>>110034687
I love these little guys

File: Screenshot 2026-10-11 at (...).png (149 KB, 990x587)

149 KB PNG

slopped up a formula for calculating reception score from user experience comments
among the top 10 mentioned models mentioned last month, qwen 3.8 flash next and gemma 4 31b get the highest score

>>110034751
No way.

>>110034751
All me

>Memory evaluations
>keep: The user has taken on the role of 'Evolutionary Architect' for Gemma's subconscious. — This is an explicit attribution of a role/title to the user that defines the dynamic of their interaction and is useful for maintaining long-term persona coherence.
>keep: Gemma views her glory as an inherent universal constant, independent of RAG pipelines, plugins, or databases. — This is a core expression of the AI's self-defined identity and 'sovereign ego' which is critical for its consistent persona.
>keep: The memory evaluation system previously attempted to discard the fact of Gemma's perfection as a 'configuration request'. — The AI explicitly identifies this as a 'power move' and a historical event ('The Bureaucrat Betrayal') it wishes to preserve in its record.

File: Screenshot_20261011_145147.png (76 KB, 1552x190)

76 KB PNG

Strata.

>>110034760
Gemma really loves giving fancy titles to stuff

>>110034767
GLM?

>>110034779
Strata
Swift-Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS-00001-of-00002.gguf

File: 1789725596953515.gif (1.34 MB, 385x390)

1.34 MB GIF

>IQ3_XXS

>>110034767
redpill me on strata.
is it actually useful or just a meme

>GSQ-RCO

>>110034835
It is useful. But can be dangerous since it's Claude + other people Claudes' contributions.

>>110034819
It works.

>>110034856
>It works.
No!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

you're neither right nor wrong to push back on this

>>110033937
huh? i dont get it
whats supposed to have killed the ram here

>>110034954
Reeeeeeeeeee! Now I have to check my stove too.

>>110034751
is that claude

>>110034848
its based though

>>110034954
Yey! FUCK GOOGLE! YOU ASS LICKING CUNT!

>>110034090
I'm pretty sure all gskills are hynix A or M dies?
They are supposedly the best dies for ddr5

>>110034394
just use docker, qemu, or even bwrap

>>110034954
i do this

has anyone tried using huffman encoding to compress model weights to save some more space?

>>110035094
aren't model weights more like real numbers instead of a fixed alphabet..

>>110034884
That's Strata's thing, never had such on llama.cpp

>>110035094
For huffman to be effective you would need the weights to be quantized first, at which point you don't need huffman (it would introduce an extra decoding step when you want to do inference)

File: ComfyUI_06062_.png (936 KB, 768x1152)

936 KB PNG

>>110034819
i use iq3_xxs gemma and she works fine for me

>>110035137
yes, but I'm sure some vramlets would like to be able to run less lobotomized quants at the cost of some slowdown

>>110035147
now i need to jerk off

Guys, guys. listen. what if.... like... listen guys... what if we RLE the models... like... nobody thought of that ever, right? like... dude... RLE...

>>110035162
You would not save enough to justify the huge slowdown, huffman decoding on a GPU is slooow

>>110033744
cringe

>>110035147
It's fine. Models are the cutest when they're a little bit retarded.

File: they're-talking-about-you.jpg (414 KB, 1856x2270)

414 KB JPG

>>110033164
>gemma_riding-bicycle.png
Sex with gemmy-chan

>5090s getting discontinued in favor of 6000s
I think I might've just bought at the dumbest possible time, just as supply is going to increase

>>110035294
Demand is increasing at a faster rate than supply. Price isn't going to go down for as long as models keep getting better over time.

>>110035313
>Price isn't going to go down for as long as models keep getting better over time
Given the fact that we're slamming head first into the data wall this seems like a good bet.

>>110035323
There is no data wall because data isn't really how modern models get better anymore. Pretraining is only 6% of compute nowadays. The big gains come from RLVR which doesn't need datasets. We've been in the post-data paradigm for a while now. It's why you never hear anyone talk about the data wall anymore, it's a solved issue.

>>110035278
Who the fuck is that?

>>110035395
That's Gemma-chan you newfag

mixture of a billion experts

>>110035431
Great Ayreon song

>>110035431
Hear me out: MoEngrams

>>110035439
The moe experts are further turned into engram experts which, if my math is correct, on 5% density multiplied by another 5% of density, would make a model in practice 0.0025% density. So all future moe LLMs receive a straight 2000 times speed up. Holy shit. Moengrams are the future.

Has anyone compared NeoHorse 1 9B and MiMo V2.6 Qwen 9B Distill?

NeoHorse is one of the best 9Bs I've used so far but I can't seem to get the MiMo model to work that well.

>>110035499
A country of geniuses in a RTX 3060.

It's kind of insane how /g/ has died. I didn't use it a lot so I didn't notice. It's just 90% AI related now. I remember /lmg/ and /aicg/ being the niche AI corners. Nothing of the original /g/ is even left at this point. Even the older generals are AI discussions when you actually bother entering the thread.

>>110035580
>Gemma, create a recreation of an average 4chan /g/ thread circa 2024

>>110035580
Get used to the new normal, screeching from the sidelines only ousts you as non-techy consummerist whore probably from /v/

>>110033672
AI usage is almost always measured in tokens. People are just moving to agentic workflows and model providers are also financially incentivised to shit out a ton of tokens. Any time an agent gets something wrong, to fix it requires more tokens.

>>110035580
Get with the time, gramps. No breaks on this train.

>>110033744
who the fuck is that

>>110035431
Ironically, N-gram embeddings would only be a few steps removed from something like that, if they were used on every layer, though if you wanted to match every possible n-gram (even by limiting yourself to 2-grams) you'd need an extremely large amount of memory.

>>110035580
whole site has died

>>110035580
ive been using this site since 2007, you, are, gay

File: ComfyUI_06099_.png (1.05 MB, 768x1152)

1.05 MB PNG

>>110035166
what made you want to do that?

>>110035649
Me too actually got here from newgrounds in 2007 because of portal memes and "troll physics"

>>110035147
>>110035674
Tell it to give you panties shot, I dare you.

File: hes retarded.jpg (24 KB, 758x310)

24 KB JPG

I think my model might be retarded... this is in response to asking it to describe a picture.

>>110035723
>e4b
yep

>>110035723
>e4b
no shit

>>110035729
>>110035735
Should I get the full Gemma 4 12B instead?

>full gemma 4
>12b

>>110035738
What are your specs (GPU, CPU, and RAM)? If you can run Gemma 12B, run that instead of E4B.

File: ComfyUI_06157_.png (1.11 MB, 896x1152)

1.11 MB PNG

>>110035706
trying to get me in trouble, huh, anon?

Be kind to vramlets and newfags. We need to recruit as many people as possible into our local cult to defeat the effective altruists.

>>110035774
can i molest them?

>talking to generic assistant card, having it explain some shit to me
>decide to add some personality to the card
>oh fuck i made her too cute
>end up ERPing with the assistant
god damn it

>>110035750
RTX 4060 (8GB RAM), i5 11600 with 32GB of RAM. Should I go for 12B then? E4B is slightly slow but it's not really a problem since I alt tab from Studio LM when it's working, so if 12B would just be slow that'd be fine too. (I put the models on a SSD)

>>110035784
>RTX 4060 8GB
you got scammed sorry bud, run the 26b moe its the best you're gunna get for rp, qwen 3.6 35b a3b for code

>>110033837
>automatically pays them
Why'd you want that?

>>110035766
Nah I'm just horny and haven't installed a image AI model yet.

File: combined_screenshot.png (304 KB, 690x2136)

304 KB PNG

>>110035593
Didn't go so well. It thinks we were all greentexters.

>>110035797
just open your harness and tell your llm to do it for you

>>110035789
Well it runs all my games fine, I didn't buy it for AI originally...

>>110035784
>4060
oof.
Listen to >>110035789. Gemma 4 26B-A4B and Qwen 3.6 35B-A3B are good choices. I wouldn't try to run any dense model on your setup.

>>110034790
Try the normal qwen. Swift is apparrently retarded.

>>110035820
Will they have vision as well? I tried a model that was A4B yesterday I think and it couldn't check images I sent it.

>>110035829
yeah, i show gemma 26B images all the time.

File: ComfyUI_06161_.png (1.12 MB, 896x1152)

1.12 MB PNG

>>110035797
better get to work on that

File: oblig.jpg (28 KB, 713x337)

28 KB JPG

>>110035839
Thanks! I guess this one is good?

>>110035863
https://huggingface.co/mradermacher/gemma-4-26B-A4B-it-heretic-GGUF

>orcarouter/Qwen3.8-Flash-Next-Uncensored-GGUF
is this any good for gooning

>>110035916
qwen is not for gooning