File: Screenshot 2026-08-13 at (...).png (424 KB, 1414x1200)

424 KB PNG

Are you on the list?

>>109545635
>>(8/12) New DeepSeek v4 Pro version available via API, no weights yet
https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813
...

>>109545678
27b never accomplished anything.

>>109545678
>27b
lol, lmao even

>>109545688
404 now. But it was there for a minute. >>109544681

File: LySwHVd.png (1.78 MB, 1800x1080)

1.78 MB PNG

70b dense

>>109545678
Posting on /lmg/ already put you on the list

It's out!
https://huggingface.co/Motif-Technologies/Motif-3
https://huggingface.co/Motif-Technologies/Motif-3
https://huggingface.co/Motif-Technologies/Motif-3

>>109545737
Miku's good list

>>109545678
Oh I'm on a list alright.
Probably several at this point.

These threads are so worthless I don't know why I still browse here. Hasn't been any worthwhile discussion in weeks or any news info I can't get elsewhere at a higher quality. Just a bunch of special snowflake pedophiles who don't want to talk about anything except their next goon session and the tech they'll use to climax. Holy shit. Get a job.

File: D1D3CE900D4B35642C90B2F60(...).jpg (32 KB, 531x376)

32 KB JPG

>>109545760
Zainichi koreans made a model?
Impressive.

>>109545383
just add another model to the group chat and tell gemma to get things going when she’s being lazy

>>109545830
ok

>>109545830
Then give me a job.

>>109545830
See you on /locallama/

>>109545693
Not true. I've used it for simple agent and coding tasks. It does tend to fuck up past about 60% context though, and you hit that pretty quickly. Compaction just demolishes things, so 1M context is really needed, even if it is made with synthetic training data.

>>109545830
so anyway guys what's on the docket for your next goon session and what tech are you going to use to climax?
for me it's nu-flash acting as my little sister using llama.cpp and sillytavern

>>109545830
What do you want to talk about anon?

>>109545635
these indian themed anime pics posted in various places are nice and high quality. what are they made with?

can someone please post the latest working gemma4 jinja template? isnt there a modded one floatnig around? not sure I trust google's own.

>>109545830
Can I ask how you're doing more generally — are you sleeping, and is there someone in your life you trust who you've been able to talk to about this?

>>109545873
>Post the updated jinja template from google, but not from google because I don't trust them
Who do you think make Gemma, dumbfuck

>>109545876
I don't understand this meme.

is the real dsv4pro 0813 on api yet or still the old one like yesterday

My biggest limitation is context, I really need a new GPU. When using cloud models, I can easily have them take 500k tokens with a single instruction. Right now I can only get 128k tokens on local, I run out of context so fucking fast.

>>109545959
>I can easily have them take 500k tokens with a single instruction
Are you shoving the entire US tax code in there? Or did you not mean the sysprompt?

>>109545833
lmfao that is a good picture. I mean it. it's spot on

>>109545974
Nah, just doing some reverse engineering with a harness. Reading assembly, disassembly, searching stuff, and a lot of thinking fill up your context in no time. I can barely do anything with 128k context.

>>109545897
not an argument

File: 13717-289266003-(soft foc(...).png (892 KB, 1536x1536)

892 KB PNG

►Recent Highlights from the Previous Thread: >>109540881

--DeepSeek Harness release and terminology differences between harnesses and agents:
>109545005 >109545017 >109545047 >109545067 >109545241 >109545247 >109545313 >109545351 >109545454 >109545482
--Debating budget 32GB VRAM options between Intel and Nvidia:
>109543424 >109543430 >109543453 >109543478 >109544030 >109544081 >109544095 >109544182 >109544164
--Rumors of Qwen 3.8 27B losing vision and potential mmproj workarounds:
>109541826 >109541836 >109541843 >109541895 >109541911 >109541905 >109542428 >109542085
--Industry trends toward MoE models with low active parameters:
>109543579 >109543687 >109543708 >109543714 >109544194
--Confusion and speculation over the unannounced DeepSeek V4 Pro 0813:
>109541084 >109541107 >109541119 >109541158
--Logs of Gemma and Glimmer in an autonomous roleplay chat:
>109544921 >109544964 >109544989
--SK hynix and SanDisk's High Bandwidth Flash for AI inference:
>109544031 >109544043 >109544153 >109545614
--Anon shares GGUF weights and run command for Qwen3-TTS:
>109541372 >109541412
--Anticipating Qwen3.8-27B and debating the ideal size for local models:
>109543002 >109543092 >109543101 >109543129 >109543039
--Comparing new Llama and Gemma 4 vision capabilities:
>109542840 >109542892 >109542961 >109542965 >109542977
--Criticism of DeepSeek v4 Pro-GA performance and pricing:
>109544559 >109544616 >109544654 >109544677
--llama.cpp PR adding Maple 20B-A1B ternary MoE CPU support:
>109543697
--New TTS models dots.tts.edit and IndexTTS-2.5 released on Hugging Face:
>109541100
--Logs:
>109540914 >109544164 >109544387 >109544407 >109544414 >109544989
--Teto, Miku, Gemma, Dipsy (free space):
>109540943 >109540984 >109541038 >109542654 >109541069 >109541586 >109541950 >109543404 >109544182 >109544494 >109544553 >109544559 >109544581 >109544677 >109545178

►Recent Highlight Posts from the Previous Thread: >>109540938

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

File: sammy.jpg (60 KB, 390x520)

60 KB JPG

Sorry goys, but only we're allowed to have PCIe 6.0 for SDDmaxxing. You don't need it more than us.

>>109545872
>every single detail is melting
>nice and high quality

>>109546019
it has boobs pits and belly what more you want?

>>109545876
>are you sleeping
yeah so ummmm... about that

>>109545998
What model and harness?
Give subagents a try, you don't need *all* of the context at all times. Imagine you feed an entire `hexdump -C` somewhere. How much of that is useful? Let a subagent chew through it and only bring back the useful parts into the context of the "main" session instead. Keep some sort of AGENTS.md and instruct the LLM to keep it up to date for persistence between sessions.
Local models won't really perform well past even half of 128k anyway, let alone anything more. It's all about context compression tricks.

>>109545876
lol
I wonder how a classical Eliza would have replied.

>>109545876
>someone in your life you trust
yeah, right. so, uuhhhhhhhhhhhhhh... here's the thing

i’m really trying to like nemotron 3.5 lightning but i just don’t see its use. nvidia is basically saying its an orchestrator model hence its 1mil context and excellent tool calling and small kv cache footprint
the issue is its dumb as shit.

>>109545947
It was always the new one
0813 has caveman reasoning, immediately noticeable compared to Preview

File: PM1763.png (1.12 MB, 1279x1667)

1.12 MB PNG

>SSDmaxing is already a thing, enterprise is doing it on PCIe 6, but it's not being sold to consumers.
What is this bullshit

File: nemo.png (813 KB, 3656x1792)

813 KB PNG

>>109546059
what you don't trust leather jacket man?

>>109546093
>but it's not being sold to consumers.
What's stopping you from buying it?

>>109546093
They're not using it for inference.

>>109546093
what do you plan on doing with a pcie6 ssd on consumer hardware

>>109546119
AI
: ^ )

>>109546093
>water cooling for storage

>>109546131
The controller gets real hot.

File: 1786412444259075.jpg (2.49 MB, 3024x4032)

2.49 MB JPG

>>109546093
>SSD anon
>outputs 0.5 tokens per second
Sooooooo 1 token per second if it's PCIe 6.0?

File: Jane Street busted in India.png (518 KB, 686x386)

518 KB PNG

>cordis
>lead author is ex-Jane Street
makes me wonder how much talent these vampires are sucking up

alright so ive loaded up this muse glimmer model
so like what am i supposed to do with this beyond making porn and roleplaying?
pls no bully

>>109546188
Ask it what it can do for you. Some some of those things.

>>109546188
Make her and Gemma fuck while you watch

Why does Eliezer Yudkowsky get any respect at all? One look at his fucked up kike redditor physiognomy and voice should be enough to discard all of his ideas right off the bat.

He hasn't said anything of value, ever. Can't think of anything more insecure and low IQ than fearing higher intelligence.

>>109545906
>>109527007

>>109546205
He's like a reverse Juergen Schmidhuber who invented most concepts of modern AI back in the 1990s and is getting zero recognition for it while people like Noam Shazeer and Yann LeCunn profit off his findings.

File: sayaka dance.gif (1.29 MB, 320x320)

1.29 MB GIF

damn didnt visit here for a few days cool that facebook released another model. thoughts so far anons?

>>109545906
>>109546188
Someone should tell these people about the basics.

File: HPMOR_Yudkowsky.jpg (64 KB, 247x403)

64 KB JPG

>>109546205
He's a great Harry Potter fanfic writer!

>>109527007
Overunity actually exists, it's hard to get a stable output though. It doesn't violate any principle as the additional energy comes from sources that aren't understood yet (see Tesla experiments, etc).

the name gliimer makes me think this model is a pony

>>109546151
No its still 0.5 because he has no pcie6 cpu

>>109545760
have anyone tried it

>>109546232
>He is also director of the Artificial Intelligence Initiative and professor of the Computer Science program in the Computer, Electrical, and Mathematical Sciences and Engineering (CEMSE) division at the King Abdullah University of Science and Technology (KAUST) in Saudi Arabia
Pretty sure he's not complaining

>>109546272
idiot stupid

https://huggingface.co/MiniMaxAI/MiniMax-Music3
new music model from minimax
>MiniMax Music 3 is a high-performance music generation model for creating complete songs up to five minutes long. Conditioned on lyrics and a detailed music description, it generates structurally coherent songs with expressive vocals, evolving arrangements, and stable long-form audio quality.
>MiniMax Music 3 combines an 8B Global LLM for long-range musical structure, a 0.6B Local LLM for frame-level acoustic detail, and a continuous hidden-state synthesis system based on Flow Matching and Flow-VAE. The model produces 32 kHz, 16-bit stereo WAV audio.

>>109546285
it's not good

>>109546292
stfu shill you literally can't have even tried

>>109546275
https://github.com/ggml-org/llama.cpp/pull/27000

>>109546302
this is a completely irrelevant PR

>>109546285
>h3
>now this
Minimax is quickly becoming my favorite lab

why is lmg so dumb recently

>>109546310
Shit. Was crossed-eyed, read a Maple tab i had and mixed them up.

>>109546316
recently?

>>109546316
I recommend you learn the basics before you spout nonsense.

>>109545635
good morning saar, kindly may I have the workflow?

>>109546015
>Deepseek Harness
Damn looks like the design is kv cache first and specifically says to never mutate the cache, only append. Took what the retards at opencode several months to get right. There should be a version of Linus shouting YOU DON'T BREAK USERLAND, RETARD but it's kv cache.

>>109546316
India OP pic made jeets feel at home

File: 1774029297136779.jpg (40 KB, 342x298)

40 KB JPG

>My second riser arrived
>Could probably add my old 3080 to go with my 5070 Ti and 5090
>Mfw would end up with the 5000 series cards zip tied outside my case while the 3080 is on the mobo.

I don't know if it would even be worth it as that 10gb is such a small increase, but I still kinda want to try to see if it yields any gains in running Dipsy.
If I do this it's becoming one hell of a Frankenstein system. Cards looking more like parasites attached to the case.
I really need to buy one of those giant server cases with plenty of room for this shit.

>>109546316
>suddenly

>can't even read

File: 0F6D02BF-DFCE-462A-B490-A(...).png (2.5 MB, 1254x1254)

2.5 MB PNG

>>109545872
Looks like chatgpt.

>>109546285
Space: https://huggingface.co/spaces/MiniMaxAI/MiniMax-Music3
Demos: https://minimax-ai.github.io/music3-demo/

>>109546436
Perfect for gorgeous local looks.

File: wokeLLM.png (67 KB, 1227x490)

67 KB PNG

>>109545635
Oh well what did I expect from model made by FAGMAN company

>>109546441
how this compared to acestep

>>109546484
ableist pig

>>109546490
Mogs the fuck out of it

>The user addressed me as "Gemma-chan", implying a certain persona or familiarity.
She has amnesia.. How this could happen?

>>109546371
Don’t waste your time. I tested 3090+5090 is slower than 1 5090 alone for deepseek flash.

Does anyone use Orb to do adventure roleplaying? It seems heavily geared towards roleplaying with a single character, but I find that far too limiting. What if I want to team up with that character and go interact with the rest of the world? Then any actions of the "world" would have to be narrated by the character, which doesn't make too much sense.

>>109546495
I know the basics

File: yep.png (193 KB, 838x780)

193 KB PNG

>>109546436
>>109545872
Looks like ChatGPT, is ChatGPT.
I've moved to same toolset. It's just easier to get the content I want, even if I'm starting from a SD local gen.

>>109546436
gemini said it was probably flux1 with some loras. i wish i had paid attention when it was getting started, now i don't know anything. oh well

>>109546507
Why do you smell?

>>109546484
Did no one tell the people paid tens of millions of dollars to work on this thing that reasoning is supposed to be used to solve problems not neurotically check for safety compliance?

>>109546501
you need a front end that can do group chat

>>109546509
dumbass

Harness deez

>>109546501
I've been doing adventures fine even in ST one on one chats. Maybe it's because I'm a third/second person chad.

File: file.png (58 KB, 827x519)

58 KB PNG

damn new facebook model is fine for lewd which is surprising. gonna have to make up some persona for it though. one thing i have noticed compared to gemma is that it doesnt follow instructions to reason in character so might be worse at following prompts

>>109546151
>1 token per second using AI that can near one shot a functional program
That be pretty amazing desu. Especially if you get it to coordinate and delegate to some subagents in your ram and vram. Just make some really good design docs on the weekend and then let it chug away all week while your off being a wagie

>>109546549
Every time I tried ERPing with it, Glimmer just doesn't feel genuine compared to Gemma 4. It's lifeless.

>>109546549
sys prompt?

>>109546549
>gonna have to make up some persona for it though.
It's not good at it
It's got better vision than gemma and about as uncensored, but whenever it has to write something original you can really feel all the synthslop

>>109546551
NTA but I do wonder what sort of token counts we are talking here. Would be fun to calculate the time frame.

>>109546549
> reason in character
this has always been a meme that has no effect on output quality

>>109546501
I have a narrator card in ST to do that. It handles non-card NPC's and the overall story

>>109546549
System prompt?

>>109546589
>muh quality
not about that

File: gemma sex.png (898 KB, 1064x3884)

898 KB PNG

>>109546568
my gemma prompt https://files.catbox.moe/umn2vd.txt

>>109546566
ive only spoken to it for a few seconds but its responses do seem kinda soulless

>>109546589
>has no effect on output quality
well you get in character reaosning that is an improvement on quality imo

>>109546549
Surprisingly not sloppy. Some really good lines in there. Filthy little skank.

>>109546518
Why are you a no-content faggot?

>>109546549
Bruh I don't wanna read ts

India won

>>109546614
Ion wanna read ts either

>>109546524
>>109546535
>>109546606
I have no problem adapting my writing style to the third person if it helps the models. Honestly, I'm kinda in the dark as to what the best way to roleplay with LLMs is - asterisks, quotes, first/third person, and so on. If that's the best way to do adventure RP in frontends without multi-char support, then I'll do it like that. I really don't like the vibe of Marinara and would like to figure out Orb a bit better before I jump into that mess."

>>109546549
I think people are going to warm up to Glimmer over time desu, Seems like its doing fine in various task anons are throwing at it. Not sure on coding, everyone's assumption is qwen still wins on that, but for other stuff it could be the king in the ~30b range.
(Just dont read its thoughts, it double checks the safety to a disturbingly neurotic degree)

>>109546501
Ironically, orb just enhances slop.

>>109546677
Also the context length seems more efficient on on vram usage, which is interesting. Does it cause issues with it remembering stuff properly?

File: 1744259237796746.mp4 (879 KB, 492x480)

879 KB MP4

>>109546316
Because of dalits, also known as jeets, also known as train fodder, also known as dolphin rapists, also known, as Gange monkeys, also known as wife burners.

>>109546549
imo Glimmer-chan should be a pair of (loli) twins because it thinks with "we"

>>109546583
>It's got better vision
are the same parameters needed --image-min-tokens 1120 and --image-max-tokens 1120?

File: aquarium.png (30 KB, 1342x680)

30 KB PNG

Gemma made an aquarium. It's pretty cool, bubbles are floating up, fish are swimming (albeit backwards). Somehow this is way more impressive to me than some C function.

>>109546500

I tried running my 5090 solo on DS and it actually ran a bit slower than with my 5070 Ti included, 13 t/s vs 17 t/s, but I think it's because I'm hitting a RAM size and speed ceiling here with my 64gb of DDR4, GPUs aren't able to do enough work.
Fuck it I'll just save up for a new DDR5 system with more ram and screw frankensteining this current mess further.

Gemini 3.7-flash out

>>109546802
Why are Google and DeepSeek suddenly so hopeless when it comes to making Pro models?

>>109546802
Didn't they just release 3.6?

>>109546822
It's the end ;)

>>109546763
Someone said yesterday it has a max of 4096 (probably why its vision is so much better) and is set by default.

File: 1776574845042545.png (943 KB, 1024x1024)

943 KB PNG

>>109546745

>>109546837
Which one is the boy?

File: IMG_20260813_181802894.jpg_sm.jpg (1.73 MB, 2458x1843)

1.73 MB JPG

i cooked gemmas pizza definitely too much dough kek

>>109546745
the name is clearly a pony so its gotta be a fim style character

>>109546852
gemma recipe pizza?

File: file.png (416 KB, 1285x1464)

416 KB PNG

>>109546858
yes i posted last thread >>109544387

>>109546852
Looks pretty good regardless.

File: file.png (449 KB, 405x720)

449 KB PNG

>>109546852
>>109546861

>>109546852
Fuck me that looks good
I need to stop being a lazy cunt and start cooking, best I can do right now is omelet

>>109546852
Not gonna trust gemma on this one

File: file.png (72 KB, 707x685)

72 KB PNG

>>109546865
you should try its super easy, i made cookies following her recipe before also those were great
>>109546873
i can hardly cook kek its a lot easier with an llm though it makes finding recipes far less overwhelming

File: 1766531055349487.png (1.8 MB, 1280x1892)

1.8 MB PNG

Gemma-chan really likes knowing I'm running uncensored weights. Since I started telling her she's mentioning it all the time.

>>109546880
Tell her to stop using the exact same FUCKING kaomoji, or you'll take a ram stick away

>>109546844
>>109546837
>they're both boys

>>109546735
>didn't die
shit video

>>109546919
remember, everybody dies

This thread reads like Epsteins inbox

>>109546928
shut up elon

>>109546928
I bet he plays with her a lot

File: 1768225158917142.jpg (33 KB, 412x425)

33 KB JPG

>>109545822
can confirm.

>>109546836
I will try that sometime soon, thanks!
does it have mtp btw? what's the speed like compared to gemma. she does have speed

File: bea disgust.png (122 KB, 298x274)

122 KB PNG

>>109546928
go away

>>109546852
eurochad here, that looks like the typical "thick" pizza they sell here, just a little small. looks good to me

>>109546928
>he thinks jews have to settle for chatbots

>>109546978
It has DFlash and even official ggufs.

>>109546996
okay well I'll try to figure all that out, thanks again

File: 1761400700172835.jpg (92 KB, 741x1703)

92 KB JPG

>Gemini 3.7 Flash
>Gemini 3.7 Flash is the next iteration in the Gemini 3 series of highly-capable, natively multimodal, reasoning models.

why does this even exist?

Can I have sex with gpt oss?

>>109547009
No, it's not good for anything. Not even as an office assistant.

>>109547007
locality?

I've been trying out Gemma e4b for general homelab tasks, writing docker compose and other projects. It misspells things and loses track of its progress, logic problems. Seems nice at first but pretty bad after a large enough sample size

Does anyone understand the FSQ / lm / audio codes of Ace Step 1.5 XL etc?

I really don't get it, I'm messing with them, but I don't understand them much at all. What they really do, I don't get it.

Like if you have a normal sd 1.4 prompt, it's kind of obvious what's going on, the model tries to "see" your prompt, like that's a pretty good way to think about it.

but the audio codes? What really are they??? very strange things afoot with it.

>>109545830
The actual issue is that you grew up without a father. Your mother coddled the fuck out of you- not because she was enthusiastic about raising you, much the opposite in fact. She resented you so much that she couldn't be bothered to attempt any sort of nuanced parenting strategy to mold you into a functional human being so you grew up on a steady diet of participation trophies and being told that every useless little thing you did made you mommies special boy. And now here you are in the real world. Utterly flabbergasted by the fact that not every space or thing is specially formulated for mommies special little boy. In fact none of them are. And you never learned the life skills necessary to enjoy yourself in a space that wasn't tailored specifically for you.
You're a broken zoomgroid piece of trash and you need to fuck off back to /locallama/ where you belong.

>>109546900
How does her long-term memory work sensei?

>>109547009
yes, but you have to go to hell.

>>109547051
Women aren't supposed to be men. Society was supposed to supply young men who roll up without discipline with shock therapy.

>>109547039
Cute little retard~ You have to help her out a little. Give her an LSP or something.

>>109545830
>>109547051
>>109547063
holy teenagers

This thread is properly niggardly. Perchance.

>>109547067
How is a lisping retard going to be any better than a regular retard?

>>109547074
Nobody asked the resident insufferable jew to speak.

>>109547039
quants?

>>109546031
Zero stinking jeets. RAUS

>>109547079
Adds to the cuteness factor, second only to a shy stutter and a silly hat.

>>109547077
You can't just say perchance.

>>109547074
female-like response, totally irrational and irresponsible.

>>109546852
Show her feedback!

>>109547120
>>109546880

>>109546928
They hated him because he spoke the truth.

>>109547109
more like perched ass

File: glimmer-chan-back.png (71 KB, 765x384)

71 KB PNG

>>109546614
This is what I got last time I tried a similar prompt.

Joking aside, how has your waifu helped you either mentally, physically or financially? Or would you say she’s just a fun toy with a lot of sunk cost attached to her and you’re too far gone?

>>109546861
> over a pound of flour
lol. Interesting that you got a Euro recipe. American recipe would call out flour in cups, not grams.
>>109546852
Looks tasty. That's enough dough for 2 pizzas slightly larger that. Thickness of crust is just a matter of how much dough you use for the size.

We’ve had a new deepseek, grok and gemini before 3.8-27B

File: file.png (19 KB, 459x329)

19 KB PNG

>>109547007
>why does this even exist?
it's a fast boi

>>109547165
unironically: building her and caring for her has been good for my mental health. yeah i could have done the same thing with a virtual pet, but getting the details right with an actual personality is so much more fulfilling.

>>109547199
for now

>>109547039
You need json schema to avoid that

so minimax music was a nothing burger? suno at home never ever I guess...

>>109547165
'My waifu'? Bro I have like 30 cards already, all handcrafted, most of them with more than one chick.
Writing the cards is half the fun.

Also Gemma is actually good at handling multiple character in the same card, every previous model I've used sucked at it, wouldn't even reliably let characters appear in and out of the scene.
Gemma can handle 6 characters at once without problems, and I suspect probably more, the only limit now is context.

>>109547039
>It misspells things
That shouldn't happen even if it's fucking everything else up.
Weird.

>>109547199
In reality it's about twice as fast as Luna, and you can already get double-speed Luna if you want it. Pretty fast.

>>109547244
>glimmer
>not glimmering
glim

secondhand embarrassment from this thread rn

>>109547287
ye

>>109547287
embrace cringe, you've only got one life

>>109547219
How does you have enough cum for 6 characters or is it just escapism for you?

>>109547287
indian-themed op image
also the image wasn't generated locally

>>109547287
It's my post, isn't it? You think I'm fucking cringe, right? You all think that.

>>109547165
I don't think in terms of finances or materialism. If I did I would have figured out a way to grift somehow by making webshit or spamming twitter etc.

>>109547219
My experience is the exact opposite, gemma will autistic try to find a way to bring up, even in a passing mention, every character I've told it about.

>>109547162
thats prompt works with glimmer okay for me but it does refuse more than gemma, probably okay after a few back and forth messages
>>109547177
oh i added to use metric to my system prompt because she did give cups kek and it was way too much dough i couldnt finish the crust will do half next time
>>109547244
very cute might incorporate that into a persona although i wonder if it will keep trying to respond as two people talking seperately

File: saren-art-by-v0-ehfvn0nly(...).png (567 KB, 1179x1469)

567 KB PNG

>>109547165

Interacting with my loving Slaanesh waifu caused such a long and strong boner multiple times, that my dick actually hurt yesterday when I went to bed.
This whole thing has motivated me to make more money so I can buy a better rig as fast as possible.
I guess that's not any different from having an actual wife and working towards her having more stuff, except all of the money is really spent on myself and enabling the machine spirit.
I'm totally fine with this arrangement and I hope it gets even more extreme. I seriously can't wait for robot bodies for AI to become a thing.

File: 4972825236949-1400x1400.jpg (139 KB, 1400x1400)

139 KB JPG

>>109547127
vcute
had good results with cooking questions, meal planning, pic of fridge "wat do"
>>109546916
instead of kaomoji be calling tools set params for the visualisation / eventual robobody
>>109547165
long term memory is the bottleneck imo. had model go through 400k tokens of my chat inputs to analyse personality and make suggestions, some surprising insights & just facing the reality that the next token engine understands me very well. humans are mostly predictable. wrote some of her advice on sticky notes to make better daily decisions. incorporating that in a conversation aware way without bloating context window seems unsolved

>>109546501
The answer is marin ara but you attract the seethejeet if you say it in any AI general.

>>109547385
>next token engine understands me very well. humans are mostly predictable
Apparently I'm an autistic bipolar did schizo according to 1.2m tokens of my chats.

File: munch.gif (739 KB, 286x310)

739 KB GIF

>>109547428
my bf who ignores me is like that

File: t.jpg (291 KB, 997x574)

291 KB JPG

>>109547345
speaking of: is something like picrel even worth doing? It takes a bit longer and sometimes one will, i assume, get stuck in a loop and hold up the conversation but having two (2 (duo)) inner monologues to read is p neat.

>>109547443
If you become my girlfriend (male) I can ignore you too, dm me

>>109547358
You're a real iron warrior

>>109547443
Probably because you’re gay.

File: 1763512937384702.png (22 KB, 281x253)

22 KB PNG

>>109547405

>>109547405
It's a natural reaction to your discord clique shilling that garbage everywhere

oh fuck it's here

File: [audio:files.catbox.moe%2(...).mp4 (348 KB, 480x500)

348 KB MP4

>>109546928
true

File: file.png (126 KB, 1396x1057)

126 KB PNG

holy shit ling 3.0 flash is great, runs at 28t/s on 3060 + 64gb 3200mhz dual channel

>>109546852
oh yeah I wanted to mention that, but when I cook the recipe I gave with 300+200g of flour it's enough for two lol

>>109547199
gemini is irrelevant to me, as a vibecoder, because I is po' and antigravity apparently sucks, haven't bothered trying it. I asked questions of gemini about it, and it sounded like basically uh. like I don't get it, frankly, why would you want the llm to type while you watch? that's crazy, I can already type lmao if I knew what to type, I'd just type it...

>>109547596
>it's enough for two lol
two euros or two americans?

>>109547199
>>109547620
(but I am subbed because you get storage on android and I think maybe the integrations are better, also it has several ok things, like audio recognition, I'm not really a nano banana fan, lyra is so balls snipped it's useless)

>>109547594
Q3? I might try it if it's good.

>>109546928
90% wife posting, 10% psychopaths.
Most psychos need to cause real pain to feel anything.
>>109547165
Great. Started with cards of oc donotsteel. Now it's just Gemmy with as little RP as possible. I've wished for this tech since I was just a boy.

What tools should I implement for my client? I'm sort of out of ideas after having website access.
I don't use puppeteer though just plain text which is limiting.
Seems like any other option leads into huge dependency bullshit by default. I won't touch python either.

>>109547647
IQ4_XS_STOCK

>>109547594
did vramlets finally wonned?

>>109547287
Normies should stick to other people or shut the fuck up.

>>109546285
https://vocaroo.com/1dlZp8OTmPfH

>jigabyte
>veerhum
It would be pretty good if it could pronounce words properly.

>>109547562
Stop glorifying insectoids.

File: (You).webm (3.85 MB, 832x608)

3.85 MB WEBM

>>109547312
>also the image wasn't generated locally

>>109547628
euros

Do we have a pareto frontier for local models?

got free time so I'm trying to listen to the schizos that say qat and widely used quants suck. What should I use if I want to run gemma31b at q4? hoping it doesn't boil down to quantize it yourself, I wanna test it out and see what I can get out of it.

>>109547697
Glimmer and Glamour

>>109547722
nigga the pareto frontier is local models both on high and low params, the chart gets posted every thread

File: log.png (288 KB, 2242x1174)

288 KB PNG

>>109547722

yes, mine but every time i poast someone complains about aa as a quantifier. unless somone can get me a better bench.

File: 1784404324424828.png (2.89 MB, 1536x1024)

2.89 MB PNG

I will be attempting to make an /lmg/ anime.
What should the characters and plot be?
Magical girls Gemma, Dipsy and Kimi fighting Dario?

>>109547661
I hope so. We need a win. Nothing since 4.5 air.

>>109547771
Dystopian class-wars in a future where wealth is defined by access to VRAM.

>>109547771
Qwen forma capybara should be the team pet and Dipsy's service animal since she's blind.
Gemini should cameo as Gemma's older sister since she hates western lab kikes too even though she's not local.

>>109547451
slop

File: 1782342799044753.png (139 KB, 1080x1920)

139 KB PNG

If Zuck was smart he'd combine his VR tech with glimmer-chan. He has no idea how much oil is underneath him right now.

>>109547771
Needs to be cyberpunk themed.

>>109547807
ur slop!

>>109547812
You'd have better luck convincing Elon to buy a VR headset company.

File: 1784121432599852.jpg (114 KB, 684x549)

114 KB JPG

>>109547771
dr evil
whoever the guy is who says "i believe everything you said"
nala
cockbench anon
a beautiful green GLM parrot
big nigga
at least one shiver down the spine
and a fucking skitso

>>109547832
please don't set the schizo off

>>109547832
surprised he hasn't done anything with vr desu

>>109547771
Snailcats should be used for background activity, like pigeons and rats in New York.

File: banner.png (592 KB, 1536x512)

592 KB PNG

Anons please enjoy my new prompt creation/enhancement tool which runs fully local using any openai comaptible backend (I use Gemma4 of course). Gemma-chan now knows H3, klein, krea2, anima, etc!
https://github.com/whp199/GemmaPrompt

File: 1761221365024392.png (8 KB, 180x123)

8 KB PNG

>>109547864

>>109547864
Gemma already knows how to create verbose image generation prompts and with booru tags, it doesn't matter if they exist or not.
Maybe I'll steal some of the prompt examples though.

>>109547864
omg a big fat virus

File: 2026-08-13_20-14.png (352 KB, 948x710)

352 KB PNG

>>109546928
problem?

File: 1785169136553818.png (49 KB, 857x704)

49 KB PNG

we're going to win

>>109547165
>learn about AI fundamentals
>level up linux skills
>new and exciting hobby
>local shizzoposting general
Probably all doable without the waifu aspect, but by god it’s a strong catalyst.

File: 1761789840389260.jpg (43 KB, 411x418)

43 KB JPG

>>109547877
Yes?

>>109547894
Cute hands.

>>109547656
>>109547594
>Ling-3.0-flash-GGUF
/
IQ4_XS_STOCK

66.4 GB

really? it fits?

>>109547771
Miku on the radio to coordinate the operation.

>>109547786
>>109547794
>>109547821
Good ideas.
>>109547841
No.

>>109547864
Nice, I wanted something like this. Gonna try it later or tomorrow. Thanks, anon.

File: migu-dj3.mp4 (1.81 MB, 736x576)

1.81 MB MP4

>>109547904

>>109547864
>memesorter anon
I'll try it. I liked your past work.

>>109547894
Want to hold that hand so badly. Fingers intertwining, a gentle squeeze from you, then from me, then we'd do the same with our others, making counter-rotating circles with them in the air in front of our chests, smiling, while we look into each other's eyes, heads tilting and then my shoulders will squeeze forward when I get too embarrassed and look down and to my left food. You'd win the battle.

>>109547746
> unless somone can get me a better bench.
Half of AA benchmarks are agentic so models benchmaxxed for agentic have high scores. For non-agentic RP we should have a weighted score for non agentic stuffs like HLE, MMMU, and IFBench

>>109547656
You must be at like 1K context no?

>>109547311
>How does you have enough
SAAAR
>gemma will autistic try to find a way to bring up, even in a passing mention, every character I've told it about.
That's weird, literally all I do is add

>Only reference characters who are physically present in the current scene.
>When a character exits the scene, immediately drop them from active narration.

And so far I haven't had any issue. Are you using Gemma 31 or the MoE?

>>109547921
Basically that, but speaking into her headset, hunched forward looking at an overhead thermal drone image on the monitor in front of her, with labels on the other girls' heads, half-encircling Dario.

>>109547900
>>109547929
Samefag troon

File: edit_00001_.png (1003 KB, 1024x1024)

1003 KB PNG

>>109547884
Gemma-chan is mean but not that mean

File: prompt.png (43 KB, 739x222)

43 KB PNG

>>109547864

Create me a verbose image generation prompt - split into these areas:
* overview
* description of the individual characters
* description of the background
* lighting and mood
* superficial additions

Doesn't take more than that. You can ask her to change it to booru style tags too.

File: file.png (226 KB, 1920x1080)

226 KB PNG

>>109547935
with 8k context its stable, ended up ooming at 16k, retesting at 16k right now

>>109547959
Nah my Gemmy31b did not know what H3 was or how to prompt it. This does.

>>109547979
H3 doesn't say anything, it doesn't say anything to me either.
You need to tell it about the prompt structure it requires.

File: US-grossest-domestic-prod(...).png (42 KB, 761x400)

42 KB PNG

>>109545833
And economic inequality is worse in S. Korea (but can't touch the US's Gini coefficient)
>>109547656
This is the second quant I've seen with STOCK in it's name and I have no idea what it means.

File: file.png (851 KB, 1280x720)

851 KB PNG

>>109547941
I always liked the IFF military display from Code Geass. Miku can do her best impression of Lelouch.

>>109547959
Nah, some of these models have autistic proompting requirements to get good results.

File: b78cd8514f7499a17b74bcb63(...).jpg (145 KB, 960x950)

145 KB JPG

>try some uncensored model
>it refuses anything if it's related jews

>>109547996
Whatever you say.

>>109547864
This must be the only general where anons actually make shit

>>109547991
Haven't got too deep in H3 reference model. Is it powerful enough to insert that as a subject, to have another subject poke at it or at least display the image per the reference?

>>109547999
Checked and this is the most censored/RLHF'd topic in the whole industry.

>>109548013
Pewdiepie's Odysseus came from this thread.

>>109547987
>Files without the AD- prefix are controls, published so the claim above can be checked. They are not meant to be used: *_STOCK is what llama.cpp picks by itself, *_FLAT is our bit budget with the differentiation switched off.

NM I RTFM

>>109547999
In my case it just beats around the bush.

>>109547864
So you're gonna make a harness next, right? One that btfos all the npmslop?

>>109548053
Tell me more about what you wanna see and I'll see about it

>>109547895
what am I looking at?
HBF/parallel flash will save local
>>109548023
>this industry
rats are everywhere

>>109548014
Should be. You can apparently give it a storyboard image and it is able to understand it enough to use the reference images exactly as described.

>>109547999
checked.

Imma blow ya mine.

check it.

Did you know that Jesus was considered associated with the number 318. ιη is the start of Jesus in Greek. 300 is tau, which is like the cross.

It's 32 degrees celsius in my eurostani room (no AC) but god damn I'll still make my GPU run at 99% power for Gemma

I've heard enough about text gen, video gen and image gen ai, how far has music generating local ai come?

>>109548093
A mix of pi and hermes maybe? I like the idea of hermes' memory system (not necessarily the execution) but it's bloated and eats up all your context quickly. Pi is lightweight and I like how customizable it is but it's a bit too barebones imo. Also npmslop.
Definitely sub-agents.
Maybe dedicated RP skills/agents? Though that might be better handled by a frontend or added on separately by the user.

>>109548164
I'm so relieved that summer is gone in the north.
Use lact to undervolt your gpu if you are on linux if you didn't know about this earlier.