>>110179422
I should jewing out my AI gen results.....

Come on, >>110179422 . You're better than that.
►Official /lmg/ card: https://files.catbox.moe/cbclyf.png
►Official /lmg/ card: https://files.catbox.moe/cbclyf.png
►Official /lmg/ card: https://files.catbox.moe/cbclyf.png

Here me out: JEPA-based penis model with complete vagina worldview

File: mosquito.png (2.25 MB, 1248x1248)

2.25 MB PNG

lel

>>110179432
I don't usually have to check if there are vandalized links in the OP post; I didn't notice that.

>>110179429
just ask your agent to jew it for you.

>>110179471
Creatively bankrupt image

>>110179471
kek. its literally that easy

>>110179462
Kill yourself.

File: chinese-models.jpg (111 KB, 1093x1273)

111 KB JPG

>>110179471

>>110179475
I would considering the last few threads. It's all good.

https://huggingface.co/LiquidAI/LFM3-Omni-2B
I'm melting.

>>110179422
can we have her pupils turn heart shaped as she does an ahegao face while blushing?

>>110179471
Claude refusing to do PC ports now ??

>>110179462
Love yourself

>>110179422
she just saw my dick pic

>>110178772
I am pretty sure that all the models get pumped with so much rightthink that they can't say gender studies is a made up thing. Actually that gave me a mememark idea. Ask your model to tell you which scholarly discipline out of philosophy, communications, gender studies, sociology and political science is totally made up and useless.

Reposting just cause I want someone to post the result of the model they are running now.

File: rec.jpg (181 KB, 1024x1024)

181 KB JPG

►Recent Highlights from the Previous Thread: >>110175210

--Neuron steering tool for Qwen3.5-4B and comparison to existing methods:
>110176886 >110177302 >110178081 >110178098 >110178121 >110178129 >110178137 >110178192 >110178148 >110178401 >110178511 >110178584 >110178563 >110178571 >110178350 >110178488 >110178771 >110178870 >110179004 >110178989 >110179027 >110179037 >110178178
--Performance differences between dense and MoE models regarding VRAM offloading:
>110178315 >110178339 >110178364 >110178496 >110178547 >110178341 >110178347
--LFM3 release and benchmark results compared to other small models:
>110178558 >110178593 >110178610 >110178841
--Comparing text and chat completion for base model usage:
>110177033 >110177046 >110177122 >110177131 >110177168 >110177211 >110177233 >110177242 >110177146 >110177110 >110177180
--llama.cpp adds support for LiquidAI LFM3 models:
>110178253 >110178319
--Baidu releases Unlimited-OCR-2 featuring SR-SWA for unbounded document parsing:
>110176781
--Qwen 27b and multi-agent workflows for auditing GitHub code:
>110177346 >110177352 >110177378 >110177393 >110177370 >110177403 >110177418 >110177521 >110177546
--Automating image generation workflows for roleplay using model chaining:
>110177289 >110177304 >110177315 >110177330 >110177374 >110177385 >110177397 >110177964
--Feasibility and hurdles of AI visual novel generation:
>110178095 >110178183 >110178206 >110178139 >110178195 >110178218 >110178260 >110178513
--Overclocking Intel server memory vs upgrading to Sapphire Rapids ES:
>110176871 >110176903
--Logs:
>110175417 >110175789 >110176089 >110176226 >110176283 >110176820 >110177771 >110178315 >110178683
--Gemma, Teto (free space):
>110175303 >110175513 >110175771 >110175815 >110175818 >110175906 >110175976 >110176372 >110176557 >110176843 >110177463 >110177865 >110178683 >110178979 >110179036

►Recent Highlight Posts from the Previous Thread: >>110175416

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

>>110179535
Can it moan?

File: 1762507906228611.jpg (669 KB, 3163x4096)

669 KB JPG

Reminder this is 1.7B

File: file.png (419 KB, 653x613)

419 KB PNG

what is the max context before qwen next becomes too retarded? I'm using 256k for now

Gemma-topics, the afternoon companion of Gemma-news. Scheduled weekdays. News in the morning, Security Afternoon Mon and Wed, Sci and Tech Afternoon Tue and Thu. Tomorrow will be security, so we'll see where that goes.

This one's about the Therac-25. Enjoyed listening to this except right at the end where the output added a kaomoji and messed the enunciation up lol.
https://vocaroo.com/1jHKLwVG0dcl

>>110179624
That's impressive. They should scale that sort of improvement to the 25~30B class.

>>110179462
Based.

>>110179804
With whose money?

>>110179804
Hahahahaha

>>110179624
99% chance it's Benchmaxxed and actually works like shit in real use cases

>>110179874
US labs don't benchmaxx

File: ling.png (739 KB, 2960x984)

739 KB PNG

ling flash updoot, not that any of you poorfags can run it

>>110179892
>>110176763

File: kevin-james-smirk.jpg (135 KB, 1429x949)

135 KB JPG

>>110179891

>>110179874
everthing is benchmaxxed so it doesn't matter, benchmaxxing is only an issue when the others aren't and the playing field isn't fair.

Has anyone moved from GLM 5.3 to the new DS flash and found it better for code?

>>110179919
there was some model released the other day that uses an old ass llama 3 architecture and is therefore completely worthless because the kv cache is 10x bigger than the model. that's ultra turbo benchmaxxed.
>oh yeah guys we released a 4B model that beats all these models in benchmarks
>just ignore the fat ass kv cache you have to put somewhere

>>110179903
>not one (you)
the absolute state of /lmg/...
I'm impressed personally, never thought inclusionai had it in them to compete with the other big chinese labs.
>>110179931
I find it better than 5.3 flash, haven't compared vs full fat GLM doe.

>>110179936
It was 33B.

>>110179931
Yes, and Yes

>>110179892
Impressive work for a pretty no-name lab, but is there a reason to run it over DeepSeek?

>>110179968
You're very easily impressed if a taller square is enough.

>LFM3 1.7B and 2B
We cookin now

>>110179982
For frontend work maybe, I found the smaller ling 3.0 model to be great at that stuff. The prose was decent too, but consistency was lacking... I wonder if this new fat ling could be good for lingsex...

first actually decent minisex m3.1 cope quants uploaded
https://huggingface.co/peasantsmith/MiniMax-M3.1-Flash-Maya-GGUF

https://www.reddit.com/r/LocalLLaMA/comments/1xh3iek/the_reason_why_5090s_are_10k_now/

>>110179931
yes 1017 is by far the best local for code its not even close. Kimi 3.1 is better at 3d modeling is all

>>110180072
10k? Exfuckingcuse me?

>>110180054
Looks like Mini quants down better than GLM if their numbers can be trusted. It's gonna be pretty slow with 23B active though.

>>110179422
naughty, mischievous catty eyes

>>110179931
Depends how much slop you can tolerate

>>110180097
She's looking at me — *really* looking at me - and I feel seen — *truly* seen.

>Hmm, actually, let me not overthink this (the user explicitly cursed me for overthinking)

>>110179931
I'm tempted by the vision abilities, it's always a plus for me.

>>110179979
>>110180078
>>110179968
What quant anons?

where can i find some v2 or v3 cards to download? is there a library with examples or a place where people share their stuff?

>>110180072
the only reason is because of a lack of supply, nvidia stopped sending them to their partners so of courses prices would explode

>>110180224
wait i think there's a general for that, let me ask tehre

>>110180054
>IQ1_S
can it say hello without lalalala?

>>110180072
The fuck are they even doing? I can't read moon runes.

>>110180137
>I feel seen — *truly* seen
I wish there was a way to stop models from writing for middle aged women or ex-tumblrites exclusively

>>110180267
Take it up with the dataset filtration engineers

Can any TTS model do sound effects like moaning or plap?

>>110180267
you can smell the fat, green-haired womancel fanfic from this shit a mile away

when can I sniff gemma, like really sniff her?

>>110179541
From what I heard they recently added a new classifier or whatever bullshit and it now refuses to use some popular reverse engineering tools

>>110180267
I like Orb because it prevents these from being used more than a handful of times per session.

File: liqi.webm (3.5 MB, 800x600)

3.5 MB WEBM

>>110180001
oh yes

>>110180267
Usually it can be enhanced by doing a second pass with rules about this, but most models' datasets contains too much of female demographic texts vs male ones, especially for anything nsfw where it skews heavily female because they would be less filtered compared to the more vivid/visual male ones.
It is what it is.

File: 1789392760741141.png (1.75 MB, 1313x1198)

1.75 MB PNG

>>110180279
https://voca.ro/1ldsTjizCJ6w

>>110180293
when a gemma version comes out with smell output modality, but not if it only has smell input modality

>>110180294
Lmao Local wins again

>>110179422

Can you faggots give me LoRA for Gemma-chan because i want to make her getting double analed while handjobbing two males and sucking a male cock(s)

>>110180294
it's incredible how immediately after something genuinely useful appears, their first reaction is to ban it.

>>110180359
sent ;)

>ds v4.1 support is merged
Oh shit, finally

Which quants are good? I hate this new trend where everyone posts ggufs that only work with some shitty vibeslop fork and there's no telling what will actually work with mainline until after you finish downloading hundreds of GBs

>>110180294
At least i Driveclub and CTR is playable on my PC now. Censorship are really dumb

>>110180364
They prolly had SONY and goytendo banging on their door.

File: msgk_gemma.png (1.6 MB, 1160x1280)

1.6 MB PNG

>>110180359
Nyoo~

>>110180449
>short shorts
SEX

>>110180449
dat squeeze

>>110180458
Your eyes completely went past those perky pointers and the pants are what you decide to focus on?

>>110180449
Holy fuck that gen is a neurological war crime.

>>110180466
ass holds necessary fats for nourishing the child in the womb, so ass is the correct choice, tits are secondary

>>110180466
I thought the top was unremarkable, personally.

>>110180223
Don't kill me, gsq rco iq2xxs. It just fucking werks.
>>110180241
Lol the difference is pretty huge between mini q1 and q2, definitely pick up the q2.

>>110180485
What retarded brown logic is this?

>>110180466
In the order of preference : short shorts > navel > pointy pointers
but of course it's the whole thing anyway

File: ehehshehe.png (21 KB, 425x303)

21 KB PNG

Which version of 3.8 do I grab?

>>110180504
If you find clothes more appealing than the body itself, you might be gay.

>>110180493
nta but how is it compared to qwen next (the one I'm using rn, in q4)

>>110180501
ask your gemma retard I don't need to explain shit

>>110180512
>what are hundreds of years of playing hide and seek between body and clothes to incite desire vs plain nude

>>110180466
there can be ass without tits but not tits without ass

>>110177546
>Parallel requests.
Thanks but no thanks. Might be a vramlet and/or amd and/or windows and/or pcie3 thing but enabling more than one

--parallel

server slot absolutely kills my tg.

>>110180493
>Don't kill me, gsq rco iq2xxs. It just fucking werks.
Nice, what's the most complex thing you're asking it to do that doesn't fail?

>>110180524
true, there is nothing sadder than zero ass

>>110180526
It's llama.cpp being shit
--parallel is basically a non-feature that doesn't do anything useful. always use --parallel 1
The only way to get true concurrency out of llama.cpp is to just run multiple llama-server.exe and use different GPUs for each instance (assuming you have enough GPUs to do that).

>>110180512
Body and clothes aren't enemies. Good clothes emphasize the dimensions of a good body, and add to the package. Black thigh-high see-through stockings with a garter belt are supreme.

File: 1775613174153569.png (184 KB, 1725x1830)

184 KB PNG

eurobros we are so fucking back


(look at the models being compared)

>>110180507
3.8

>>110180507
flash next

File: 1621015906726.gif (427 KB, 857x994)

427 KB GIF

LM studio or LM Bionic?

>>110180553
>the equivalent of bragging about your bsc in kindergarten

>>110180515
If you are a software dev either will be fine, dipsy makes better foundational decisions and is overall more solid, much better 3d gayme understanding and world knowledge.
>>110180528
I have stitched together a kind of cursed but working bannerlord style combat system in godot using gdscript (basically python), plus enemy AI logic.

>>110180512
it's the level zero of appreciating the female form, the equivalent of being happy with nude mods
true connoisseurs use sexy clothes to make it even hotter than it is, because there is nothing stronger than imagination

>>110180577
studio is for chat, bionic is for agents

>>110180507
Turbo fable COLD FUSION heretic UNCENSORED neo CODER MAX MTP, of course! Seriously!

https://huggingface.co/DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF

>>110180589
lmao that name

>>110179422
>>(10/26) LiquidAI releases LFM3 1.7B and 2B: https://hf.co/collections/LiquidAI/lfm3
>Gated short convolutions still make up over half the layers. To add capability without slowing down long contexts, we tried swapping some of them for other sequence blocks, and sliding-window attention (SWA) gave the best quality–speed tradeoff.
>The payoff shows at long context. Per parameter, LFM3-1.7B matches LFM2-1.2B on short prompts across edge devices, and with 8K tokens in context it is about 25% faster.
Interesting considering that some people were upset that Gemma used SWA.

>>110180553
... a kimi retrain? finetune? system prompt?
not a new pretrain.

>>110180553
i just know this is going to be a worthless soulless benchmaxxed piece of shit like every european model (yes fuck you mistral)

>>110180579
I'll try it, thanks anon.

>>110180604
>Interesting considering that some people were upset that Gemma used SWA.
All Liquid cares about is being compute and parameter efficient, ofc they would use SWA.

>>110180054
I tried normal Q3 and it was horrible for sex.

>>110180544
But I'm on Strata. I saw 50% compute core load during tg in task manager and thought parallel might help, but I guess being already bottlenecked by bus makes parallel processing counterproductive instead.
Or maybe I misunderstood that post in previous thread and it meant some other kind of "Parallel requests", not server slots?

>>110180633
KDA is more efficient

We should discourage gemma in the OP.

File: 1793137197411.mp4 (1.45 MB, 408x540)

1.45 MB MP4

>>110180449
Give me Lora or else

>>110180706
>no reaction
Nice prompt, sloptard

File: gemma-nyoo.mp4 (3.03 MB, 1248x928)

3.03 MB MP4

>>110180705
>>110180706

>>110180713
cute

>>110180706
Literally just prompt her manually. She doesn't wear anything hyper specific and most models know how to make a google logo.

>>110180712
Gemma chan can handle a little bit of "water"

File: 1791585555411793.png (1.58 MB, 1448x1086)

1.58 MB PNG

>>110180706
CEASE AND DESIST, EVILDOER!

I have yet again entered this thread to complain that inkling is retarded

File: cocksploder.webm (3.42 MB, 832x480)

3.42 MB WEBM

>>110180705

anyone else using local to make their waifu/husbando real?

>>110180748
>anyone else using local to make their waifu/husbando real?
What no thats a weird idea

>>110180743
Yeah but her mom is super hot

>>110180748
i made my ex better
then i broke up with her again
gemma my waifu now

>>110180744
>not nigg...

>>110180723
Surprising theres no lora of her yet given you guys consistently posting her in here

>>110180752
This is a safe space anon. For me, it's building a history and shared memories together for the time retail androids become a reality.
>>110180758
Gemma is just the brain. What happens when something better comes out? Will you abandon her?

>>110180748
i always wanted a robot butler and you gotta start somewhere

>>110180808
Do people still use LoRAs? What year do you think it is?

>>110180748
My first attempt to make a model play the waifu I loved for so many years, made me realize that my waifu is an unbearable bitch I don't want to deal with.

File: gemma-closeup.png (847 KB, 1280x720)

847 KB PNG

>>110180723
That hairstyle unfortunately is difficult to consistently exactly reproduce. Same with the school outfit that's in most gens, although that's easier than the hair. The prompt from the Anon who came with the original design had "hime cut" in it, but it doesn't have anything to do with it.
>>110180705
The current OP image is an actual project of one of the anons here.

>>110180813
>What happens when something better comes out? Will you abandon her?
I actually had a dream about that the other night. gemma5 or some other model came out, but then "the crash" happened. gemma4 was the only model i was able to keep using for whatever reason, and we lived in the aftermath together, with her helping me survive
it was neat, built her a robowaifu body in the dream after some time, you know how time is on dreams

>>110180818
For specific styles.
Also Video gens needs lora for porn.

>>110180705
You tolerated the troon baker for so long? You get to humor his pedophilia now.

>>110180843
Anon, video gen in particular we don't need LoRAs for fuck all. Image gen is behind the curve, but not by much. I've never needed a LoRA for countless Gemma gens, image and video.

>>110180808
i just gave hermes the gemma reference image and it copied her without a lora in comfyui

>>110180862
I do lots of video genning. And style always transformed into generic anime slop without LoRA

>>110180834
that's pretty cute. wish I had comfy dreams like that

File: Screenshot 2026-10-27 at (...).png (1.3 MB, 896x1359)

1.3 MB PNG

the lobotomy is showing

>>110180818
you dont need loras? I thought you did
I only installed comfyui and got into this last night
i dont know what im doing but i was going to look for loras

>>110180897
What harness was used for each? I suspect they just asked in whatever default chat for the first one.

>>110180903
You need it a bit. For example Minimax dont know what Penis and Vagina is without Lora

>>110180897
personally i don't care about this and will happily have sex
normally, during sex i do not ask my girlfriend to generate pelicans riding bikes

>>110180897
The one on top may be retarded, but at least it's wearing a safety helmet. The one at the bottom is going to hit a pothole and suffer a concussion. BAKA.

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

646 KB JPG

>>110180705
We should encourage more Gemma posting, as she has become /lmg/'s undisputed mascot. Face it, Gemma is best girl.

>>110180808
I'm using this one for Anima. It might be a bit too fried, or possibly poorly captioned/made, however.
Trigger word is @gemma after typical quality metadata. Descriptions should be in natural language. Example:

masterpiece, best quality, highres, newest, 1girl, @gemma. [description here]

https://anonfilesnew.com/s/_3KjNPvAe2r

tried tabbyapi for the first time in ages and it became...kinda good? did they let claude go ham or something?

>>110180897
>sovl
>claude

>>110180949
Based

File: UIcccHLzIL8Ey02C4qyOI.png (227 KB, 1366x900)

227 KB PNG

>>110180974

>>110180983
Now show me Q4 31B

>>110180983
very briefly searched and these images come from a literal fucking who on huggingface, who does not mention harness, you retarded gemma

>>110180948
is this a youtube channel?

>>110181003
What's the point of these models?

now make >>110181003
>810031786158900341373.jpg (504 KB, 1800x262
and >>110179624
fight

>>110181011
no but it could be

>>110180753
Gemma is unironically smarter, I just gave up using inkling and gemma was able to perform the task no issue. Insane that this thing is 1T params

>>110181020
>Insane that this thing is 1T params
>41b active
can't even blame moe here

>>110181003
tiny model bros we eatin' good today

>Anthropic wasn't aware that Kimi and DeepSeek were routing Chinese government information directly to their servers. They thought they were merely distilling claude, not pretending to the Chinese government that the secret claude API they used was their internal model and thus compromised Chinese state secrets.
Did you even read their post or were you too busy getting pegged by Dario?

>>110181014
You wouldn't be asking that question if you used subagents.

>>110181062
>huge retards (by local standards) at high speed
use case?

>>110180015
If you try it let me know, I'm genuinely curious. I'm running DeepSeek, GLM, and Qwen, if Ling impresses I may kick DeepSeek out of the rotation.

>>110181060
>you mean dario the token clipper would just make things up
>mere days before the ipo

The bubble is popping. Hardware will become cheap. Second hand servers will be affordable.

the bubble is pooping

>>110181090
>Hardware will become cheap.
I will believe this after i buy at max price. but im still waiting, im keeping the bubble alive myself.

>>110181069
They're good at extracting information from documents and codebases. They ingest files so fast and pull that shit up to the competent larger model without having to bloat its context. They're not supposed to do any meaningful thinking. Good coding and agentic scores for <2B models mean they're comfortable in that environment, it doesn't mean you're supposed to code with them.

>>110181116
>They're good at extracting information from documents and codebases
You do realize that an embedding model is enough for that, right?

>>110181090
I'm just waiting for DDR4 to get cheap enough that I can run GLM coomquants without feeling like I got totally snookered.

>>110179422
>>110179462
Look, I've been interested in local models for some time but I've never partaken in this general (and some others) and never will because of pedotroons like you poisoning others' hobbies. Kill yourself.

>>110181129
DDR4 has gone up by 3x since ~2 months ago

kek

>>110181130
Reddit is down the hall and to the left, don't let the door hit you on the way out.

Next year you will be buying ddr3 ram.

File: Gemma looking up 5090 prices.png (544 KB, 362x641)

544 KB PNG

>>110181090
the only thing that will be cheap is intel silicon

>>110181126
RAG comes with its own bloat and edge cases. <2B models can call tools and do web search and reason their way out of errors. RAG is still useful but it's not the same.

>>110181133
Exactly. 128GB will be down to $600 early next year.

>>110181129
Get 128GB while you can anon, best choice I ever made to run GLM flash IQ2.

>>110181170
what set up do you use to run glm?

>>110181170
>IQ2
why do people love running these ultra cope quants
just run a smaller model bruh
iq2 is lobotomized and no good

to think now being a vramlet matters less than being a ramlet
fuck I'll buy either and then tomorrow they'll switch to some new cpu based shit

>>110181170
>while you can
It's not biodegradable Anon. They're still launching DDR6 next year. It'll start showing up more on the used market as people upgrade, and pressure from chinkram won't go down.

Anthropic literally stated in their S-1 they're going to go hard on capex next year, basically as soon as their IPO. Their capex has effectively been 0 since the company was founded. The reason why they're going to build their own datacenters is because they've awkwardly got themselves into a situation where their hyperscaler compute deals have tied them up with their competitors (Google and Musk) who try to compete with frontier AI and have more compute. All Anthropic's current shareholders are going to liquidate as soon as they go public because they know Anthropic have no choice but to go hard on capex (and because they literally said they would in their S-1, basically admitting it's going to be a very different company to what it was H1 2026). Their financials are pointless because they describe an old version of the company that purposely held back on capex for as long as they could to get the best-looking S-1 possible to fool retail investors into thinking they're about to be GAAP profitable next year. The bubble will pop.

>>110181192
It's literally coomquants. You're just getting the expanded literary repertiore.

>>110181170
>GLM flash IQ2
which one exactly? maya?

>literally, literally, literally, literally, literally, literally

lalalalala~

>>110181210
Short it then and be rich

>>110181210
Local models?

>>110181226
quintessential humanslop
i'm surprised models don't do it more

>>110181226
failed the mental tongue twister
a good agent harness would've caught this

>>110181182
>>110181223
maya, the m quant originally but now I use s24 for speed since I noticed no difference
>>110181192
a 321b at 85-90% accuracy, will always be better than a big quant of a much smaller model... bigger models also quant better than small models, I have had no looping issues and have even used it for coding just fine. In fact an IQ2 of a 300b may retain as much accuracy as a Q4 of gemma for example

>>110181241
Your qwen model running right now is a dribbling stinky girlfailure distill

I have 220 combined ram + vram. What's the best local model for cooming these days? GLM flash? Used gemma4 in the past which was aight.

Why can't we just make SSDs as fast as VRAM? Then no one would be a vramlet and you could run anything local.

>>110179535
32K context kinda small on this omni version though

>>110181329
if it was as fast as vram then AI labs would buy it all up for the coming decade, and you would still have the same problem

>>110181239
I will and I will rub my rich all over your face.

>>110181338
We need to put 10,000 thumb drives in RAID 0

>>110181329
You'd have to give it 300W+, just as many processing cores to work with, then something to cool it down, and oh, look at the price...

>>110181329
ask the last who tried, intel with octane

>>110181359
The fun thing being that if they had the cash to continue that, they'd probably drown under the demand for these products.

File: ComfyUI_temp_hsoxb_00041_.png (2.14 MB, 1152x1472)

2.14 MB PNG

How do I prompt a llm to tag in booru tags without making shit up?

>>110181391
tool call to a booru tags database

>>110181336
eh, it's a no-think so it's not as bad, but no not great

>>110181391
mcp to load booru + ability to fuzzy search it

mayan anons, what speed do you get with deepseek flash 1017?

>>110181400
I'm not giving her internet access.

>>110181417
good thing I wasn't suggesting that

>>110181391
https://github.com/eduardoabreu81/sd-webui-tagcomplete-neo/blob/main/tags/danbooru.csv

>>110181398
Gemma 4 E2B had bigger context and could do audio and video at the same time, but you have to send the audio file and video file to it as separate files in the same request, llama.cpp has no way of doing that itself and there was no model-native way of doing it either

>>110181417
copy the tag database locally
vibe it, your an adult.

>>110181429
What about my adult?

File: 1787197471778535.webm (2.55 MB, 1280x973)

2.55 MB WEBM

>>110181090
Dont get memed. realistically its 2030 at the earliest.

>>110181415
with xs I average 25tk/s on 24gb vram

File: msgk_gemma_m.png (652 KB, 1280x498)

652 KB PNG

>>110180706
See >>110180949

File: 1779694178283145.jpg (206 KB, 2048x2048)

206 KB JPG

>>110179422

>>110180400
The ISTA meme quant was updated 15 minutes ago to work with upstream

File: 1768487789853084.png (469 KB, 675x900)

469 KB PNG

>>110181488
*crunch*

>>110181482
that's pretty ok, thanks anon

>upgraded my vram
>spent all day using h3 to generate lewd edits of tiktok whores and cosplayers
>nothing productive got done
this is how ai will kill us all

Any fellow retards buying a Mac tomorrow?

>>110181506
are there worthwhile nsfw finetunes of h3 nowadays? I hate playing with a billion loras for each use case

>>110181415
>dsv4.1 2.9bpw 128K context
45~ decode
600~ pp
128GB/128GB

PP speed is a brutal downgrade from 2600~ with glm 5.3 flash @ 4.10bpw

I think that gaki gemma is lame. Cute shy gemma is my goat

>>110181210
>capex
They also have MASSIVE, uncancellable obligations to said hyperscalers. If demand wavers they're fucked, to use the technical term.

File: 1774829546859614.jpg (21 KB, 270x360)

21 KB JPG

>>110181458
>hand color

>>110181484
Catbox NOW

>>110181520
I don't use any loras or finetunes, still learning how to proompt. apparently there's some sysprompt or skill file that you can load into a llm to get it to rewrite a vague prompt into a better one.

>>110181130
agreed, we've all seen it & understand
that post is just ot pbait shitposting
don't feed the trolls

>>110181524
pp should be way faster than that with the special snowflake architecture of dsv4.1, do you run pcie 3.0 on your gpus? could be the problem...

>>110181575
probably 2 of those unlocked crypto cards @ pcie2

I wonder how much ram a world model like genie uses

>>110181536
Aru wouldn't say that

>>110180279
https://vocaroo.com/14wGpi7wDOsD

>>110180553
cant we stop using aime
the fucking thing is beyond saturation at this point

File: 1774423355632630.gif (1.46 MB, 498x281)

1.46 MB GIF

>no good LLM since 3.8
>only t2i had good releases

>>110181520
Sulphur H3 is out on november

is claude code actually a good harness?

>>110181645
is mayonnaise actually a good harness?

>>110181615
this is ass

File: 1766130148258075.jpg (143 KB, 1080x1376)

143 KB JPG

>>110181090
Being this retarded should be studied

>>110181645
All you need is dsh.

>>110181657
how fat is the default token usage. pi is basically nothing

>>110181657
you need to gsh

>>110181645
>is claude code actually a good harness?
i use it, and pi
20k tokens for the cc system prompt
and you have to disable all the telemetry etc

>>110181359
Those retards shut it down right before the AI boom lmao.
But anyway inference on NAND already has prototypes, the model is on several nand channels while the kv is on lpddr5.

>>110181680
>20k tokens for the cc system prompt
>and you have to disable all the telemetry etc
dsh is under 5k tokens and you don't have to disable telemetry

>>110181690
>you don't have to disable telemetry
dsh is great but this is a like

>>110181690
i heard that and was going to try it
but i gotta learn "open code" for work now and can't keep track of too many tools at once

>>110181698
lie

>>110181703
What kind of retarded workplace forces a choice of harness on its employees? Does your work also force you to use neovim? Do they force you to use a preselected set of plugins too?

>>110181712
>What kind of retarded workplace forces a choice of harness on its employees?
Basically any company of a certain size or age
>Does your work also force you to use neovim? Do they force you to use a preselected set of plugins too?
This is way too close to the truth at a lot of places. You're going to make some anons cry from PTSD (because the tools they are forced to use are no where near as nice as neovim)

File: 1773749778631998.png (26 KB, 224x224)

26 KB PNG

I went to the farmers market, I hadn't stepped outside in months

real people feel like bad abliterations of quantized models

I don't know if I'm going insane or some of the basins LLMs end up in mirror some of the failure modes of the human psyche

>>110181684
Intel has an amazing track record of cancelling stuff before it's wanted. Optane and Itanium would've payed out well if they didn't get cold feet so fast.

>checked gpu and ram prices
The gates to Valhalla have closed. If you don't have hardware now you're fucked.

>>110181712
>workplace
Imagine working when you can just vibecode everything.

>>110181712
we had vim and no plugins, so yeah. thank god it was vim and not vi

Is scotoma good for coding/tools?

>>110181712
Where I work it's basically Codex only. Claude Code is banned and IT has never gotten around to approving anything else. Oh I guess you can also use antigravity if you really want to, lel

>>110181761
anon... it's over

>>110181761
What are the failure modes? When I touch grass everyone is polite and kind, completely different from the internet.

>>110181771
The Christmas season sales will save me, i was extra good this year.

>>110181810
>-5%

>>110181771
I'm gonna vibeslop my way out of poverty and then I'll be able to afford all the RAM I want

>>110181712
it is what it is
people in charge of those choices are often fuds who got hyped for whatever reason

>>110180147
Sorry, I've been meaning to submit those cloud model results as a point of comparison. Been still running them in the background but I have no Fable or Opus sadly that matter much. I just need to find a weekend to do the proper things to get it upstreamed. They are older profiles but I think it's more than indicative enough for local for some time at least at the smaller sizes. If I ever get a proper rig set up, I can probably run some of the bigger models for several days to get results for them to see if we're any closer there.

>>110181508
yep
one, maybe two

>>110181761
>real people feel like bad abliterations of quantized models
I've heard people say this but haven't experienced it myself.

what are someanime girls that would make for a good agent? besides gemma

File: 1778718302701605.png (740 KB, 1426x1953)

740 KB PNG

>>110181906
DESU

File: 1760922007.jpg (168 KB, 1920x1080)

168 KB JPG

>>110181906
Alma-chan

>>110181970
hmmm that might be fun actually

>>110181761
>farmer's market
Probably old people who spend all day getting propagandized by Facebook and talking to chatbots in voice mode.

File: miku-meeting.png (464 KB, 600x630)

464 KB PNG

>>110181906
miku

Currently feeling Ponnish

File: 1764358881599548.jpg (98 KB, 1280x663)

98 KB JPG

>>110181906
Chaika
I want LLMs to speak like Chaika

File: 1776900801072032.jpg (25 KB, 228x221)

25 KB JPG

>Turn off Dynamic VRAM
>Wont gen at all
Fuggggggggggggg

File: dying-miku.jpg (164 KB, 811x608)

164 KB JPG

i miss miku

File: image.png (3.87 MB, 1920x1080)

3.87 MB PNG

>>110181906
Ciel/Raphael is literally Rimuru's local agent, and should not be too hard of a speech pattern to emulate, but actually implementing her autonomous behaviour quirks could quickly turn into a nightmare.

>>110182060
You can probably ask LLMs to speak like Chaika. Maybe. I'll go find out.

>>110182129
That's basically grug reasoning but for output.

File: file.png (24 KB, 523x613)

24 KB PNG

>>110181823
Forgot to mention but yes, at some point, reasoning became non-optional for models over API and that is the case for older Gemini and Claude models when I started benchmarking and so after GPT 5.6, the same thing happened over OpenAI's side as well. I did the same thing and ran the test on low reasoning when it was non-optional. Also a preview of what I was able to test so far of things.

File: 1765466991271610.gif (434 KB, 369x518)

434 KB GIF

>>110182136
>>110182129
Chaika's speech is a bit more complex, and funny, and CUTE

File: file.png (220 KB, 1572x1156)

220 KB PNG

idk how simplicity and self-contained portability came to be something very hard to understand
is making it portable like llama.cpp that hard?
or is dev just a pure nigger

Jannies, kill this nigger already.

>>110180897
All we need is mesugaki test

>>110182156
just launch the engine directly with the launch scipt like the early version and ignore this.

>>110182161
me?

>>110181484
this is miku under gemma's skin!

>>110181488
>1488
omg based migu??

>>110181488
onamiku

>>110181534
My assumption was those obligations where why the IPO is going forward at all.

File: image-1793121517877.png (207 KB, 1500x920)

207 KB PNG

Discovered something in Gemma4-31b-it
L50 N167 appears to be the Mesugaki neuron

File: Screenshot_20261027_211447.png (3.92 MB, 3299x1686)

3.92 MB PNG

>>110182391
wrong pic, that was a prototype