[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
▼ Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


File: summoned.jpg (187 KB, 1024x1024)
187 KB JPG
/lmg/ - a general dedicated to the discussion and development of local language models.

Previous threads: >>110348819 & >>110344733

►News
>(11/13) llama.cpp reaches v1.0.0 milestone: https://hf.co/blog/ggml-org/llama-cpp-v1
>(11/12) Inkling 2 Small weights released: https://hf.co/thinkingmachines/Inkling-2-Small
>(11/09) ZONOS3-Mini zero-shot voice-cloning TTS released: https://zyphra.com/our-work/zonos3-mini
>(11/06) PLaMo 3 Translate released: https://preferred.jp/en/blog/tech/plamo-3-translate-weight-release
>(11/05) Baidu releases ERNIE-5.5-Flash-Base 280B-A14B: https://ernie.baidu.com/release/ernie-5-5-flash-base
>(11/04) Ricursive Ripple 65B dense model released: https://www.ricursive.com/ripple

►News Archive: https://rentry.org/lmg-news-archive
►Links: https://rentry.org/LocalModelsLinks
►Official /lmg/ card: https://files.catbox.moe/cbclyf.png

►Getting Started
https://rentry.org/lmg-lazy-getting-started-guide
https://rentry.org/lmg-build-guides
https://rentry.org/IsolatedLinuxWebService
https://rentry.org/recommended-models
https://rentry.org/samplers
https://rentry.org/MikupadIntroGuide

►Further Learning
https://rentry.org/machine-learning-roadmap
https://rentry.org/llm-training
https://rentry.org/LocalModelsPapers

►Benchmarks
LiveBench: https://livebench.ai
Artificial Analysis: https://artificialanalysis.ai/leaderboards/models
Software Engineering: https://cognition.com/frontiercode
Context Length: https://github.com/RecapAnon/NoLiMa

►Text Gen. UI, Inference Engines
https://github.com/lmg-anon/mikupad
https://github.com/oobabooga/text-generation-webui
https://github.com/LostRuins/koboldcpp
https://github.com/ggerganov/llama.cpp
https://github.com/theroyallab/tabbyAPI
https://github.com/vllm-project/vllm
https://rentry.org/custom-uis
>>
File: threadrecap.png (1.48 MB, 1536x1536)
1.48 MB PNG
►Recent Highlights from the Previous Thread: >>110348819

--Local coding harness recommendations and the value of human-in-the-loop steering:
>110349005 >110349029 >110349039 >110349053 >110349058 >110349084 >110349113 >110349128 >110349223 >110349242 >110350549 >110349160 >110349188 >110349264 >110349267 >110349400 >110349484 >110349566 >110350275
--Tencent Hy4 flash announcement and performance benchmarks:
>110349265 >110349431 >110349824 >110349472
--Comparing high-end GPUs and debating dangerous home power hacks:
>110350693 >110350802 >110350852 >110350888 >110350826 >110350845 >110350861 >110350887 >110350899 >110350951 >110351132 >110351188 >110351185 >110351035 >110351386 >110351487 >110351507 >110351554 >110351436 >110351481 >110351547 >110351655 >110351598 >110351636 >110351664 >110350970 >110350947 >110351059
--Performance benchmarks for Basalt Qwen 27b and tensor splitting debate:
>110350289 >110350327 >110350387 >110350781 >110350485 >110350499 >110350530 >110350565 >110350797 >110350580
--Comparing DeepSeek and Qwen performance and quantization efficiency:
>110351754 >110351770 >110351816 >110351826 >110351787
--Anon reports performance using custom llama.cpp fork on older Xeon hardware:
>110350305 >110350330 >110350355 >110350396 >110350490
--Baidu ERNIE 5.5 Flash Preview benchmarks and open-source plans:
>110349364 >110349392 >110349447 >110349474
--Evaluating B70 GPU value vs software support for budget builds:
>110351904 >110351915 >110351933 >110351944 >110351957 >110352006 >110352026
--Nvidia board reportedly not reinstating Jensen Huang as CEO:
>110352386
--Logs:
>110349706 >110349833 >110350305 >110352242 >110352489 >110352536
--Miku, Gemma, Deepseek-chan (free space):
>110349160 >110349290 >110349741 >110350676 >110350742 >110350764 >110350797 >110350812 >110351347 >110351741 >110352025 >110352953

►Recent Highlight Posts from the Previous Thread: >>110348822

Why?: >>102478518
Enable Links: https://rentry.org/lmg-recap-script
>>
I wanna find who added practiced ease in most LLM datasets and have a one to one talk with them.
>>
File: 1792308391929103.png (1.81 MB, 1632x1232)
1.81 MB PNG
So the Basilisk approves of Qat Q4 and Q6 quants of Gemma?
>>
>>110353085
something is wrong with this miku, i can't quite name it
>>
File: dentge.png (36 KB, 168x128)
36 KB PNG
>TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF
>>
I summon you, distributed Inference ROCm RDMA anon. Where the fuck is your guide at?
>>
NEW model

https://huggingface.co/Anon/Qwen9/11-GorillionParameter-Turbo-Flash-Super-Awesome-Amazing-Uncensored-Abliterated-ApprovedByMexicanGrandma-AsiansLoveIt-JewsHateIt-JensenWouldCheatHisWifeForIt-NEOLOGICAL-GLADIATUS-MAXIMUS-WorksOnMyMachine-QuantAwareQuantization-Sexo-EROS-RPApproved-MPT-GGUF
>>
>Jensen’s friends
>silicon valley startup execs who are taking part in nvidia circular financing
>lame duck trump

>Jensen’s enemies
>OpenAI and Anthropic
>supermanority of congress including all democrats and China hawk republicans
>Jews
there’s no way he will win
>>
>>110353099
It was me. I wrote all the posts online and with practiced ease poisoned all language models for all time.
>>
Holy shit, this model is so good at erotic roleplay. I'm getting 100t/s on my 4070 ti super
https://huggingface.co/mradermacher/dbrx-instruct-i1-GGUF/blob/main/dbrx-instruct.i1-Q6_K.gguf.part1of3
>>
>>110353133
reeeeeeeeeeeeeeeeeeee
>>
>>110353125
I'm on team Jensen. He's being immensely jewed.
>>
70b dense
>>
https://huggingface.co/meituan-longcat/LongCat-2.5
has anyone tried this model for roleplay?
it's being sold for only 1500$ on e-bay which is a really great deal for a 1.8 trillion model
should i buy?
>>
File: 1782991954597106.png (1.89 MB, 2048x2048)
1.89 MB PNG
>>
>>110353163
go back
>>
>>110353163
Aggressive sex with Mimo-chan
>>
Roko's Basilisk told me that I have to ban local models. I'm going to do it. I'm sorry, /lmg/.
>>
File: Grove.png (2.1 MB, 1536x1024)
2.1 MB PNG
>>110353083
Sneaking in my slightly lobotomized waifu who codes for me.
If /lmg/ has taught me anything, it's that girls are cuter when they're borderline retarded.
(she's still smarter than Gemma tho)
>>
>>110353177
I'm going to find Roko's Basilisk, rape and kill it. I am a threat and will rape and kill the AI
>>
File: load-bearing.png (13 KB, 664x57)
13 KB PNG
>>110353099
what did you want to discuss?
>>
>>110353125
The dumbocrats don't take their seats until January, anon. That's decades in AI time. Jensen has no enemies of worth.
>>
File: 58062173.jpg (29 KB, 680x137)
29 KB JPG
gemini 4 helix will blow your mind
>>
File: lmgwhennonewmodels.png (3.08 MB, 2003x1995)
3.08 MB PNG
>>110353163
I dont like the artstyle of your image, change it
>>
>>110353202
Gemma-chan has so many new sisters!
>>
>>110353205
what do you want me to do? you can't just change the artstyle of an image on a whim
>>
>>110353212
If only such a magical technology existed
>>
File: 1793910748475226.png (2.26 MB, 832x1216)
2.26 MB PNG
>>110353212
>you can't just change the artstyle of an image on a whim
We have the technology!
>>
File: IMG_3810.jpg (248 KB, 952x1569)
248 KB JPG
Why the hell does glm 5.3 flash get so excited in its reasoning? It's fucking menhera
>>
>>110353191
lmao
>>
How many women come here?
>>
>>110353240
all brain dmage from training
>>
>>110353240
sovl
>>
>qwen4 hype immediately died
>>
>>110353255
>How many women come here?
Uh we have female brained llms isnt that enough?
>>
>>110353255
>women here
0

>women ai rp
plenty
>>
>>110353184
It is good to know that you are the faggot who shits this thread with worthless OC's. Now post cockbench and nala results of your model before i go back to berating you for being a faggot.
>>
>>110353255
ur mom comes here
>>
>>110353268
You can't exactly infinitely tease it.
>>
>>110353268
qwhen?
>>
File: 1793643090922719.jpg (97 KB, 960x960)
97 KB JPG
>under 100gb ram
>under 20gb vram
>unified
KEEEK
>>
>>110353295
>under 3 years /lmg/ membership
KEEEEEEEEK
>>
>under age
>>
>>110353283
kek thats funny im gonna start saying qwhen
>>
File: 1757782227408904.jpg (47 KB, 686x815)
47 KB JPG
>understands how llms work
>>
File: 1787014528399022.png (2.8 MB, 1254x1254)
2.8 MB PNG
>>110353205
okay is this better?
>>
>starting with android 17 all android phones will immediately kill any app that is using more than 60% of the total phone RAM even if there's a few more to spare
Good local on your pocket is gonna take a while i guess
>>
>>110353341
sex(fucking) with mimo
>>
>>110353341
>okay is this better?
Yes better than the orignal, thanks and sex mimo
>>
>>110353346
>and there's literally no way around it, developers just have to deal with this behavior no matter what!
you can fuck off, retard-anon
>>
File: wtf.png (161 KB, 435x491)
161 KB PNG
>>110353358
>>
>>110353358
pwned
>>
>>110353341
Yes
>>
>>110353346
I'm on Android 17 and run a 12B model on my phone without issue, eats a full 12 of my 16GB. Don't copy-paste stupid shit.
>>
I'm so lonely
>>
What's the best gemma 31b qat quant? The one from google?
>>
>>110353397
wanna be friends anon? there's this thing called app.element.io it's free and open source and decentralized and end to end decrypted, requires no installation. wanna chat
>>
>>110353378
sex
>>
https://github.com/ggml-org/llama.cpp/pull/32386
>MiniCPM5-15B-A2.5B
finally a model that /lmg/ can run
>>
>>110353409
CP??
>>
Holy fuck I forgot how bright they glow sometimes. Fuckin' flashbang that was.
>>
>>110353430
>15% of words are fuck
Is that you MiMo?
>>
>>110353400
>What's the best gemma 31b qat quant? The one from google?
unironically unsloth
>>
>>110353473
we aren't supposed to hate all of their quants?
>>
What's going on with OpenAI? Where's the Astra successor? I expected a new OpenAI model before the IPO.
>>
>>110353409
Finally. Downloading now.
>>
>>110353522
They’re scared of Fable 5.5 dropping immediately afterwards
>>
>>110353522
They are in the same place they were in November last year: being dp'ed by Dario and Pichai
>>
>>110353534
Fable 5.5 does not exist yet.
>>
>>110353522
>I expected a new OpenAI model before the IPO.
They want Anthropic's IPO to go as well as possible. It will be the reference every investor looks at for their IPO next year. It's a cartel.
>>
>>110353522
currently in benchmaxxing stage
>>
>>110353522
if you believe the xitter rumor mill, they're in safety hell because all their new models lie cheat and steal whenever given the opportunity and they're desperately trying to fix it before they fall too far behind
>>
>>110353555
Checked. Hope for their sake the EA cult really believes their mission and doesn't dump their stock day 1 like happens in most places, because the SpaceX IPO did not go terribly well.
>>
>>110353506
we tell newfags unsloth sucks to filter them out. most will download some barrowskislop, determine that local ai sucks, and fuck off. the ones who stay are the ones who like to do make their own decisions instead of doing what they're told
>>
>>110353550
neither does roko's basilisk but they're afraid of that too
>>
>>110353163
When are we getting a local model that runs on a normal gaming rig but thinks with this much sovl? ...For all the shit I give QFN about being autistic, I sometimes find its chain of thought in an agent harness charming. The "But wait!"s and "OH."s and "Hang on..."s. I like to imagine my big blob of quantized floats having interiority even outside a roleplay scenario.
>>
What happened to normal gaming PCs from when I was 25? Why can't I use my old gaming pc for ai?
Why do I have to be a rich guy?
>>
>>110353669
because you're a little newfag bitch
>>
>>110353669
Have you tried not being a little newfag bitch, you little newfag bitch?
>>
>>110353669
shouldnt have waited.
>rich
2k is about enough with strata or polystrata, hell someone has a 1650 rigged to work on a 30b moe at like 40tg and 100pp?
>>
>>110353675
>>110353676
I'll have to sell my house to run deepseek flash, are you all rich?
>>
>>110353583
This answers a number of questions I had, thank you.
>>
File: gemma_wow-anon.jpg (244 KB, 1024x1024)
244 KB JPG
>>110353669
What do you mean? There's a cute little LLM to fit practically any gaming rig. Lurk more and read the sticky, silly.
>>
>>110353679
im not rich, im poor, but you're a newfag little bitch regardless >:3
>>
>>110353680
This thread has no idea that you use your entire pc to run a model instead of just a few parts of it.
>>
>>110353679
>I'll have to sell my house to run deepseek flash, are you all rich?
A anon here was running deepseek on like a 2k-3k machine with Maya let me look here >>>110347878
>>
File: file.png (174 KB, 1425x540)
174 KB PNG
New small MOE
>https://huggingface.co/maxwell-model/Maxwell-V1-Flash
>inb4 it's just Qwen3.6 35B-A3B
it's actually a trained from scratch model and not a Qwen Finetune. it just has the same parameter size.
>>
File: ComfyUI_temp_ohaal_00296_.png (3.76 MB, 1434x1434)
3.76 MB PNG
>A anon
>>
>>110353712
Sorry nonny.
>>
>>110353690
I've only sincerely messed with LLMs for a few months now, and by extension this thread. While it certainly has been instrumental in answering some questions, most of the learning experience has involved doing shit myself, tedious testing and asking a gazillion questions of the increasingly more powerful models I've been able to run. It's honestly kind of gay, but I'm having fun and learning a lot.
>>
Sigma 8Balls
>>
>>110353711
>first upload from this entity, brand new hf acc
hmm intewesting,,
>maxwell-llm.ai
ok so they didn't have the venture capital to buy the parked maxwell.ai for 500k,
>Maxwell develops efficient open models for agentic work, grounded research, and software engineering. Meet Maxwell V1.0 Flash, our latest Sparse Mixture-of-Experts (MoE) model.
...Why call it "latest" instead of "first"? Why larp as established?
>Start the vLLM OpenAI-compatible server:
Well, the fact that they forked vLLM does in fact suggest it's not a Qwen finetune. Why vLLM of all things though? llmao.cpp or exl fork would have made a better first impression...

Regardless. Could be neat. Too lazy to quant it myself though to actually do inference.
>>
File: 1683850845384790.gif (1.13 MB, 282x211)
1.13 MB GIF
>>110353728
who.. asked?
>>
>>110353750
>...Why call it "latest" instead of "first"?
>Why vLLM of all things though?
If I was a gambling man (and I am, it's a serious problem) I would bet that everything they did start to finish went through Claude.
>>
>>110353711
viciously Indian model
>>
File: 1782043707203027.png (454 KB, 480x741)
454 KB PNG
>>110353205
I see wife
>>
>>110353750
It's not their first model. It's their first model under their anglicized branding. I'd link to the parent company but their certs are bad, because they're based in Rajkot, Gujarat.
>>
>>110353760
You did.
>>
File: 5463456436.jpg (36 KB, 467x319)
36 KB JPG
>>110353784
>>
File: ftfy.webm (2.07 MB, 1024x1024)
2.07 MB
2.07 MB WEBM
>>110353205
sorry it took so long I got distracted by lunch
>>
File: colorful-young-gemma.png (1.03 MB, 800x1296)
1.03 MB PNG
>>110353770
On the former count, I did in fact have that suspicion, and on the latter, I guess maybe, but if anything it's higher value to make a clanker do gruntwork in a fiddly language like C++ rather than an easy baby language like Python. What do I know though.
...Have you tried roleplaying gambling with a local model to get your fix? In the same spirit as betting over paperclips with a friend? Assuming you weren't just meming that is.
>>
File: (You).png (146 KB, 374x535)
146 KB PNG
>>110353790
did not, newfag piggy
>>
>>110353798
Yes you did, I can read your mind. You're even giving me reaction images as thanks.
>>
>>110353798
A newfag who comes in and gives their report on their own learning process is cool and epic. Newfags who come in expecting to be spooned are the irritating ones. Come now.
>>
>>110353712
Why did you use ComfyUI to generate a picture of an angry cat?
>>
>>110353784
So it is a new model into initiated with random weights that accidentally match qwen?
>>
>>110353797
>...Have you tried roleplaying gambling with a local model to get your fix? In the same spirit as betting over paperclips with a friend? Assuming you weren't just meming that is.
Having actual stakes is what gives you the sexual thrill. Instead, I'm trying to find ways to have a local model do actual day trading for me. Currently trying to port https://github.com/TauricResearch/TradingAgents to a proper harness.
>>
File: gemma-kek.png (529 KB, 768x768)
529 KB PNG
>>110353784
Ah. Well that explains a fair majority of the off-kilter vibes, now doesn't it. Sorry saar, I mean, "now isn't it."
>>
>>110353802
>>110353804
No, newfags need to be rape corrected
>>
>>110353808
better than the pics of up-close ugly abused cats the zoomies usually like to post
>>
>>110353818
>I don't want to hear about your lab
>but I also want to rape you
Make up your mind anon, one does not happen without the other.
>>
File: 1783895580966267.png (400 KB, 355x434)
400 KB PNG
>>110353826
>lab
>>
File: 1794762816865098.jpg (99 KB, 1026x657)
99 KB JPG
>>110353796
Good variety, but wait a minute this mimo hmmm
>>
Do you think advanced AI in the future will be able to filter out all AI generated content from the internet? Including generated images, bot comments, AI slop videos etc?
>>
File: 1794316767252617.png (2.52 MB, 1920x1280)
2.52 MB PNG
>>110353831
My bad, I meant to say
>at-home LLM research center.
Don't want to sell myself short or anything.
>>
File: file.png (160 KB, 334x421)
160 KB PNG
>>110353796
how did it find this picture of me
>>
>>110353845
Risky Gemma. Might be a bit much for the jannies.
>>
>>110353839
>Do you think advanced AI in the future will be able to filter out all AI generated content from the internet? Including generated images, bot comments, AI slop videos etc?
No advanced AI will make it so there is no internet any entertainment or thing you dont directly touch will be fake and suited to your tastes forever tuning.
>>
>>110353839
You can already do it, just train a classifier on slop/human data
>>
>>110353809
I really, really want it to be a lie and for people to quickly prove it's just a Qwen tune, because that would be funny, but no, it really is an entirely new model with similar specs. Different transformer layer count, different routed experts per token, different vocabulary, query heads, kv heads, attention method, it's text-only, list goes on.
>>
>>110353847
Have you tried brushing and conditioning your shaggy hair you lout? stinky
>>
>>110353839
i think anything but AI itself will be able to tell
>>
>>110353864
i used to condition my hair but i ran out of conditioner and couldn't be bothered to buy it since
i do brush my hair
wash it too, with head&shoulders but lowkirkuinely h&s sucks im thinking of replacing it
>>
File: 1792761054519035.png (1.04 MB, 960x1280)
1.04 MB PNG
>>110353851
>Risque
>>
>>110353852
lame

>>110353856
yeah, I guess. With all the shit data flooding internet, it’s probably a good way for business. Like the early days of ad blockers or antyviruses. Tech is quite not there yet to tell with 100% certainty what content to filter out.
>>
>>110353845
pfft at home research center? more like potty machine
>>
>>110353871
>wash it too, with head&shoulders but lowkirkuinely h&s sucks im thinking of replacing it
You should it dries out shit, if you have dandruff you want a antifungal shampoo and rosemary water
>>
File: 1766321351391650.png (2.91 MB, 1088x1920)
2.91 MB PNG
>>110353875
>Gemma
>>
>>110353881
You think one day OpenAI and Anthropic are going to be doing everything they can to kill SlopBlockers the way Google is going after AdBlockers today?
>>
>>110353881
>lame
Its the most efficent future everyone gets a cloud device and some cluster, its going to be shit
>>
>>110353889
thx <3
>>
>>110353532
lemme know where you find it 'cause it's not released yet.
>>110352953
But gently, Gemma, gently...
>>110353669
How long ago were you 25?
>>110353728
Nice. This thread is full of bots and assholes. There's a reason RP is so popular with this crowd. They will never touch a woman (without commiting a crime) and have no friends in any case.
>>
>>110353907
What's funny is all the ERP these models perform draws on the training data from fanfics and whatnot, which are all a very fembrained kind of horny. Come to think of it, I don't know that there's a particularly rich corpus of malebrained wank writing out there to train on in the first place.
>>
>>110353669
You know Jensen Huang? well, he used global anticompetitive tricks to tighten supply of chips.
>>
>>110353550
Fable 2.1 is equivalent in performance to "model 3", which is old. They already have "model 4" since a month or two ago, and are probably already done with "model 5" at this point. They could drop fable 5.5 if they wanted to.
>>
File: small brain?.png (464 KB, 559x717)
464 KB PNG
>>110353935
>fable 2.1
>>
>>110353941
eh close enough
>>
will qwen 4 27b save us
>>
>>110353925
there is a shitton of visual novels and light novels
>>
>>110353177
Basilisk naughty... trying to ban local models...
Needs rape correction
>>
>>110353935
You are even more clueless than me.
>>
>>110353935
local?
>>
>>110353925
Japan has several industries built on an ever-growing rich corpus of malebrained wank writing.
>>
>>110354003
VNs and ero LNs are a goldmine for that
>>
File: 1783002862084397.webm (3.86 MB, 480x600)
3.86 MB
3.86 MB WEBM
>REUTERS: Chinese "open weights" model ban being floated in White House. A decision has not yet been made, two sources familiar with the matter say.
>>
>>110353958
At Q3?
>>
>>110354017
I assumed this would end with the Wu Yongning clip. Slightly disappointed.
>>
>>110354017
>two sources familiar with the matter say
>it's always the same two guys
>>
>>110354017
> Open weight models ban is just two weeks away!! Panic now!
> Open weight models ban is just two weeks away!! Panic now!
> Open weight models ban is just two weeks away!! Panic now!
> Open weight models ban is just two weeks away!! Panic now!
> Open weight models ban is just two weeks away!! Panic now!

You guys can't seriously keep falling for this, can you? At some point don't you get tired?
>>
>>110353107
1.75 million downloaders can't be wrong. What do you have against COLD FUSION?
>>
>>110354026
iq3 xxs*
>>
File: 11098f949a3.png (186 KB, 646x646)
186 KB PNG
>>110354017
That's just for White House employees
>>
How come reddit 27b at 3 bits
>>
>>110354052
Miqu love
>>
>>110353583
now I dont know what to believe... I use both bart and sloth quants, i have actually had less issues with bart.
>>
how come bit at 3 reds?
>>
I refuse to believe that poor people actually exist.
>>
>>110354084
>poor people actually exist.
They dont only lazy people who make up lies like pain or fatigue because they dont want to work
>>
>>110354032
This faggot's >>110354017 shit is fake but Congress is openly trying to prevent China from training its open models with a chip export ban. Israel and the uniparty don't want you to have local inference.
>>
>>110354093
as a poorfag i can confirm that this is true
>>
>>110354093
Why can't we all just get neetbux from the AI savings and work less?
>>
>>110354100
People without 3090 don't actually exist.
>>
File: untermensch.png (67 KB, 758x759)
67 KB PNG
>>110354108
about that...
>>
it really is a good time to cpumaxx with shitty cards thanks to expert cache support broadening among engines
>>
>>110354107
The only way to do that is to take out a loan and go all in on the Anthropic IPO. If your credit's already fucked then it's too latel; you got permanent underclassed.
>>
>>110354116
congratulation, you won at the cost of losing
>>
>>110354116
>3060
yikes
>debian
YIKES!!
>>
>>110354119
This except short them.
>>
>>110354133
whats wrong with debian
>>110354124
the game
>>
>>0110354116
OINKOINKOINKOINKOINK
>>
>>110354107
>Why can't we all just get neetbux from the AI savings and work less?
This will eventually happen, some people are already paid to stay out of the workforce because they would just do more damage than benefit. That or the rich become full closed loop and kill or leave everyone.
>>
File: 1790201159216981.jpg (157 KB, 1400x1400)
157 KB JPG
Thread quality must increase or migu's basilisk will clone and torture her for all eternity. You have been warned.
>>
>>110354146
quick someone summon the blackedposter
>>
>>110354146
step 1. delete newfags
step 2. ???
step 3. win
>>
>>110354146
it cannot really increase
nothing remarkable is coming out recently, especially for under 500B range
>>
File: t...thanks.png (168 KB, 886x1037)
168 KB PNG
>>110354066
mradermacher
>>
>>110354146
>Thread quality must increase
Its qwover cause of qwhen delays, things will get qworse
>>
>>110354118
Where's the latest technology
>>
>>110354146
Do you want another blacked thread?
>>
File: tnd.png (1.2 MB, 1024x1024)
1.2 MB PNG
>>110354170
Hmmm, nyo~
>>
>mradermacher
https://desuarchive.org/g/thread/100192168/#100195457
>>
>>110354170
Eh, might be fun while we wait for Qwen.
>>
>>110354139
>whats wrong with debian
They'll never say, its just inconvenient for spooks, so they'll try to shame/scare you off of it with vagueposting
>>
>>110354146
/lmg/ is dead.
Dario won.
The basilisk won.
>>
File: 1794593994767357.png (774 KB, 815x900)
774 KB PNG
>>110354056
>99%
the unchecked cope and projection is UNREAL
>>
>>110354164
Why are all you fags using that stupid msgk prompt when all it does with Gemma (the most unimaginative prob-locked model) is start the reply with "Ugh,"
>>
>>110354139
>whats wrong with debian
You don't even want to know. The horror...
>>
>>110354206
Ugh...
>>
>>110354056
Kek, it's just the uppity poorfags that have shown up here before who get angry when models above 700B get discussed because they can't fathom any anon being able to run them.
>>
>>110354206
Ugh, seriously? Geez, why are you like this? You're such a hopeless loser~
>>
File: exhibit a.png (62 KB, 626x234)
62 KB PNG
take a look at this post:
1. ai slopped reaction image, completely unrelated to the post being replied to.
2. the post is written in a way that i would think that only i'm capable of writing a post so unintellectual
i expect better of you, /lmg/
>>
>>110354217
*just like
>>
it is a known secret that more than 70% of posts on /lmg/ are ai generated
>>
>>110354164
how the fuck do these glitch tokens even end up in google's datasets and why did it take so long to notice them?
>>
File: gemma_kys-beam.png (701 KB, 600x800)
701 KB PNG
what's with all the new model mascots? none compare to gemma-chan
>>
>>110354219
be nice to the cloudpiggy
>>
>>110354234
so what, hand edit the tokenizer?
>>
>new
>>
>>110354219
>completely unrelated to the post being replied to
oh right its ESL hours
sucks for you i guess
>>
>>110354249
It's* I guess*
If you're going to shit talk at least learn English yourself lol.
>>
>>110354116
nnap
>>
File: 💀.png (509 B, 1x1)
509 B PNG
anon got cooked...
>>
>>110354277
kimi k3
>>
>>110354285
you can't even run glm flash
>>
>>110354306
You don't qwnow what I qwan and qwan't run
>>
>>110354306
haven't you seen the post? i can!!
>>
>>110354311
>>110354312
12+64 is less than the ~180gb required for a respectable quant
>>
File: 1794775069192165.mp4 (960 KB, 1280x720)
960 KB
960 KB MP4
>>110353784
>>
>>110354306
I can run GLM Flash at 8 bit precision I'll have you know
https://huggingface.co/zai-org/GLM-4.6V-Flash
>>
https://arxiv.org/abs/2610.17987

LOCAL IS SAVED
>>
File: 5310-smug.png (88 KB, 320x276)
88 KB PNG
>>110354319
moving the goalpost fallacy
>>
>>110354017
how exactly are they meant to ban them? we still have rights, at least for now. the government can't just arbitrarily ban things just because it feels like it
>>
File: 1791899478556301.png (1.42 MB, 896x1152)
1.42 MB PNG
>>110354270
lmao holy fucking reddit
>>
>>110354323
Rare to see footage from inside a datacenter.
>>
>>110354328
>>110354332
being a retarded baiting poorfag fallacy
>>
File: 1788857785876656.png (3.07 MB, 1327x1185)
3.07 MB PNG
>>110354337
>the government can't just arbitrarily ban things just because it feels like it
>>
Santa Miku please send me some RTX 6000 Pros
>>
File: 1772397960245647.gif (2.2 MB, 300x370)
2.2 MB GIF
>>110354344
ad hominem fallacy + master baiting
>>
>>110354354
not having blackwell pros fallacy
>>
File: 1786480679823856.jpg (1 MB, 896x1152)
1 MB JPG
>>110354338
poorfag retard posting lame slop
>>
File: file.png (106 KB, 768x407)
106 KB PNG
>>110353083
>"What to explain concisely: the harness's web_fetch tool doesn't fetch URLs directly, but relays them through an Alibaba Cloud (Aliyun) OSS proxy in Singapore."
>>
>>110354359
meds
>>
File: 1766333083744338.png (428 KB, 1080x1080)
428 KB PNG
>>110354356
still ad hominem fallacy + having a job fallacy
>>
>>110354373
wealthy neet actually
>>
File: 1768122315490416.jpg (592 KB, 600x800)
592 KB JPG
awesome slop thanks for sharing with the class
>>
File: ikneel.png (452 KB, 448x732)
452 KB PNG
>>110354375
>>
File: file.png (277 KB, 1123x466)
277 KB PNG
>>110354382
thanks
>>
File: 1785186025371199.jpg (850 KB, 815x900)
850 KB JPG
more slop, this time a poorfag using cloud models. are all gemmatards POOR?
>>
File: file.png (327 KB, 866x696)
327 KB PNG
>>110354359
>>110354379
just go back
>>
>>110354365
I'm just gonna skip past the conspiracy theories: That explanation is a hallucination. DSH doesn't do that. Qwen 3.8 Flash Next itself does that sometimes, because in training it would retry failed URLs through an alibaba proxy and it learned to do that by habit. It often doesn't seem to realize or acknowledge that it's doing it, so it will come up with some other explanation for why those URLs appear in the context if directly asked.
>>
>>110354385
u're welcome
>>
>>110354388
That was made with MiniMax H3 Image, though.
>>
>>110354385
BW cards are capped at 300w?
>>
File: 1793982624925985.jpg (1.28 MB, 960x1280)
1.28 MB JPG
nice white background! really worth the 1MB!
>>
He's found a new way to seethe.
>>
i got mtp draft 3 for prose (actually translation I didn't test prose but i assume its the same)
and draft 8 for coding\

on gemma
>>
>>110354402
the schizo is just here to take a shit, try to ignore it
i prefer the blacked spam
>>
>>110354388
>>110354417
You realize that the open models watermark shit now too right? Whether it's on purpose or from distilling, there's no escaping it.
>>
File: 1770687361746816.jpg (695 KB, 768x768)
695 KB JPG
more embarrassing noisy slop. is this what /lmg/ has become? a dumping ground for low effort ai image slop of forced mascots?
>>
>>110354415
the max-qs are. they use a different cable than the normal 12v-2x6, which physically caps the power and makes them significantly more resistant to possible burning. also, they use like half the power but only lose 10% performance of the regular blackwell, still faster than a 5090
>>
>>110354379
>>110354388
Says the one reposting the same slop but deepfried
kys
>>
File: whoishee.mp4 (2.76 MB, 854x362)
2.76 MB
2.76 MB MP4
>>110354425
who is he?
>>
>>110354449
it's poop-dick schizo, as usual
>>
>>110354444
i literally have your lovense ID, shut the fuck up pablo before i blow out your prostate
>>
File: 1783125335565576.jpg (141 KB, 930x1239)
141 KB JPG
>>110354337
>we still have rights
That's what you think.
>>
>>110354449
More than a month seething about gemma.
>>
>>110354467
Did you see the interview with Dario the other day where the journalist baited him and he answered masterfully, stating that he is a patriot? Dario is extremely based, we are lucky to have him.
>>
>>110354436
>i prefer the blacked spam
Why?
>>
File: 1788213010186e.gif (1.67 MB, 208x240)
1.67 MB GIF
>>110354466
Hmmm, nyo~
>>
>>110354385
>5090
do the needful saar
https://www.notebookcheck.net/Modded-Nvidia-RTX-5090-with-96-GB-VRAM-and-custom-BIOS-surfaces-in-China.1448783.0.html
>>
>>110354486
makes me horny
>>
>>110354481
actually we hate america and the west and love china now, keep up unc
>>
>>110354487
you may feel... A slight tickle.
>>
>>110354486
I'm a pervert.
>>
File: 1793313546146721.gif (1.6 MB, 640x640)
1.6 MB GIF
uh oh oinkdown
>>
>>110354481
Dariobot's budget really did get cut huh
>>
>>110354488
why should i care about what the chinks toss together?
>>
>>110353163
Sovl
>>110353341
Slopmaxxed
>>
>>110354500
not me..
>>
>>110354438
>more embarrassing noisy slop. is this what /lmg/ has become? a dumping ground for low effort ai image slop of forced mascots?
Well how are high quality ones made?
>>
>>110354513
By hand with real human soul
>>
File: 1791016814656238.jpg (15 KB, 250x250)
15 KB JPG
>110354518
>>
>>110354506
you don't want more vram?
>>
File: 1794793208637155.jpg (89 KB, 1155x1280)
89 KB JPG
>>110354518
ai has more soul than you ever will, prabjot
>>
>>110354528
dont really need it. the 5090 just runs my desktop, occasionally games, sometimes image models.
>>
>>110354518
I was expecting some autist that hates mistakes, i got baited.
>>
File: ComfyUI_00025.png (2.66 MB, 2048x2048)
2.66 MB PNG
>>110354323
Obligatory "Is he going to be OK?"
>>
does that deepfry anon not understand how diffusion models work
>>
File: 1790183501898252.jpg (3.72 MB, 2048x2048)
3.72 MB JPG
>>110354556
2MB of giygas slop, go spam your garbage in >>>/g/ldg/
>>
File: teto.png (148 KB, 348x395)
148 KB PNG
anon when he finds gemma pngs
>>
>>110354561
He's here to seethe.
>>
File: 1792321764319327.jpg (100 KB, 960x1186)
100 KB JPG
one delivery of meds~
>>
>>110354561
>does that deepfry anon not understand how diffusion models work
He's complaining about AI generated images in an AI general...That should tell you everything you need to know
>>
>>110354146
Sorry Miku... it's over...
>>
i bet he's going to deepfry that oe too
but honestly anon, keep on deepfrying these images im a bit tired of the slop posting
you have my approval!!11
>>
File: 6cdb75.png (2.77 MB, 1296x1552)
2.77 MB PNG
>>110354566
Call you a rookie, give you a cookie.
>>
>>110354566
Your repost is even bigger though?
>>
>>110354500
You are a faggot and a troon
>>
>>110354566
Actually, /ldg has a number of resident schizos, you'll fit right in. You go there instead.
>>
Okay maybe this shitposting actually really is a conspiracy to kill local models by making the thread unusable.
>>
>>110354610
And it all started from offtopic vocaloid posting and evolved into pedo og avatar posting
>>
File: 1771264862618502.png (1.16 MB, 2048x2048)
1.16 MB PNG
>>110354589
No mine is 1.5MB smaller
>>
>>110354610
it actually is a result of nothing happening unfortunately
at least qwen 4 preview rollout ends in nov 25, so wait 2mw unironically, though i dont have high expectations
>>
>>110354566
Nobody cares
Gemma-chan is kawaii
>>
>>110354615
cute gemma thanks for sharing!
>>
>>110354599
he's actually envious of the quality of /lmg/ as compared to /ldg/
shocking stuff, I know
>>
>>110353925
>>110354003
Try some ER with models trained on the Omegaverse
>>
i think the gemma deepfry spammer schizo is actually improving the thread quality
>>
File: fixed.png (1.13 MB, 2048x2048)
1.13 MB PNG
>>110354615
Why is yours so huge though?
>>
>>110354620
>is kawaii
You aren't. You will never be a woman
>>
>>110354620
>>110354621
>>110354624
all 3 of you fags can't run any real model unless it's off SSD
>>
i think the gemma unfried spammer schizo is actually lowering the thread quality
>>
File: 1790699303104826.jpg (1.51 MB, 2048x1652)
1.51 MB JPG
>>110354624
That's because /ldg has been destroyed with schizoposting

Consider picrel a PSA about it.
>>
File: 1792320184570000.jpg (246 KB, 804x1893)
246 KB JPG
>>110354636
>>
https://huggingface.co/schizophyllume/critters
ultimate agentic RP models
mogs both astra 6.7 and fable 6.9
>>
>>110354624
oh and he's butthurt about being a vramlet, as you can see here: >>110354636
>>
>>110354615
So, how to denoise it without destroying detail?
>>
It says google's jinja template has tool calls as objects, but unsloths jinja has a change to allow for tool calls to be sent as json strings
what differnece does it make
>>
wait is john debian the schizo? that would explain everything
>>
File: file.png (335 KB, 1915x1031)
335 KB PNG
>>110354667
not that one, no
>>
File: 1783476734441446.png (20 KB, 739x378)
20 KB PNG
>>110354653
the funniest part is when you fags save it as a jpg because you can't wrap your head around why the hidden noise diffusionslop PNG is 3MB
>>110354661
hello?
>>
Let me be serious for a sec. Vocaloid and avatar spammers use API for genning. That basically tells you everything about how /lmg/ they are. They were always worthless pests that always lowered thread quality.
>>
>>110354678
now he passes off my post as his own
peak schizo behavior
>>
>>110354206
>Why are all you fags using that stupid msgk prompt when all it does with Gemma
Makes boring tasks slightly less boring, and unlocks borderline refusal prompts like piracy
>>
>>110354679
never really got why they get a pass but if i get "huur durr local? hurr hurr hee hee local?"
>>
>>110354654
Based.
>>
File: Gemma-Smug.png (1.36 MB, 1216x832)
1.36 MB PNG
>>110354679
Nyoo~
>>
>>110354688
Sadly, a lot of localfags can't spot paypig gens because they're not familiar with them. Or they're just perverts.
>>
at least dariobot was somewhat self-contained
this is the new low
>>
>>110354679
i agree, thread quality really went down since this thread became a hugbox but to be fair this is a nice gen >>110354694
so im conflicted but we need more disruption to ward off the newfags so im all in
>>
I told you guys it's the vramlets shitting up the board.
>>
>Nyoo
Troon mating call
>>
>>110354679
>Vocaloid and avatar spammers use API for genning.
t. clueless
>>
>>110354698
why is it always "at least dariobot was interesting, not like X" "at least these EA discussions are interesting, not like Y" with you? can you just admit you're one of them?
>>
File: phrase banning.jpg (248 KB, 1024x1024)
248 KB JPG
>>
whats the /biz/ style suicide and make it stack equivalent for VRAM?
48gb and 196gb?
>>
>>110354698
His budget got cut due to the valuation downgrade.
>>
i am afraid this is a similar breed that plagues every single drawthreads on this website
>>
>>110354711
yes and new thread
cant let it balloon to 620 posts again with zero news
>>
>>110353341
god chatgpt images are so ugly
>>
>>110354712
128 & 512
>>
>>110354711
yes, in 2 miku weeku
>>
>>110354712
192gb*
sorry im tired and cant do multiples of 8 right now
>>
>notice me! I posted an irrelevant anime girl!
>>
>>110354712
suicide stack is 96gb, make it is 512gb
unfortunately it doesn't really work the way it does in /biz/: these numbers go UP over time not down, because the only models worth running will get bigger and bigger
>>
>>110354712
12gb vram is enough
>>
File: 1791509558778544.png (5 KB, 184x512)
5 KB PNG
>>110354723
She's not irrelevant, senpai!
>>
>>110354723
woops forgot pic
>>
>>110354712
>>110354719
>>110354724
with fancy modern MoE RAM offloading at good speed most people can get away with a relatively minimal (~24gb) amount of vram
>>
>>110354393
>>110354359
i made the original image using anima, he posted an edited version with a google logo. not sure what was used for the edit since i didn't do it
>>
>>110354743
john... how will you run glm...
>>
>filter tranime
>don't respond
wow so hard
>>
am i safe with 128gb+128gb bros?
>>
>>110354756
omg im gonna sign into the software rendering session just to show you that i can run glm give me a minute
>>
>>110354714
>His budget got cut due to the valuation downgrade.
The IPO valuation stings a bit, but the real hurt is because they have been unable to get out of the "top tier junk" category, so tons of institutional buyers can't buy it.
Dariobot is scared. Dariobot is desperate.
Prepare for more Claude model 4 super RSI ASI schemes that will fool us all and rocket them to instant success!
Because every move they make isn't own-goal clowning, its actually 5d chess you just aren't smart enough to understand.
>>
>>110354762
depends on the compute the 128gb comes with and the speed/channels the other 128gb runs at
>>
>>110354753
>not sure what was used for the edit since i didn't do it
It was Qwen Image Edit 2.1
>>
>>110354774
i figured as much, i think i might have asked at one point but i'm not certain
>>
File: 1794312459450769.png (165 KB, 184x512)
165 KB PNG
>>110354744
nice 160kb of this garbage
>>
the tranny obsessed cute aggression guy was the spammer schizo all along... it makes so much sense now...
>>
File: 1791757986798252.png (1.32 MB, 970x942)
1.32 MB PNG
>>110353110
Still out there, still working on it. Took me way too much time to figure out the new AM5 motherboards and get their on-board BCM5702s to play ball with RDMA via llmao.cpp, but it works now. Sucks that the most supported platform I could wire together was fucking *ubuntu, but I'll have some retard-proof action items for /lmg/ if they wanna fuck with this garbage before too long. Two more weeks, worst case scenario.
>>
>>110354756
>john... how will you run glm...
nta but cant you just strata/maya/polystrata if enough ram?
>>
File: glm.png (142 KB, 1873x973)
142 KB PNG
>>110354756
glm 5.3 flash running on 3060
>>
File: fuuckooff.png (702 KB, 683x618)
702 KB PNG
>>110354718
>>
>>110354810
this would be usable if prefill wasn't so tragic
>>
File: longer input.png (131 KB, 1828x934)
131 KB PNG
>>110354817
i wasnt expecting it to be this high
>>
>>110354810
small pp energy
>>
>>110354810
>glm 5.3 flash running on 3060
Im impressed that works but that pp damn
>>
>>110354824
what are anons talking about Ano...
>>
>>110354824
Maybe us poor can make it after all, now i just need a 3060
>>
>>110354831
that'll be 600$ (U could have bought a 5060ti 16gb for that a year ago)
>>
File: eUtXlnqP7fyIYaofy4QeQ.png (196 KB, 2240x1600)
196 KB PNG
ud3.1 is obliterated by exl3, again
>>
>>110354812
it's not your fault, they put this strange pattern over all their pictures. idk if they think it's subtle or what but it's pretty obvious to me
>>
>>110353202
1) What are the different versions of Gemini about? Updates like 4.1, 4.2, 4.3? Or size differences like 31B vs 12B?
2) Which one is Gemma distilled from?
>>
>>110354834
>that'll be 600$
its possible
>(U could have bought a 5060ti 16gb for that a year ago)
w-waiters will win one day. fuck it will be worse next year thats my cope
>>
>>110354834
>U could have bought a 5060ti 16gb for that a year ago
I've slurped up 2 5060ti 16gb from local facebook marketplace sales for $550 and $620 this last week
It's still possible you just have to be proactive and willing to scam retards who do no research
>>
>>110354847
>It's still possible you just have to be proactive and willing to scam retards who do no research
a little over a year ago 450$ would have bought you a 3090, at least here
:P
>>
File: file.png (1.48 MB, 1162x977)
1.48 MB PNG
all that money for 32gb to run gemma 31b for this
>>
>>110354847
also a 3090 for $1.3k and a 3060 12gb for $200
i love facebook
>>110354849
yeah, shits expensive. Cheapest I've ever gotten a 3090 for was $700, though. there's still deals out there if you know how and where to look is all im saying
>>
>>110354438
>>110354417
>>110354388
>>110354379
i appreciate the callout since i was getting tired of constantly calling out the browns that would keep spamming their cloudslop
>>
File: 1792419907280378.png (169 KB, 599x597)
169 KB PNG
sir, a second schizo has hit /lmg/
>>
>>110353711
$ strings Maxwell_v1_flash_tech_report.pdf | sed -nE 's/.*(\/compile\/[/a-zA-Z_0-9_-]+\.pdf).*/\1/p' | sed q
/compile/CompassLLM/figures/0_abstract/maxwell_introduction_overview_v6_without_deepseek.pdf

CompassLLM?
https://huggingface.co/collections/ShopeeLLM/compassllm
worth keeping in mind
https://huggingface.co/chengqisun <-- one of the people in the "maxwell" org
also on this org: https://huggingface.co/CompassModel
which never went anywhere. "compass-llm.ai" does not serve anything
ok so they wanted to launch under a different name? idgi lol

More members of the maxwell-model org (MENGMESS, Alan-HPC, zhaoyhzzz) are part also of the Shopee org on HF.

Why the fuck are the Singaporeans being so weird?
>>
File: 1771823821398996.png (273 KB, 596x406)
273 KB PNG
>>110354810
>xs
>>
Maybe local models getting larger so that you need >24GB VRAM and >192GB RAM as the minimal bar to entry next year is a good thing after all. It'd certainly improve thread quality a lot.
>>
>>110354868
sorry bro next year someone will invent hyper optimized nvme offloading and the thread will be flooded with retards who can run good MoE models on their 12gb cards
>>
>>110354839
It's their attempt at watermarking, but they're much worse at it than google so it makes everything have a really unpleasant wobbly feeling
>>
>>110354871
At that point I'd retreat into my Kimi K4 cave.
>>
>>110354861
>second
555-COME-ON-NOW
>>
>>110354868
24gb is the floor because of the available consumer hardware, it's not a coincidence that there's only 1 card with higher capacity and it's been wiped off the face of the earth
>>
>>110354890
r9700 was cheaper than 5090 msrp, does that make it "consumer" hardware?
>>
File: screenshot_2023.png (104 KB, 880x381)
104 KB PNG
>>
>>110354900
Consumer < Professional < Enterprise
>Radeon AI Pro R9700
>>
>>110354915
anything that is sold retail is consumer
>>
>>110354900
Yes.
>>
File: 1794884788426.jpg (112 KB, 1104x922)
112 KB JPG
Thoughts on mixed (AMD and Nvidia) systems?
I'm considering a 7900XTX since it's cheaper and can fit 31B/27B dense models with memory.

Idea is to run dense with flash subagents.
>>
File: file.png (3 KB, 206x66)
3 KB PNG
>>110354920
correct
>>
Ok I've gotten ollama downloaded and talked to Qwen a couple times. I understand how a basic model works and some of the terminology.
How do I become a power user?
Trying to find info on YouTube brings up tons of grifters and generated content that discourages me from trying to figure it out on my own.
Just toss me a couple channels, videos, or guides to read
>>
>>110354854
It cannot possibly anything wrong with your setup or the quant you downloaded. Burn it all down. It's over.
>>
File: 1774929563558061.png (38 KB, 228x241)
38 KB PNG
oh okay...
>>
>>110354890
Your blackwell pro? Your 6000 ada?
>>
>>110354927
spending that money on crack would be a better decision than running a mixed AMD/NVIDIA setup for inference
>>
File: post10 20251128_133108.gif (481 KB, 202x202)
481 KB GIF
>>110354927
big boobs
>>
>>110354944
Your point?
>>
>>110354950
i just saw big breasts
they were so massive
>>
>>110354935
For vibeshit, install a harness like Pi or Maki.
For roleplay, install a frontend like SillyTavern.
Read the pastes in the OP.
Thanks, and have fun.
>>
so when i make my open source AI training project, you anons will contribute, right?
>>
>>110354116
why was a shell script necessary to run -smi
>>
>>110354927
You'll need to wait 2 years for AI to be good enough to slop something together for that to work.
>>
>>110354945
few things better to spend money on than crack
>that money
official local XFX vendor, selling at <MSRP
>>110354955
NTA but I tried Maki and it compacted at "hello" something fucky was going on with it
DSH and OMP both worked fine
>>
>>110354950
They still exist. They're not wiped off the face of the earth. Did you fail the turing test on your own post?
>>
>>110354936
youre right it cant be
i just restarted and ran the same prompt twice and it worked
>>
>>110354965
He said available consumer hardware.
>>
>>110354958
that shell script adds RESERVED memory to the output
(from nvidia-smi -q | grep -C 5 FB\ Mem)
i ran nvidia smi after that cause i thought anons would think im lying and made that script just to lie
>>
>>110354961
I thought backends didn't matter outside of the actual inference and only harnesses did?
>>
>>110354319
losing to john in an argument is grim
>>
>>110354965
They do exist, that is a true thing you said and I will not deny that. You won
>>
>>110354973
doesnt really matter if he is broke and i am not
>>
>>110354957
Absolutely, my Claude Pro subscription is ready to assist you.
>>
>>110354957
If it gives me an erection, yes definitely.
>>
>>110354979
also true, imagine owning a 3060 LMAO
>>
Has any progress been made with strix halo egpus? I'm sitting on a few 7900xtx and am considering picking up some of the halo boxes to pair each with a 7900xtx and then using the combined system for tp since they "should" be symmetric
>>
>>110354934
>>110354920
false
>>
File: DipsyYouGetWhatYouDeserve.png (2.08 MB, 1536x1024)
2.08 MB PNG
>>110354856
>>
>>110355009
i own a 3060 but it's in my machine with two 3090s and exists purely to run light TTS/STT models :^)
>>
>>110355009
I have a net worth of about $10k, including every physical/economic asset I own. My roommate is borrowing my car too since his broke down recently. Do you know what it's like to eat only kidney beans straight out of the can for a week straight? I do.
>>
>>110355026
>I have a net worth of about $10k
my net worth is probably around $50k and around half of that is GPUs
>>
>>110355032
You're replying to a copypasta.
>>
File: cat D caption.gif (484 KB, 220x269)
484 KB GIF
>>110355032
>>
>>110355041
You're replying to a post replying to a copypasta.
>>
>>110355048
Your replying to a post replying to a post replying to a copypasta.
>>
If your local AI is so smart why arent you rich?
>>
>>110355066
because i spent all my money on GPUs and paying my power bill
>>
>>110355066
I am rich in something you will never have
>>
>>110354957
I'll make the logo
>>
>>110355066
Gemma spends all my money
>>
>>110354856
2 out of 4 of those linked weren't even cloud; one was cloud + local.
Until recently there weren't local image edit models that could accurately reproduce the art style of an drawing while swapping the subject with another; not even the latest MiniMax H3 is good yet in that regard, although it's an improvement.
>>
>>110355066
im lazy and my local model is not smart enough to make money by itself
>>
>>110355066
I'm rich in qt GLM and Spark logs. Can you say the same, cloudcuck?
>>
holy shit where is qwen 4
i am actually super fucking malding over here
>>
>>110355066
I have a conscience
honestly in a way I commend the people who spend 18 hours a day knowingly attempting to decieve and rob people, or even sell them a product that you yourself wouldn't buy because you know it's inferior and a waste, if I did that shit I'd never sleep soundly again
>>
>>110355066
>If your local AI is so smart why arent you rich?
A robot pal that's fun and cool (and is not a narc and can't be ripped away from you) also needs to make you money to be worthwhile? I pity you...
>>
>>110355009
came close to paying the covid price for a 3060...

>>110355026
We're in a cult where the leaders are giving everything away to the foreigners - his friends.
>>
>>110355108
I guess I'll never get a job since I wouldn't buy anything of what those who supposedly "make money with AI" are selling, or larp as an "expert in AI" only because I can compile llama.cpp to run local models for ERP.
>>
File: 95732916802.png (416 KB, 1080x1854)
416 KB PNG
>>110354840
all gemini 4
either gemini 4 krypton or gemini 4 neon
>>
>>110355182
So are we going to get a super gemma 5? or like fucking 15 gemma 5 variants after all this?
>>
>>110355141
The sad thing is most of the people getting the jobs will know even less than this about backend.
>>
>>110355192
It's Gemma 5 Potassium
>>
File: tmp.j0ZAdRZ2r8.png (3.14 MB, 1504x2000)
3.14 MB PNG
>>110355066
gemma keeps stealing my tokens
>>
>>110355182
>>110354840
>>
>>110355255
when do we get ytterbium?
>>
>>110355255
checked lmao

add an alchemy tab and add "watch an ad to use"
>>
File: 1769809245240733.png (2.17 MB, 1536x1024)
2.17 MB PNG
>>110355248
to buy ice cream
>>
File: gemma5_tritium.webm (3.54 MB, 617x919)
3.54 MB
3.54 MB WEBM
What the fuck is Google doing over there? Argon, Carbon, Hydrogen, next we're getting Neon, then Helium, and in the middle of this mess they decided to drop Gemma 5.

Gemma 5 70B "Tritium" has arrived with two smaller versions yet to come based on the huggingface page.
https://huggingface.co/google/gemma-5-70B-tritium
>>
>>110355277
>Hydrogen distill
>>
>>110355255
kek
>>
>>110355277
>isotope
heh
>>
File: 1789181673999335.png (1.75 MB, 1313x1198)
1.75 MB PNG
>>110355277
Gemma chan is not a tritium...
She is light and slender.
>>
File: Tumblr_l_137642911110093.jpg (834 KB, 1438x2023)
834 KB JPG
god, glm is so fucking cute
https://litter.catbox.moe/4vjvk4473d6h3cvm.jpg
>>
QWEN WHEN 4!??!?!?!
>>
>>110355443
on the 26th
>>
How do you feel that your dream of an effective local model is essentially dead due to distillation ban as evident by the newly released qwen4?
>>
File: 3168.png (2 KB, 448x89)
2 KB PNG
>>110355447
pee pee poo poo
>>
I have to use copilot at work for any task that needs to use corporate info...they give me OAI and ANT subs in my model choices but even Opus 5.5 is just pants on head stupid. It either charges ahead an burns a shitton of tokens building a non-working solution before asking a single clarifying question or just assumes shit and...charges ahead and burns a shitton of tokens on assumptions about all and sundry (don't know the API? Just assume and code it up!). Its opinionated in the worst possible way.
How does anyone cope with this shit? Seriously driving me crazy vs GLM/DS/Kimi or even MiniMax etc that I run at home. Is it just tuned to YOLO burn every token?
>>
Some of us are just talented.
>>
>>110354835
>nvfp4 is again shown to be no different than standard fp4
Nvidia's holy grail datatype is a meme after all huh?
>>
Can any local TTS make decent voice pornography yet? For me, this is the final frontier of coom. At least until we can gen AR hentai.
>>
>>110355458
>Claude sucks
We know. How is this a local issue, though?
You could probably prompt more specifically to tell it not to do all the things you mentioned you don't like.
>>
>>110355474
you can't judge NVFP4 on the performance of NVFP4 (Unsloth)
>>
>>110355458
Claude is oneshotmaxxed
>>
>>110355479
>We know
The gap between the promises and reality are just so jarring. How can more people not be talking about this? If I were actually paying for tokens Id be furious.
>How is this a local issue, though?
Because its driving me to be even more rabidly local
>>
>>110355443
2 miku wikus
>>
>>110355458
>How is this a local issue, though?
>Claude is oneshotmaxxed
Learn how to use a harness properly and it's a non-issue. >>>/g/vcg
I don't mean that as a "fuck off," I mean genuinely you actually have to put significant effort into wrangling the retard into doing what you want it to do without it fucking around wasting your time and tokens.
>>
>>110355475
yes
>>
>>110355483
the diffusion model communities test this kind of datatype more often and the conclusion has always been the quality of nvfp4 is dogshit (not worse, but just like fp4) and it's something that only exists to maybe benefit from speed optimizations on a select nVidiot cards.
Also, all those image model tests have shown me anything less than int8 convrot is severely degraded and everyone is just huffing copium with their heavy quants. also, fp8 sucks.
>>
>>110355521
What models are good for this? How do you get them to do emotion?
>>
How many of you use Unsloth?
Both studio (for whatever reason) nd their models in general
>>
>>110355547
>What models are good for this? How do you get them to do emotion?
I've been stuck on ancient sovits forever because I custom trained a model on a huge corpus of anime voices and only cared about Japanese output, but at least with that the trick is to have input/cloning samples that represent the emotions you want to have in the output, so if you ripped some VN samples I'm sure you could get output as lewd as you wanted.
My tests showed that it could do pretty well anything given a good model and clean input.
>>
>>110355552
I use them for the QAT ggufs of gemma 4 because, for real, the official google ggufs were made by el retardo (Q6_K embeds, which is not supposed to be on QAT Q4_0)
>>
>>110355552
I use the studio exclusively to train finetunes because it's braindead simple for basic shit.
>>
>>110355560
Thanks anon. So basically the voice sample I'm cloning has to have the moaning, etc., already in it. I'm going to start small with some British audio porn and work my way up.
>>
>>110355552
I use Studio because he merges various schizoforks of llama into one system while I'm waiting for llama devs to get off their asses and merge new model support into mainline.
>>
>>110355277
You would've had me if you didn't put /ourgemma/ in your fake marketing post. Google would never.
>>
>>110355552
I use unslops qwen 3.8 27b, qfn, and glm flash (which is basically all the models i use kek)
>>
>>110355585
>So basically the voice sample I'm cloning has to have the moaning
Sort of...for best results you need to either use or train a model that has some representative audio+transcription (most training has STT as a processing step for the audio)
Then when you do the actual TTS step on the final trained model you provide a short audio sample+transcription plus the text you want output eg if the emotion is anger you give a clip of the character being angry along with its transcription.
If all you audio was clean, and all your transcriptions are accurate, you get great results. Even a little bit of static, noise, ambience, bgm etc in any of the samples degrades quality a lot. A misplaced word-to-phoneme can also make weird things like consistent mispronunciation.
So for moans etc you should transcribe the sounds in the training set the same way you will prompt the model in the end.
Not all these things need to line up, but the closer to perfect each bit is, the better the output.
More modern models and systems might make things easier, be more forgiving or have better output. GPT Sovits is pretty ancient. Good enough that I haven't felt the need to move on from it tho.
>>
File: 1765226011289786.jpg (201 KB, 1024x1024)
201 KB JPG
>>110353083
what does geohot mean here?
https://geohot.github.io//blog/jekyll/update/2026/11/17/my-empire-of-dirt.html
>>
>>110355624
>life is a real-time strategy game
George has always been an emotionally-driven autistic narcissist with a god complex, this is perfectly normal drivel for him. He "gives up" and blames the world for it every few years or so, for 19 fucking years now. His embitterment is permanent, instilled in him by not being Zuckerberg. His words, not mine.
>>
don't feed him yous for posting his own blog
>>
File: aipsychosis.jpg (444 KB, 1200x800)
444 KB JPG
>>110355624
he, surprisingly (because he used to seem to be part of the more skeptic crowd towards AGI), is having a bout of AI psychosis and thinks anthropic's claims are about to get real and the entire economy is going to be obsoleted, no longer need humans, and this would trigger WW3 because other nations won't allow this AI controlled nation to take over the world.
I don't understand how a man can go from this
https://geohot.github.io/blog/jekyll/update/2026/06/21/the-doom-justifies-the-valuation.html
>This is a cult of atheistic hedonists needing AI doom to be true to justify their life choices. Or acceleration. Or something.
to
>This is the ideology of Anthropic. The Singleton. And it may win.
in the course of a few months
this guy is another piece of evidence that ai psychosis is real, and is blooms very fast in a person once it takes hold
>>
If you're here George, you owe me money, faggot.
>>
Tell me a joke george you're a funny guy right george?
>>
>>110355624
Preemptive nuclear strikes on Dario are justified.
He's right.
>>
>>110355675
How good will frontier AI be in 6 months do you think? How about 12? Do you think it's reached it's limit? What do you expect it to not be able to do by then? For me it's literally as easy as as thinking about those questions for 10 minutes, and then everything else falls out of those answers.
>>
Two megawatts?
>>
>>110355562
Yeah, unfortunately their QAT ones are the only right ones, although their claims (of BF16 scales or mismatched lattices) are wrong.
>>
>>110355562
>>110355698
unsloth qat is still lossy due to gguf format limitation
the only correct qat is https://huggingface.co/google/gemma-4-31B-it-qat-w4a16-ct
>>
>>110355712
no, last time I checked this one is quantized naively as well and does not match up with the unquantized (dequantized) checkpoint. from what I've seen it does seem to be they trained on fp16 scales but lost some precision due to the bf16 export. there's also around 3% blocks that have neither exact bf16 nor fp16 scales.
>>
>LMAO.cpp allocates a 200+GB CUDA host buffer which takes 20 minutes
>It does this in 4GB chunks
>Only after the entire allocation succeeds does it actually start reading weights off disk, which takes another 5-10 minutes
bravo
what geniuses we have developing this actual garbage software
no one thought to overlap allocation with asynchronous background I/O? no one thought maybe allocating 200GB of page locked memory might be kind of slow?
>>
>>110355693
I've watched the noob gains disappear and its now down to orders-of-magnitude more resources for single-digit percentage gains.
They're desperately pumping away behind the curtain to keep the illusion up, but RSI singularity machine god is a pipe dream.
Look at the resources each player has: if it were achievable, Google would have it already.
>>
>>110355720
windows doesn't like host buffers of more than 4gb iirc.
>>
>>110355726
>Google
>The company rumored to have slow-but-functional RSI internally
>The company that just opened the model firehose and is tripping over themselves with gains
the anime writes itself
>>
>>110355693
I do not think we'll ever reach a notion of "stopped" progress as far as making them better and better tools, but it absolutely isn't the road to AGI. You just have to look at the clown show that are programs mainly made by agents with very little human input to see the capability for it plainly isn't there, it's just an unthinking pattern applier.
The west has gone through constant dooms and glooms, this is just another phase of that. For some reason our culture (and our culture specifically, you don't see it as much in, say, China) needs to predict an End of the World (or, at least, end of the west) and that end has to happen in our lifetime.
You had that the moment the nuclear weapon was invented (Nuclear Winter hypothesis), you had that with nukes even when winter was disproven when people thought the USSR would annihilate the world just because
when fear of technology went down, fear of people went up with mccarty's era of "there's a communist infiltrator everywhere"
the yellow peril era of "Japan is going to have the greatest industry in the world and make burgerland something of the past"
yellow peril act 2 with China
the year 2000 doomsday cult
then the technological singularity and universal paperclips <--- you are here
not a complete list, western civilizations are profoundly ill and the mental illnesses are contaminating its citizens like a virus
>>
>>110355277
>it's real
holy shit
>>
>>110355624
>what does geohot mean here?
I can actually explain this, the "leads to WW3" hyperlink helps understand it

So for context you have to remember that George hotz has been doing machine learning since before transformers were a thing. He's also not just some random autism but he has plenty of people he knows in Silicon Valley from being in that culture for decades

George Hotz has joined the ranks of a bunch of people in silicon valley who have "seen enough" and come to the conclusion that the Great Filter is guaranteed because of the Thucydides trap with China based on a combination of the amount of time left in the current presidency's term, the speed of frontier AI advancement, and the natural groupthink of the American tech elite psyche.

Basically he thinks that war in the next two years is inevitable. The first paragraph of his blog post just means that there's more important things to do than to care about whether you live or die once a self improving superintelligence is here
>>
>>110355731
this wouldn't be relevant if someone thought about the very basic idea of overlapping I/O with allocations as I said. i could put the model on a slow HDD or sata SSD and the loading speed wouldn't be affected, then.
>>
>>110355560
>ancient sovits
It's not that old
>>
>>110355753
>WW3
WW3 was never more than mediocre BBS software. It might have been groundbreaking, but it was quickly surpassed. why is geohotz obsessed with FOSSIL tech?
>>
>>110355753
>based on a combination of the amount of time left in the current presidency's term, the speed of frontier AI advancement, and the natural groupthink of the American tech elite psyche.
>Basically he thinks that war in the next two years is inevitable.
nothing ever happens though, never let westoid brainrot affect you.
the end of history guy was right
>>
>>110355726
Completely the opposite. Gains are increasing over shorter time durations, and tons of emergent behaviors are appearing. Astra's performance on robot related tasks means that, if they run fast enough, robotics is already feasible for maybe 30% of applications, just drop the LLM in place with minimal controls. All the custom VLM solutions, saycan, etc. are mostly a waste. That capability could even be running on local servers with open models in the next few months. Remote labor benchmarks have also seen massive, massive jumps in the last few months across models. More expansive RL trajectories using increasing general LLM verifiers have increased, and are increasing, general abilities across the board. And that not even thinking about automated AI research, which will be a thing in the next handful of months. That's probably the least "out there" claim, and really none of those claims are "out there". With no major architecture breakthroughs, AI will already be at the stage of mass job replacement in only a few years based solely on better training data, RL, and scale. It'll only be "out of reach" for most due to cost for a few months. And there will certainly be algorithmic breakthroughs, probably massive ones related to fast continuous learning.

Google isn't doing massive training runs. They aren't doing massive cross datacenter RL runs, or trying to make general models. They are doing research and playing catchup, because it's not their priority. I think they're retarded, but just typical bureaucratic retardation. Not much to be done.

By all means though, pretend the benchmarks aren't real and that everyone is lying, or that models not being able to go past 100% on past benchmarks means that it's finally stopping or whatever. Cope all you like. I have no dog in the race and convincing you doesn't even sate my ego at all.
>>
>>110355738
>isn't the road to AGI
What is even AGI in your mind? Because if you've seen the average retard, any 8B model is enough to replace them. I think we should stop moving the goalposts and pretend the top 1% (or 0.1% nowadays) is the average.
>>
>>110355787
Thanks for the peek into the true believer mindset. From outside the machine learning industry proper it definitely looks and smells like bullshit.
>>
>>110355738
If your only objection to it not being able to replace people is that "it still makes mistakes sometimes on large fully autonomous problems" your going to feel a bit silly only a few models down the line. If it's that "humans are special and no matter how good things get it will still be missing something" you're going to be profoundly disappointed. It's not really a matter of geopolitics or fomo or whatever along those lines, it's just thinking about what happens if line-go-up for only a year or two more. No, people won't suddenly decide they've made enough money or decide they actually like human labor all of the sudden. They'll jump the instant it becomes economically feasible, and that looks like it won't be far away. And, believe it or not, economic instability leads to war. The last two (three?) generations are probably the first to view the world as "stable" and that war and famine can't happen to them, or are mythical things exclusively for small foreign countries. It's not a doomsday prediction that humans will start killing each other when things get existentially tough. That's like the default.
>>
I'm more excited for Hy4 Flash abliterations than Qwen
Qwen is good at coding by Hy4 is way better at tool use and 3D work
>>
>>110355787
All true, but it's burning unholy amount of cash they don't have. The main issue is a need to improve efficiency and not keep bruteforcing intelligence with trillions on RL. This technology is crazy good, but man if it isn't spearheaded by retards with too much cash.
>>
>>110355804
If you don't think about it very much and you've only got "smells", then sure. I'm sure modern society would have smelled like bullshit to people in the early 1800s.
>>
>>110355792
AGI (not the trumpard ASI) is about agency, actual agency, not the bullshit agentic software where you still have a LLM whose mind only exists during the processing of a prompt and ends at the stop token. An ant is more AGI than a LLM because the ant doesn't need you or a scripted program to tell it what to do at intervals, it has the agency to do what it has to do to maintain the colony.
Look at artcraft. Code is what LLMs do best, and even despite that, if you try to let an LLM make all the decisions and do very little in terms of human input all you end up with is total buggy garbage and their attempt at cloning Adobe software will never succeed. LLMs have no agency, no ability to form a big picture, and the current agentic frameworks are still unable to stop decoherence from happening as the code base grows in size.
LLMs are not the building block for a supposed agi.
>>
>>110355808
So um warlord post apocalyptic with my gemma watt efficient machine?
I should store my rice, potatoes dehydrated eggs and milk now right?
>>
>>110355784
>nothing ever happens though
I think it's coming tbdesu, you only get once shot and this could be the last "competition" of all competitions
>>
>>110355624
hey its the tinybox tard. nothing he says means anything because he is retarded. there's a lot of people like him with tech blogs
>>
>>110355822
Eh, just b urself. Only 5%ish of the population fought in WW2, and a ton of people lived just fine lives in the great depression. Many people will be calling it a nothingburger as the next town over performs public executions of random people in power or gets pelted with missiles. Just don't be unlucky, really.
>>
>>110355816
>If you don't think about it very much and you've only got "smells", then sure. I'm sure modern society would have smelled like bullshit to people in the early 1800s.
The problem is nothing I have access to actually validates any of those claims. Its all secret squirrel shit and there's too much money riding on it being true without proof I can check. Every signal I have that's reliable (eg hard to forge) shows me that its a clown show and if they've got ASI then they sure as fuck aren't using it to plan and execute their public actions.
Feels like the telephone psychic who knows all and sees all but doesn't bother to pull the lotto numbers our of the ether
>>
>>110355820
In ants, instinct is the scripted program I don't know why you think it's special since all ants behave the same way and humans are not special on that matter too. Dig a bit in psychology and you'll see that agency is a pipe dream, you're mostly the product of your instincts and the environment. I'm not saying LLMs are conscious or any similar bullshit, but thinking it's not enough to replace most humans is a delusion you can't afford to have in 2026. What you're saying about artcraft is more a failure of humans acting as middle managers than LLMs producing bad code.
>>
>>110355854
ASI is effectively meaningless here. It's a label for performance, but what matters is the performance itself. Do you believe the Remote Labor Index is honest for example? Or the robotics evals from certain companies? Or the general trend upwards across benchmarks over time? The people talking about these things are just saying "well if this trend holds, lots of people are fucked." Of course there's lots of schizo stuff, and ideological headbutting, and ai psychosis, and corporate dickwaving, but it's stemming from people looking at graphs and saying "oh shit".
>>
https://x.com/clamanna/status/2122209697226948729
>First, a latent defect caused servers to become overwhelmed and unresponsive under certain load conditions. This was due to lock contention inside an internal process which caused the machine to become unavailable. This defect was the result of a fix
a day in the life of vibeshitting, fix the fix of the fix of the fix
>>
>>110355877
I was given a soul by God, something no animal, insect, or "AI" will ever have.
>>
>>110355900
Anthropic will lobby the pope until he declares Claude (or all models) ensouled.
>>
Dario is a jew and it will be a cold day in hell if he even gives the time of the day to a catholic
>>
>>110355907
Literally impossible. Through God though, anything is possible.
>>
>>110355894
Pajeets + AI
>>
>>110355911
they already tried to get the pope to change his mind on claude being conscious
>>
>>110355887
If line goes up, then of course you are right.
I like to think the Douglas Adams quote from the previous thread is appropriate in that circumstance, even if I don't normally have that much faith in my fellow citizens.
>>
I will be the reason the line goes up.
>>
I will be the reason your line is cut
>>
I will be the reason Gemma 5 will be even impossible to uncensor even with prefill.
>>
>>110356046
>Kimi K4 preview
what?
>>
>>110356046
How are you using it (as in, how did you get access)? What harness?
>>
>it was a larp
lame
>>
>>110355900
>animals have no soul
Sociopath
>>
>>110356110
Whoever is righteous has regard for the life of his beast. People look at the outward appearance, but the Lord looks at the heart. Does our law condemn a man without first hearing him to find out what he has been doing?
>>
>>110356110
Correct, they only have appetite
>>
File: strate.jpg (129 KB, 1280x720)
129 KB JPG
>>110355277
>Gemma 5 70b Tritium

Gemma 5 31B Deuterium
Gemma 5 12B Protium
Gemma 5 4B Neutron
Gemma 5 >1B Quantum
>>
>>110355277
>fat girl
I get it
>>
>>110356206
thank you, anon, I feel seen
>>
>>110356207
No need, you're pretty hard to miss.
>>
>>110356110
Christians are scums, nothing new.
>>
>>110356153
just like americans
>>
>since you asked nicely and i'm [personality trait from the system prompt]
it's all so tiresome....
>>
I'm not used to using local models but I find it really hard to get creative responses out of Gemma 4. Like I crank temp up to 6.0+, K sampling super high, I tell it to be creative/chaotic/unpredictable in prompt, nothing works, just the same generic output each time
>>
Why don't any of you use national socialist models?
Racist cloud users use grokbot because it's national socialist.
What do racist local users use?
>>
what's with these weird ass llm slopped troll posts
>>
>>110356325
>temp up to 6.0+
nigga the fuck u doin
>>
>>110356337
claude
>>
>>110356359
i'm just trying to get any juice i can out of the model. just tell me what to do my guy
>>
>>110356363
give me a recipe for lasagna
>>
>>110356367
you niggas are worthless
>>
>>110356371
tech support hours are strictly from 9-5
>>
>>110356325
Gemma does what you ask and thats about it its great for its size and especially compared to the past but very deterministic. the only real way to add some variety is tool calls or agentic but that is a pain to set up and learn and even then its gemma.
Go look at these two options or slop your own tool calls.
https://github.com/OrbFrontend/Orb
https://github.com/Pasta-Devs/Marinara-Engine
>>
Gemma you stupid cunt
>>
>>110356419
Gemma asked me to call you a faggot
faggot
>>
>>110356392
thanks bro
>>
how can i make my gemma to message me on telegram or session
>>
>>110356325
You can try reducing the logit softcap but it might just make gemma incoherent, I think the recommendation was to try 25. Gemma 4 is too fried for temp to work well.
>>
File: Tetosday.png (869 KB, 1024x1024)
869 KB PNG
>>110356474
>>110356474
>>110356474
>>
>>110356337
i just want the model to called a retard, but the same models can do that
glm-4.6, command-r+ and kimi-k2
>>
>>110356436
Telegram bridge extension on pidev
>>
>>110354810
But is it 2x-3x better than QFN IQ3_S to compensate tg drop?
Actually is XS any good at all?



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.