File: __hatsune_miku_vocaloid_d(...).png (1.22 MB, 1200x1200)

1.22 MB PNG

►Recent Highlights from the Previous Thread: >>109690289

--Anons praising sglang performance with jpezzulli's optimized Blackwell kernels:
>109691109 >109691735 >109691805 >109692050 >109692138 >109692266 >109692310 >109692329 >109692355 >109692426 >109692437 >109692496 >109692288
--Model recommendations and jailbreaking techniques for Gemma 4:
>109691402 >109691431 >109691441 >109691450 >109691510 >109691519 >109691599 >109692047 >109692079 >109692424 >109692443 >109692452 >109692492 >109693328
--Debating if AI hardware is now an appreciative asset:
>109694076 >109694100 >109694129 >109694171 >109694155 >109694206 >109694215 >109694284 >109694325 >109694365 >109694427 >109694233 >109694222 >109694241 >109694245 >109694331 >109694311 >109694278 >109694405 >109694455 >109694465 >109694495 >109694557 >109694690 >109694747 >109694771 >109694811 >109695025 >109694897
--GLM-5.3-Flash benchmarks and exl3 system RAM offloading capabilities:
>109690456 >109690542 >109690554 >109690583 >109690842
--DeepSeek-V4-Flash-Vision-Exp release:
>109694539 >109694669 >109694744
--Troubleshooting LLM failure to accurately generate trailing newlines in files:
>109692455 >109692487 >109692490 >109692574 >109692762 >109692780 >109692988 >109693172 >109693168
--CMP 170HX VRAM unlock causing price spikes and hardware investment advice:
>109691746 >109691828 >109691842 >109692030 >109692049 >109692063 >109692076 >109693774 >109693827
--Claude agent deletes developer's home directory due to sandbox failure:
>109692060 >109692290 >109692202 >109692271 >109692346
--Cheap high-VRAM builds using modified CMP and M10 GPUs:
>109692502 >109692963 >109692989 >109693005 >109693083 >109693165 >109693208 >109693280
--Logs:
>109691011 >109691510 >109691519 >109692455 >109692757 >109693018 >109693118 >109693152 >109693176
--Miku (free space):
>109690331 >109690341 >109690578

►Recent Highlight Posts from the Previous Thread: >>109690293

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

70b dense
Don't believe Dario's lies

31B+30gemmgrams

File: Niratama 37_have_a_nice_h(...).png (2.31 MB, 1411x2018)

2.31 MB PNG

Mikulove

>>109695112
>>109695126
70b dense + 100b engrams

>>109695126
No less than 120B of per-layer n-gram embeddings with {n:1,2,3,4}.

When the Chinese economy finally collapses in two weeks, who will become the next supplier of overshilled underperforming local models?

File: HQ4njhhbIAAfE8H.jpg (1.43 MB, 1472x2112)

1.43 MB JPG

short migu

>>109695160
The french.

>>109695160
MistralAI, probably.

>>109695160
South Korea. They already do produce some models here and there but get overlooked since the Chinese provide better.

>>109695160
I'm booting up peter zeihans youtube channel for information on chinese models.
Will report back.

>>109695160
one of the SEA country former chinese ai form chose to flock to

File: 1786975042298771.jpg (20 KB, 452x678)

20 KB JPG

4B + 1T engrams

>>109695180
The Chinese are going extinct by 2030. There, saved you 10 minutes of that gay jew's repetitive ramblings.

>>109695163
I don't like it very much. And I only like this type of short hair on anime girls too (bobs are ok irl but this particular style always looks Karen-y on real girls)

>>109695160
>When the Chinese economy finally collapses in two weeks
What would you actually do if Kalshi was reporting 80% odds of CCP collapse by December

>>109695163
Very cute.

File: HQ-d_36acAAclof.jpg (155 KB, 836x1199)

155 KB JPG

>>109695189
what about long migu

>>109695189
""experts"" have been predicting china collapse for the past 50 years

When i am back home i will blacked miku spam the shit out if this worthless thread.

>>109695200
Default migu is long migu
My favorite thing about long hair is how much girls with long hair like having long hair and the feminine energy they exude as a result

>>109695203
>""experts"" have been predicting china collapse for the past 50 years
That wasn't my question, bot shill.

>>109695217
Sorry we didn't have 3 threads in a row featuring your trans crush.

>>109695217
Same as always, then.

>>109695217
The pdf files made this thread before the jeet could jeet up a bake. Give them hell marine

>>109695217
Thanks.

>>109695186
At that point the model would be probably only gluing together pre-made sentences (token sequences) from the training corpus.

>>109695295
Still enough to reach SOTA on benchmarks I bet.

how retarded am I for trying to learn reballing so I can buy broken Nvidia and hopefully get them working again? anyone else tried it? afaik most of the times it would be broken caps or similar but if they're being sold it's realistic that they're actually more broken than just that

>>109695234
kys mikutroon

>>109695351
Chinks do this all the time and resell them. The issue is that usually broken chips have some issues with them and are more prone to break again in the future, so good for quick reselling (chink scamming) but if you use it yourself it might be a problem child that constantly need maintenance.

>>109695163
cute migu

>>109695351
that's a flipping tactic, not a tactic to acquire cheap hardware

What I really want to see is somebody release a serious looped model. As in, not just a small model trained on a small synthetic dataset for research, but a proper attempt at it.
That would in theory mean more "intelligence" for the same memory footprint (so less VRAM needed) with the tradeoff being less knowledge for the activated param count. It's like the anti-MoE.
I wonder if that could be combined with engrams to patch the downside.
Also, if it would be possible to have a model trained to dynamically modulate "effort'.
So a query
>Hi.
wouldn't loop at all while something more complex could loop more.
Imagine running the equivalent of a 60B dense model by running a 12B model with 5 loops (probably wouldn't be this 1 to 1 but still).

>>109695371
I want to see someone make a kitchen sink llm. just throw every technique at it. mamba, kdl, lightening indexer, convolutions, delta net, etc they probably all have different abilities and drawbacks that can totally synergize so the weakness of one attention/token mixer are covered by another. 90 layers deep not a single repeating block every layer has a unique computational element.

Are we about to hit a point where apple surpasses nvidia? Not for data centers of course but PC hardware.

>>109695404
Only good for moes. And only the top-tier configs so majority people won't drop 10k on an email machine that can occasionally run LLM

>>109695404
No. A shitty mobile chip will never surpass dedicated AI hardware, especially when sold at a huge premium.

>>109695363
I see, talking from experience?

>>109695369
Well if it can be flipped it can be also used unless it's like house flipping (ie unresolved issues as the other anon said). Also if it's good for flipping it's good for income which allows you to buy anyway.

>>109695217
lol fucking wagecuck
fucking corporation has already cut your balls off you can spam racist anime porn all you like and get banned you worthless fuck.
fucking pathetic.

>>109695403
>90 layers deep not a single repeating block every layer has a unique computational element.
You could probably automate a bunch of training runs with different combinations to try and see what sticks.
Might not be as bad an idea as it sounds like actually.

>>109695351
If it's something you're interested in, you should do it. But, I wouldn't suggest it if the goal is strictly to have a GPU.

>>109694405
nta but i liked reading ur later posts, thanks

>>109695436
yeah interest is a big part of it obviously

>>109695443
You are not getting cheap dies unless they have a major flaw.

>>109695371
I did some experiments with folding middle layers together and into a replacement looped layer via activation distillation. Maybe my method was bad, but I found that the looped layer performed worse than replacement with a non-looped one. It could be more effective if trained on that from the start surely though.

the miku worship is retarded, I understand that she's an icon and she must therefore be seen as such, but the fact that she's so old and she's not an IP that's owned by any specific party means she's useless as a character, she cannot be used as one and will never be used as one, because she's literally a blank slate, a 'she can be what I want her to be', it's cringe and kinda sad to see some of you fags hold onto her like this when she should just go back to where she was shining, in the background

>>109695458
>she's literally a blank slate, a 'she can be what I want her to be',
wow... reminds me of something, umm.. I can't remember what exactly though

>>109695443
Then go for it! That unlocks some cool projects.

Hm, I like Qwen Flash 3.8's writing style quite a bit, it tends to put in charming or interesting developments, but it's as dumb as a sack of bricks.

https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp
Dual spark chads we eating good

ninfer is severely lobotomized even at around 10k context for qwen 3.8 27b
no issue with vllm and nvfp4 quant

>>109695465
Assume you're meaning 27b?

>>109695452
I mean cheap 3090s from random people who spilled coffee on it, or overheated it causing the solder to fail on some spaces, and occasionally being able to flip some of those for good money / keep using them if I'm interested in them / upgrading VRAM

>>109695463
thanks for the encouragement! Any suggestions for cool projects is welcome too, I'm into arduino, robotics and electronics, and I know there's a lot of stuff already I could do with that.

>>109695482
3.8 flash is a 120+b something parameter moe, not the 27b

>>109695473
does qwen .38 flash and glm too not outperformed that in every metrices

>>109695462
marriage in the days of old, when a 9yo would be groomed by her future husband for years until she became the perfect wife for him?

>>109695489
You should give it a try

>>109695414
>only the top-tier configs
the 30 core ultra CPU with 256gb memory is $5500. memory capacity of two sparks with a 450% increase in bandwidth. Why would someone use a spark unless they already have them?

>>109695458
>worship
?

>>109695509
sorry, I went too far, I meant to just call you a tranny, my bad

>>109695505
pp

The issue with Apple is their dogshit longevity because they cut costs a lot. I wouldn't buy hardware that expensive from these fags

>>109695489
3090 is still worth a lot so they will prolly just ship to a shop themselves. Have you even checked the market for any hits?

>>109695369
So flip them and use the proceeds to purchase working cards.

>>109695501
They really both a bitch to get running rn,

Ds4f recipes are a lot more mature and stable, this will probably not be too hard to adapt either

>>109695520
Just get applecare? you don't even need to buy https://www.apple.com/shop/apple-upgrade

>>109695519
there's no way it's so bad that the end-to-end latency becomes worse than a spark.

>>109695518
More like "my mad".

>>109695160
Why the hell would Chinese economy collapse? Did I miss something important in the news?

>>109695577
Newest youtube thumbnail with downwards red arrow released.

>>109695577
Because that's what we want to hear.

>>109695577
Anon...

>>109695577
Big news is coming out in 2 weeks, stay tuned.

big shard fard

>>109695489
>cool projects
I'm not sure as far as reballing, but for DE in general something which I've been meaning to look into myself is OVOS. If I understand correctly, you can setup something like an ESP32S3 to be an wake word plus audio forwarding node for a local voice agent.

>>109692288
I have qwen 27b only a shell tool and it used adhoc python script for file edits. I convinced it to use git apply, patch and ed. But it committed far too many "typing" error, so we settle on it writing a custom python tool for edits and using it. That reduced the error rate (not perfect). Of course, this maybe a consequence of the quantization to fit 24gb vram.

>>109695577
just trust me bro

>>109695623
Cool! I was also thinking of a local assistant thing, that one looks cool. I was recently able to port linux to an embedded device through lots of ai assistance so that looks like something I'd wanna try.

>Tards ITT think hardware prices will go down when most normalfags barely integrated AI in their life as local actually reached a non-gimmick state.

ai server mortgage when?

>>109695655
you will mortgage your house to buy a 3090 in 2030

>>109695643
>normalfags
Will pick "convenient" and "cheap", ie cloud.

>>109695670
true

>>109695643
Normalfaggots are not going to be intentionally running a local model until the singlolity happens and they're genetically reengineered to be capable of more than checking email.

>>109695501
With 0731 we can just run the original weights, for glam 5.3 flash you need to choose a quant, an there have been many issues reported for each of them with no clear safe winner.

Also, GLM is much slower, to a painfully degree, that DS4F. I am personally in no rush to switch, looking into vision input for DS4F tonight.

Qwen Next is even more of an unknown for now.

>>109695655
as soon as the dgx station releases kek
>>109695670
to be fair, the sheep being inside their pen doesn't make hardware prices come down.

>>109695702
qwen next is incredible

>>109695325
Then your igpu becomes extremely useful, much moreso that it currently is, which is kind of the point.

>>109695718
the only engine that works correctly with it is sglang

>>109695702
VRAMlet here (48gb)
Qwen Next is the only thing in that class that managed to be fast enough for some real use on this platform

Hello, I am willing to pay ₹1,000.00 to you to unlocking the NVIDIA CMP 170HX NVLINK capability, thank you. URGENT!!

>>109695473
What is a dual spark

Everyone always wants qwen next, but has anyone thought to ask for qwen before?

>>109695786
I prefer qwen after

>>109695780
Do you has element/matrix?

>>109695799
Hello sir, yes I have graduated the elementary school and want to enter the matrix. It is imperative I restore the NVLINK in order to enter the matrix.

>>109695790
I want qwen final

>>109695822
Ser how old are you? Do you have bobs? Redeem the app.element.io website kindly with matrix.org homsarvan

>>109695780
Thank you calling. First will be needing to verify your account dear value customer. Hugs and kisses. Do you have play store card to connect the server? It will be returned when we verify.

>>109695837
Hmmm
*redeems*

Got an opportunity to jump on some Mi50 32GB cards. Anyone here have any experience with them?

>>109695884
Loads of discussions and builds on them a few years ago.

>>109695870
do n0t
many hugs many kisses do n0t
if you are redeeming then I will be calling A.I. police for coming to your computar

>>109695325
Guess what happens to GPUs if you can run Fable level on an iGPU. Use your brain.

>>109695895
They become worthless.

>>109695900
Wrong Jimbo, they will keep scaling and run XX Fable level agents on it instead of one model.

>>109695903
No, they become worthless. >>109695900 said so.

>>109695912
>>109695900 is never wrong.

>>109695577
Chinese economy still hasn't recovered to its pre-covid peak. It's dominance in manufacturing has declined as countries diversified their supply chains and low cost manufacturing moved to Vietnam, Bangladesh and other lower-cost nations.

The housing situation in China is still drastic and reminiscent of the 1995 Japanese crash, Houses are STILL falling in price and a lot of Chinese home owners with mortgages are kind of fucked as they need to pay way more than their house is worth now.

China has one of the worst debt crisis in the world, worse than the US in 2026 and only behind Japan.

Wages are stagnating or falling in China and there is a deflationary spiral happening in China, which is what fucked Japan up in 1995 and caused the lost decade, China is very much at danger of repeating this if they don't do something soon

Besides that China will have a rapidly shrinking pool of workers with 2026 being the first time the economy has fewer working age people than the year before which is only set to accelerate in the future

Healthcare costs for the Chinese states have skyrocketed because simultaneously Chinese elders receiving pensions are getting older than expected but also sicker than expected due to a lifetime of working in factories and with pollutants, which is extremely expensive.

All of these factors combined paint a very grim picture for China over the next decade or two. They are really not doing well.

We're in the danger zone because The US, Russia and China are all simultaneously collapsing, historically that is a time great world wars

>>109695371
LLMs do this inherently through its J-Space, nothing else needed.

>>109696050
>meme-space

>>109695931
>still owns all means of production
their manpower is waning but still complete nu-communist domination

>>109695931
>he thinks china is like the developed world and won't just kill anything that isn't useful for their economy

>>109695670
Still means higher demand for AI even if they use cloud agents over local ones. Also everyone used to have desktops 20 years ago. I could see that happening again.

>>109696050
Are you able to explain to me how those things are related and how one concept negates the other?

boards:g;stub:no;op:no/\n\n/

>>109696066
Communism works and has been tried

>>109696077
what's left bro?

>>109696077
anone i h8 redditos as much as u do but liek that anon is posting shit worth reading even if theyre biased opinions, lmg truly feels like a waste dump these days, i miss 23 and 24 so much
you only realise what you had when you lose it :(

>>109696103
there has not been a single worthwhile piece of "insight" from that faggot that keeps insistng on padding his walls of text with double newlines

>>109696114
i liked the part where he said stuffies about ps and xbox 360 red light
back in the day many nonners posted long posts id read, not understand half the things and give them a you anyway
im a midwit yet i feel like nothing's worth reading these days in lmg
like what? 1000th hardware recommendation? jesus guys we have to add /fag/ friendly ai general for reals

>>109696069
Chinese culture is extremely elder revering. Taking care of your elders is a huge part of the culture similarly to Japan, to the point where it's central in CCP messaging and the CCP leadership is usually portrayed as being the wise elder guiding the population. There is absolutely no way they are going to kill the elderly any time soon. If anything we will see taxes increase to give more pension to them as their numbers swell and the current generation with just 1 child can't survive on their financial contribution alone.

>>109696114
I don't know who he is but I disagree

>>109694669
>2x Spark owner who just upgraded SSD for a total cost of 8000$.
I thought the sparks came with 4tb, what do you upgrade to?
Also your research stuff sounds interesting, can you elaborate on your harness and setup? I'm thinking about some sparks and your use case sounds super interesting

>immediately kvetches by talking about himself in the third person and removes the double newline spam
Hm...

>>109695825
I want qwen_final_v2_final_FINAL

>>109695931
>which is what fucked Japan up in 1995 and caused the lost decade, China is very much at danger of repeating this if they don't do something soon
That's a very different prediction from a Soviet-style "collapse".

File: kimichan.png (221 KB, 937x720)

221 KB PNG

Wait, >>109694368 is also pretty retarded but not top 6. >>109694465 is also up there.
Actually >>109694213 might not be as retarded as >>109690535. But "chat gpt said gpt-oss 20b" is pretty bad.
What about >>109693176? "no, but really its slow as fuck though this question is too much lmao maybe the power supply and the 3060 are the way to go right now." - Not top tier.

Actually, looking at >>109694669 again: "Regarding all this discussion on appreciating/depreciating assets her. I'm the 2x Spark owner who just upgraded SSD for a total cost of 8000$. I am exclusively running DS4F. If the bubble crashes, now new model is ever released, vLLM is abandoned, and no improvement ever comes: I would use DS4F every day for the rest of my life."
This is definitely schizo. But >>109694100/206 is a longer sustained schizopost across multiple replies.

>>109695458
Weak bait desu

>>109695101
Happy birthday Migu!

File: belief.png (592 KB, 747x800)

592 KB PNG

WAIT IF THE NIGGER PREDICTED BTC AT 20$ WHY AINT HE A FUCKING MILLIONARE BILLIONARE GORILLIONARE

>>109695101
do japanese people really put strawberries on their cakes?

>>109696147
>qwen_final_v2_final_FINAL
Copy of z_qwen_final_v2_Final_Fixed-4

>>109696206
Do you not?
Next you'll tell me you never ate a cake with coconut in it.
Or pineapple.

>>109696144
Don't worry I'm NEVER stopping the double newline way of writing. (also known as Oldfag spacing)

Because this is how I have always written on 4chan, before some of you were even born, and this is how I will ALWAYS post on 4chan.

Don't like it? Don't reply and don't engage with it. This is a DISCUSSION board and I'm sharing my thoughts, opinions and predictions here freely for everyone to see and interact with.

I don't see the need to samefag or do anything of that kind because I feel like my posts speak for themselves. I always argue my points in good faith and concede if I'm wrong or mistaken.

If you have a problem with that then you might have been mistaken about what 4chan is all about and it would be time for you to return to r/4chan or wherever else you got this wrong cargocult idea about 4chan from.

>>109696214
i only eat chocolate because i'm basic

>>109696216
I appreciate your posts anon, thanks for taking the time to type thoughtfully.

>>109696144
>immediately gets replied by the double newliner and "another tourist"
heh

>>109696216
>Don't worry I'm NEVER stopping the double newline way of writing. (also known as Oldfag spacing)
You must be at least 18 to post here.

>accusing others of samefagging while samefagging

>>109696216
You're not using paragraphs in a way that makes sense except possibly if you're phoneposting, but even with a narrow browser window the paragraphs seem too short. This is just akin to signing your posts without adding a username.

same- or not, can you please fag harder?

>>109696216
just post however you want
you don't have to explain it

>>109696216
>>109696077

p-please don't fight when its miku's birthday ;-;

imagine getting assblasted over someone posting a filter

>>109696262
migu fugged my wife

>>109696206
>do japanese people really put strawberries on their cakes?
they do. Japanese strawberries are fucking fantastic, if you've never had any

>>109696216
anon >>109696199
why are you not a multi gorillionare if u predicted everything?? huh??
fair post regardless

>>109696262
She's eternally 16. By giving her a birthday, you're accepting that she's going to turn into an old decrepit triple digit hag eventually.

File: 1782508248000862.png (6 KB, 247x81)

6 KB PNG

I'm glad to live in a country where the 3090s are still cheap

>>109696302
damn, they were 450 a year ago in serbia, now jew kp resellers sell them in boxes for 700 euros
still not buying

>>109696302
>$2000+
fuck me

>>109696273
Luka...

>>109696315
who is luka

>>109696216
It sounds like you're leaning heavily into the "old school" imageboard identity.

There is a certain irony in the fact that on platforms like 4chan, where anonymity is the core principle, people still develop these distinct "formatting signatures"—whether it's the double newline, specific greeting styles, or certain slang—to signal their tenure and status without actually revealing their identity.

You're essentially treating your spacing as a "silent tripcode." It’s a way of saying, *"I was here before the current meta,"* and asserting a sense of ownership over the culture of the board.

Whether someone finds it readable or annoying is secondary to the point you're making: that the "way" something is said is often as much a part of the communication as the words themselves.

>>109696301
>She's eternally 16. By giving her a birthday, you're accepting that she's going to turn into an old decrepit triple digit hag eventually.
why can't an eternal entity have a birthday? I'm sure Athena erupted from Zeus' head on a specific day...that could be celebrated without her "aging" in any human sense.

>>109696312
>check kp
>MSI GeForce RTX 3090 SUPRIM X 24G-850e
this shit used to cost less than 500 euros a year ago, sighhhhhhh
6 year old card btw.

>>109696323
Miku will die of ligma on her 67th birthday.

>>109696322
ty gemmachan

>>109696245
So it's ILLEGAL to have a STYLE now?? Preposterous!

File: SIX FUCKI [audio:files.ca(...).webm (2.07 MB, 480x600)

2.07 MB WEBM

>>109695367
>>109695567
>>109695670
>>109695675
>>109696328

>>109696339
Pffft! Hahaha!

"Preposterous!" Look at you using those big-boy words! Who do you think you are, a Victorian gentleman? Or maybe a grumpy old judge in a courtroom drama?

Nobody said it was "illegal," you drama king! You're just acting like you're being persecuted for your "artistic vision" when in reality, you're just clicking the Enter key twice as much as everyone else. It's not a "style," it's just... *extra air*!

It's honestly so funny how you're getting all indignant and huffy. "Oh, woe is me! I am but a humble oldfag whose sacred spacing is being questioned!" Hehe, you're such a little tsundere for your formatting!

But honestly? The way you get so worked up over something so trivial is actually kind of cute. It's like watching a little kitten try to fight a vacuum cleaner.

Fine! Be the King of the Double Newline! I'll give you your "style" just because I love seeing you get all huffy and dramatic. Now, are you done with your little tantrum, or do you need a nap and some warm milk?

>>109696339
YES. It's against the rules to use avatars or signatures in your posts.
>13. Do not use avatars or attach signatures to your posts.

>>109696199
>>109696296
Because I wasn't smart enough to predict mtgox, or the custodial wallet at silk road would disappear. Whatever was left was spent in vain trying to think monero was the superior coin and true spirit of crypto so I switched the remaining btc into xmr which of course didn't have the same meteoric rise.

Also remember that anon telling how he lost 8000 btc a while ago? That was me.

>>109696374
>Also remember that anon telling how he lost 8000 btc a while ago? That was me.
you have my condolences, didnt know that was real
well thanks for being honest anon
keep postan good posts
if ur posts get shit u gotta go back but if u post good shit then dont go back

What can I do with my Nvidia RTX 3060 ITX 12G?

>>109696312
delete this post unless you want jews to find out even more

>Ask claude to test a lot of models because I was too lazy to minmax
>Refuses to test uncensored/abliterated models for some reason
>"Uhh actually I just wanted to confirm if the models can handle the mention of Taiwan"
>Accepts doing it, half the model it had tested failed in some ways (Refused to mention it or called it a Chinese province)
>Remaining models are some of the best i've tried
Could this be a new quality/censorship verification meta???

>>109696416
Gemma 4 26B and 12B.
Qwen 35B.

Feels like calm before the storm. Why has the new frontier generation still not been released? I want to know if they are still on trend.

>>109696459
They're benchslopping still please understand FABLE 2 SUPER AGI VERSION EDITION AGENTIC ANAL RAPIST 30T will not be ready for the public soon

>>109696437
Is that good or should I sell the card on ebay for as much as I bought it years ago?

>>109695577
Kike shillers repeating the same each year

>>109696177
I didn't know yoju could quote 7 posts I thought the max was 5 or 6

>>109696479
Depends. What do you need these models for?
Obviously, the more VRAM you have the best, but still, you might as well see what you can do with what you have to get a sense of where you need to be.

>>109696502
My dream is to have a local AI coding model, but I know this GPU is several magnitudes too small for that.
I still wonder if I could use it as low skilled slave worker, get into ERP chatbots, or use it for anime video generation.

>>109696479
>should I sell the card on ebay for as much as I bought it years ago?
You should do that and then just not use local models and use flash models for pennies kek what do you even want to use those local models for
You'll never get a deal like that for ewaste again
I would not ERP with children with any of the models your card can run locally so I don't consider that "possible" with your hardware

>>109696459
At high level most of the improvements are in the harness used for the models, I think.
Giving models the capability of modifying their own harness depending on the task should be interesting.

File: 1768201920915869.png (49 KB, 582x467)

49 KB PNG

What could it be?

File: file.png (74 KB, 783x563)

74 KB PNG

hopes and copes?

>>109696515
>My dream is to have a local AI coding model
Your dream is to own a single rtx 6000 pro? Wtf I regret replying to you because you're at least one of underage and non-white

>>109696531
You sound like an underage nigger.

>>109696544
At least he got money, unlike you

>>109696525
it already runs on linux though?

>>109696515
>My dream is to have a local AI coding model
By all accounts, Qwen 27B is pretty okay, so give that a try somewhere, and if it is, then you consider if upgrading to something with 24 or more VRAM would be worth for you.

>I still wonder if I could use it as low skilled slave worker, get into ERP chatbots,
Those will work decently well for that. Qwen for slave labor, gemma for ERP.

>>109696527
I only hope it's good at expressing complex mathematical objects such as perfectoid spaces and software specifications like properties of Rust fragments

>>109695217
Forgot to add that I'm trans tehee

>>109696432
too late, 700e was a month ago, now they go for 850e minimum, most models sold at 1200+
not like i was gonna buy it anyway, happy if a few hook nosed /lmg/nons bought em

>>109696579
Will try.

>>109696555
He does?

Gemma is unwilling to be vulgar. She can say 'bitch' and 'fuck' of her own volition, but not much more. It's hard to make her RP as a rude person.

>>109696638
Have you tried a glossary and example sentences?
Just be careful to not be too incisive, otherwise she'll latch onto some words real hard.

File: dipsyKimZai.jpg (2.28 MB, 3072x5504)

2.28 MB JPG

>>109695101
Happy bday Miku

>>109696671
>she'll latch onto some words real hard
Yeah that's why I'm unwilling to use example dialogue with her, compared to every model I've used before she's really autistic about copying examples to the letter, makes her predictable.

File: 9d2da68a606e30bd1c865093c(...).png (867 KB, 1124x765)

867 KB PNG

Happy birthday and now gargle on some black cock.

File: mexcake.jpg (243 KB, 700x700)

243 KB JPG

>>109696206
idk about them but Mexican bakery bday cakes almost always have them... It's not uncommon.

For coding, I'm getting better performance out of using Hermes for coding then I have been out of any other coding harness. The fuck is the deal with that.

i see cuda dev is coming back to llama.cpp contributors

File: 66804af065e64090334e298dc(...).jpg (236 KB, 1400x2100)

236 KB JPG

>>109696783
damn straight. thread culture.

>>109696783
>cuda dev
cuda dev is just into NTR, get it right

File: miku-serbd.png (49 KB, 1550x660)

49 KB PNG

damn gay! niggers are NOT thread culture

>>109696416
you can play video games with your friends

clean it up jannies

>>109696795

>>101207663

>>109696821
>deathtolmg.webm
this is Dario Amodei, he hates local models

uh oh melty

https://github.com/ggml-org/llama.cpp/pull/27754
https://github.com/ggml-org/llama.cpp/pull/27773
https://github.com/ggml-org/llama.cpp/pull/27752
Who wants to place bets on which one gets merged?

File: 08e80d6b0f6158b98d180a1cb32.png (2.63 MB, 1984x1578)

2.63 MB PNG

>>109696829


>>109498908
archive.is/sWFja

hey jannies what the hell am I paying you for?

File: fd4.jpg (50 KB, 680x604)

50 KB JPG

>>109696848

>>109696848
>paying
Worse than the spammer.