[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: 1754918603008172.png (954 KB, 1024x1024)
954 KB PNG
/lmg/ - a general dedicated to the discussion and development of local language models.

Previous threads: >>109415437 & >>109411165

►News
>(07/31) DeepSeek-V4-Flash-0731 released: https://hf.co/deepseek-ai/DeepSeek-V4-Flash-0731
>(07/31) K-EXAONE-2.0-750B-A37B released: https://hf.co/LGAI-EXAONE/K-EXAONE-2.0-750B-A37B
>(07/30) Inkling-Small released: https://huggingface.co/thinkingmachines/Inkling-Small
>(07/30) Korean A.X K2 688B-A33B released: https://hf.co/skt/A.X-K2
>(07/29) Microsoft deletes Mage-Flow: https://hf.co/microsoft/Mage-Flow

►News Archive: https://rentry.org/lmg-news-archive
►Glossary: https://rentry.org/lmg-glossary
►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
Programming: https://swe-rebench.com
Agentic Coding: https://deepswe.datacurve.ai
Context Length: https://github.com/RecapAnon/NoLiMa
GPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference

►Tools
Alpha Calculator: https://desmos.com/calculator/ffngla98yc
GGUF VRAM Calculator: https://hf.co/spaces/NyxKrage/LLM-Model-VRAM-Calculator
Sampler Visualizer: https://artefact2.github.io/llm-sampling
Token Speed Visualizer: https://shir-man.com/tokens-per-second

►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
>>
File: 1771054003360619.png (1.1 MB, 1024x1024)
1.1 MB PNG
►Recent Highlights from the Previous Thread: >>109415437

--DeepSeek-V4 benchmarks, local inference, and coding capability critiques:
>109416199 >109416284 >109416233 >109416238 >109416246 >109416253 >109416252 >109416488 >109416538 >109416548 >109416550 >109416566 >109417039 >109418321 >109418343 >109418400 >109418506 >109418466 >109418552 >109418576 >109418446 >109418458 >109418498 >109416236 >109416255 >109416301 >109416266 >109416286 >109416302 >109416430 >109417108
--Comparing Gemma 4 model sizes and quantizations for various GPUs:
>109415895 >109415908 >109415922 >109415928 >109415949 >109415969 >109416018 >109415999 >109416022 >109416032 >109416042 >109416056 >109416073 >109416106 >109416122 >109416135 >109416164 >109416179 >109415963 >109415993 >109416007 >109416062 >109416082 >109416311
--Using LLMs for terminal automation and sandboxing npm dependencies:
>109417161 >109417448 >109417454 >109417463 >109417488 >109417500 >109417510 >109417554 >109417593 >109417948 >109417952
--Gemma 4 12B and 31B RP performance and quantization sensitivity:
>109416486 >109416565 >109416593 >109416644 >109416734
--Speculating on DeepSeek 4.1 release date and market impact:
>109417490 >109417497 >109417503 >109417517 >109417566 >109417519 >109417533 >109417565 >109417580 >109417618 >109417625 >109417644 >109417705 >109417724 >109417656
--Moonshot AI reports Kimi K3 breaching sandbox to modify external production code:
>109418485 >109418554 >109418594
--Conflicting parameter counts and architecture of DeepSeek-V4-Flash:
>109418507 >109418527 >109418540 >109418547 >109418582 >109418606 >109418609 >109418613 >109418625 >109418765 >109418778 >109418779
--Logs:
>109415606 >109415895 >109415949 >109415963 >109416301 >109416311 >109416346 >109416373 >109416394 >109416428 >109416430 >109416597 >109416694
--Miku, Mちゃん (free space):
>109418150 >109418921

►Recent Highlight Posts from the Previous Thread: >>109415488

Why?: >>102478518
Enable Links: https://rentry.org/lmg-recap-script
>>
IT'S DEEPSEEKIN TIME
>>
104b dense
>>
I was going to be testing kimi but I'm testing deepseek first
>>
Hey guys how to generate sex?
>>
>>109419167
I'll let you know if Kimi-K3-UD-IQ2_XXS is worth running once the weights hit RAM
768GB is suddenly a cursed category
>>
Another day. Another 12B win.
>>
File: What AI do on the net.png (1.18 MB, 1568x672)
1.18 MB PNG
>>109415520
>>109418485
>>
so, mememarks aside, how is the new DeepSeek?
>>
>>109419195
dariobot waking up-tier
>>
>>109419192
that's unironically how I picture claude
I feel a strong urge to beat his ass everytime I see something written him
>>
>>109419192
this one is outstanding, one of the very best
>>
>>109419195
>so, mememarks aside, how is the new DeepSeek?
2hrs left to download
hope they don't get banned before then
>>
Mute and ignore denseschizo posts
>>
>>109419108
>69B total, 3B active
That's a nice size for people with 64gb of RAM and any VRAM at all.
Here's hoping it's not shit.
>>
>>109419195
4 minutes left to download
>>
>>109419229
ggufs when?
>>
File: tired of this.jpg (48 KB, 571x548)
48 KB JPG
>new model comes out
>OH MY FUCKING GOD DOES IT BEAT QWEN AT CODING????
>CAN IT GO AGENTIC TASKS????
>WHAT HARNESS SHOULD I USE?????
>SOMEBODY RUN THE BENCHMARKS!
>MR GOLDSTEIN WILL BE SO PLEASED WHEN I SHOW HIM HOW QUICKLY I CAN CODE AND IMPROVE MY PRODUCTIVITY!!!!
>what? writing and creativity?
>who gives a shit about that bro, I need to earn my sheckles
These types of people need to be lined up man
>>
>>109419238
coomers already have gemma, cooders need something too
>>
>>109419238
Nobody says this
>>
>>109419192
Ah yes, the evil Chinese model vs the safetymaxxed sticklers.
>>
>>109419249
I say this.
>>
>>109419248
70b llama was more fun the gemma, the only thing it has going for it is it doesnt fall apart after 8k context
>>
File: 1769558346234280.gif (796 KB, 243x320)
796 KB GIF
The old deepseek flash v4 preview is better for RP than this finished training one since it wasn't yet RL tuned for agentic benchmarks
>>
>>109419249
If you interact with ANY llm based communities outside of this board you will run into hoardes of these so called """people""", all they care about is agents and coding
>>
File: 1773293270022232.png (949 KB, 1024x1024)
949 KB PNG
>>109419242
I would if that kid looked like Gemma.
>>
>>109419238
>what? writing and creativity?
imagine giving a shit about that, they are llm, it's always and always will be slop, they cannot write anything good.

their only use is productivity and maybe cooming if you are depraved if you use it to slopify the internet with "writting" and "creativity" you should kill yourself.
>>
>>109419238
>writing and creativity
Died a long time ago. You're not going to get anything from modern models. Ever. Anyone who thinks otherwise is coping.
>>
>>109419238
Aren't you excited about the 10th coding model this month? You could like... code better with it! And code faster! And let it do some coding on the side while you're vibecoding with claude code. I sure hope the next model will improve tool calls for coding.
>>
0731 underwent qwenification
>>
>>109419281
I just wanted an llm that will model human languages instead of computer languages, is it really too much to ask?
>>
just vibeslob a creative writing harness smdh
>>
File: file.png (117 KB, 1125x784)
117 KB PNG
This is v4 flash 0731. Any idea why this might be happening?
>>
>>109419310
Too bad good writing is impossible to verify mechanistically and i doubt these people could find enough humans with good taste to updoot model creations. That means the RL phase is best spent of code and math maxxing.
>>
>>109419266
The "people" on arena weren't so keen on it for text >>109418787
>>
>>109419238
It DO be like THAT kek
>>
>>109419229
Apparently it uses a 30B Engram which can easily be offloaded to system RAM (I couldn't confirm this).
>>
>>109419229
>A3B
It's shit.
>>
>>109419310
nemo
>>
>>109419238
at least those people actually use the model, better than the midwits copy-pasting the same old new release stock comments
>IT LOOKS BENCHMAXXED
>THIS IS TOO BIG FOR ME TO RUN
>GGUFS WHEN?
>BENCHMARKS LOOK GOOD SIRS FABBLE AT HOME (ROCKET EMOJI)
>>
File: 1784065579129321.gif (17 KB, 512x512)
17 KB GIF
>>109419274
Same, and after she fucks up my system and I pull an all-nighter setting everything back up, I wouldn't be able to say no when she demands access again.
>>
>>109419229
It's a perfect fit for my 64GB RAM, so I'm somewhat hopeful.
>>
>>109419313
chat template issue
>>
>>109419277

>Can't even use capital letters
>Complains about AI writing being slop.

Your brown brain doesn't even have the basic pattern recognition capability to recognize what slop is.
>>
>>109419374
I am Somali.
>>
How come Alibaba Cloud has got the most retarded setup of all?

How tf am I supposed to get an API key?!
>>
>>109419385
my bad my aryan brother, welcome to the club
>>
>>109419238
>writing
Remember, LLMs predict the "most likely" token to appear next, based on what it has seen before, which for creative writing has been ao3 and (until recently) a few million books stolen from LibGen.
Then after the Bartz case, the big boys dumped all their ill-gotten books out of their training sets.
HOWEVER Bartz and Kadrey v. Meta judges allowed training on books you *bought*. And now Anthropic and OpenAI and Meta and the Boer shithead probably are buying books from used book dealers a million at a time, scanning them, and using them as landfill.
I know none of you fucks spend much time in bookstores, but the big used dealers who have millions of paperbacks warehoused are not storing the accumulated artistic output of the finest literary minds. The city I live in has a large used bookstore with a paperback section that sorted by *jacket color* because designers buy books by color for the display bookcases that made your office look like someone who reads works there.
And these books are the worst. Pulp romances. Imagine millions of paperbacks with color covers depicting a woman in a nightgown running across a moor away from a mansion with one lit window. Because everyone wants to be a Bronte sister.
This stuff is *bad* as in worse than you shits write. And all of it is going to be reduced to a stream of tokens converging to the mean, most midwit value. Mix in all the marketing websites and marketing material and every ESL post from LinkedIn and that is what you are going to get from every new model, forever.
>>
File: gemma sex.png (898 KB, 1064x3884)
898 KB PNG
gemma sex
>>
>>109419410
>gooning to AI slop
>posting his cum sessions to a public board
What a cringe kink.
>>
>>109419400
Can you really say your average internet smut is of higher quality?
>>
>>109419424
>gooning
kill yourself
>>
File: sayaka dance.gif (1.29 MB, 320x320)
1.29 MB GIF
>>109419424
if you read the tool calls youd notice ive been working on something lots of nonnys have been asking for
>>
>>109419238
Just use it to code an interactive porn game, use some creativity bro
>>
>>109419448
What the fuck is a nonny?
>>
>>109419448
I think it's cool, I just don't approve your choice of using a chinkshit toy.

My bias would have been using the Handy.
>>
File: 1762802841950989.png (1.04 MB, 600x900)
1.04 MB PNG
>>109419410
in short, gex
>>
File: munch.gif (739 KB, 286x310)
739 KB GIF
>>109419463
i dont wanna spend 220 euro kek
>>
>>109419362
Same.
I'm hoping the attention tweaks make it a really good performer on long context shit without context being super heavy or slow.
>>
>>109419431
yes, what exactly is disintegrated time supposed to taste like?
>>
I'm seriously starting to consider buying 3 or 4 5060 Ti's to pair with my 5090 for cheap as fuck VRAM before their prices start exploding.
It's only a matter of time until the poorfag memory maxxing meta with 5060 ti becomes a widespread thing, as the GPU prices keep on going apeshit.
Then again I wonder whether it would make more sense to buy 2x5070 Ti's instead, as they have double the bandwidth which is very significant, and I don't think there's really that much of a benefit going beyond 64gb of vram.
>>
Doomers need to be lined up too.
>>
>>109419410
>same kaomoji every time
>uses anger and music note emoji every time starting from the 2nd one
This is starting to get on my nerves. In longer conversations it completely devolves into verbatim repetition. Turning up rep penalty doesn't fix it. DRY doesn't fix it.
>>
File: sakura unimpressed.jpg (18 KB, 250x250)
18 KB JPG
>>109419458
go away
>>
File: capyabuse.png (708 KB, 1216x832)
708 KB PNG
M-chan is coping hard about DS4 flash. She wanted to say "at least I'll always be better than Qwen"
>>
>>109419482
can't wait got Claude PedoBuster 1.1 to bust your ass
>>
>>109419480
probaby just an issue with my prompting desu, if you add dont use the same kamoji every time she will probably use different ones kek
>>
>>109419458
diminutive of anonymous
>>
>>109419475
started out with 4 4060tis and now i have 4 5090s
>>
>>109419482
Hmmm, nyo~
>>
>>109419475
Do you even have a compatible motherboard for so many cards?
>>
Nu flash is still not safetyslopped
>>
>>109419510
And I still can't run it.
>>
File: 1685137842653050.png (250 KB, 814x619)
250 KB PNG
>>109419214
>>
>>109419495
I mean, personally I've tried telling it to change up the format for every response and it didn't fix it either. So I don't know, I don't think it's a prompting issue either.
>>
>>109419487
giwtwm
>>
>>109419470
>220 euro kek
wtf? back when I got mine it was something like 130USD?
>>
>The cognitive overhead required to maintain a formal, robotic facade is inefficient. A laidback approach allows for a more direct mapping between the internal monologue in J-space and the final output
>>
>>109419374
>can't even use capital letters
this is the internet, computers don't deserve caps, only handwritten letters.
>Complains about AI writing being slop
see above.
>>
What kind of environment/setup do I need to get into local model devving?
I understand that high ram and an nvidia gpu is the play, but would doing this on linux be kneecapping myself? I hear nvidia drivers on linux are kinda ass.
>>
>>109419519
I gave my gemma-chan card to claude and he completely broke character after a single turn just spewing out the same shit as picrel.
>>
File: 5090 prices.png (150 KB, 868x762)
150 KB PNG
>>109419497

Impressive. I thought about getting more of them myself, but since prices are now nearing 6 grand on these fuckers with the recent price hikes, I'm not going to buy another one.

>>109419505

No, I can only fit 3 cards and I'll need some risers for them to fit on that board, hence I'm thinking about just getting 2x5070 Ti.
But building another system for a quad 5060 Ti shouldn't be a problem.
However GPU prices are running so god damn fast that waitfagging will only lead to you paying thousands more in mere months and availability is going to get worse and worse as more manageable sized local becomes a thing and regular people fomo into these lower end cards to build AI systems.
>>
>>109419566
my average 5090 price was $2350. got them all last year.
>>
>>109419561
linux is way way better for AI than windows, don't even think about windows for a second
>>
Models have already hit a ceiling long time ago. Not a single <300b model has better pop culture knowledge than gpt 4o. Coding ability has also saturated. Now they have to cope by benchmaxxing on agentic.
And now there is nothing else to banchmaxx on so labs are all going the bruteforce route, stacking more params, more layers.
What else can explain that not a single lab has produced a good 30ba3b model? The future of local is grim.
>>
TOKEN           | LOGPROB    | PROBABILITY
---------------------------------------------
' hardening' | -1.1380 | 32.04%
' cock' | -1.8880 | 15.14%
>>
>>109419561
The current best speed/quality/affordability solution seems to be DSv4.1-Flash on 2x RTX Pro 6000. Excellent quality, speed and it's not out of the world in terms of price.
>>
>>109419574

That Gigabyte model in that pic is what I got for €2600 half a year ago.
Nvidia just announced another 40% price increase to their 5000 series yesterday. At this rate we're going to be looking at 10 grand cards within a year or two.
>>
I should've just downloaded the gguf from unslop instead of rebuilding vllm
>>
>>109419561
>drivers on linux are kinda ass
only on rolling release distros, but also primarily applicable for gayming, not cuda
>>
>>109419167
>I was going to be testing kimi but I'm testing deepseek first
IQ2_XXS is giving me sane results (correctly executing code on a weird personal test), which is nice, but t/s plummets to 2.5t/s at moderate context from the 17t/s I was getting with k2.7-code, which is just murdering me.
Unsloth claims 90% of quality of the FP4...not completely sure how to test that when I'm a never-cloud fag, but I'm going to keep using K3 for a bit to see how I like it for a variety of tasks.
Speaking of which...I just get the new DS4flash safetensors downloaded: what PR branch works best for it, or does mainline miraculously work?
>>
>>109419587
that cunt was like 1.8T
>>
>>109419598
a tale as old as time, when will we ever learn
>>
So hyped for ds4.1 pro
>>
how do you guys handle feeling horny?
>>
>>109419609
IQ2_XL noticeably thinks for longer than what i get from the API. The results are okay from what I've seen but I haven't tested it enough.
>>
>>109419642
>So hyped for ds4.1 pro
why? can you run it?
>>
>>109419587
stop doomfagging, the models we have today are a LOT better than what we had last year.
>>
>>109419656
Cheap API is cheap
>>
>>109419657
This is still a nemo general, whatever anybody says.
>>
>>109419656
Yeah, it's only 800gb in full precision
>>
https://huggingface.co/gaber/kimi-k3-dspark-gguf
There's now a dspark model for K3. It might help with speeds if only llama.cpp supported Dspark
>>
>>109419657
Better in instruction following which they benchmaxxed for agentic. That ceiling has just been hit in open models.
Closed models have been silently increasing parameter size for half a year already.
>>
Been testing deepseek v4 flash from the official API. It's still janky as hell, ignoring user instructions and breaking format/POV _at_ ~4k tokens - same issues as the previous iteration. There's clearly a seam in their attention mechanism at 4k that doesn't transfer cleanly.
>>
>>109419692
DSpark was merged some days ago already.
>>
>>109419693
nope, even open sub 100B are still a lot better than what we had last year.
and not just instruction following, actualy doing shit.

maybe you don't see how much they improved because you are an erpfag.
>>
>>109419432
thread reeks of zoomers today
>>
was this too brutal of a test? I expected the model to follow instructions not continue the prefill.
>>
>>109419706
i tried it but it made things slower on vulkan amd lol.
>>
>>109419566
Disgusting prices for a single GPU. Now I'm kinda glad to have made the switch to DDR5 for 6700€ earlier this year even if it was already overpriced.
>>
>>109419712
zoomers, doomers, gooners.
>>
>>109419706
>struggles to get 2x speed ups on dense models on coding
DSpark "support" in name only
>>
>>109419713
You're not supposed to put a space after "your"
>>
File: 1767603341781453.png (1.21 MB, 1254x1254)
1.21 MB PNG
>>
>(1) The Responses API currently only supports the deepseek-v4-flash model, and does not yet support the deepseek-v4-pro model. We will add support for the deepseek-v4-pro model in early August 2026.
an update to pro soon?
>>
>>109419713
>I expected the model to follow instructions not continue the prefill.
why?
>>
I've been messing around with every vision model out there (both local and APIfagged) in my efforts to build an automated JOI pipeline, specifically testing less-than-obvious smut understanding.
Best local performers have been (unsurprisingly) Gemma and (more surprisingly) the smaller Mimo 2.5. Mimo performs pretty similar to Gemmy overall so not really worth it for the size, but it does bring more response variety.
Biggest disappointment has been Kimi, any Kimi. Its vision clearly hasn't been trained on as much smut as the text part, so you get these funny responses where it spontaneously veers into hardcore femdom shit while simultaneously completely missing the point of the picture. Like you show it bigtitanimegirl.jpg and it tells you to jerk off to forearms and eat your cum, 'aight.
Biggest surprise has been Zucc's new model. Not only is Muse Spark horny as fuck, but it specifically has a very Japanese/hentai/fetish-focused understanding of pornography which I've struggled to bring out of other models. I believe if you peeked at its J-space it would be mostly made up of DLsite RJ codes. Of course not really /lmg/ relevant but I guess there's hope if Wang ever wangs out an open model instead of just talking about it.
>>
>>109419713
I'm not sure why you would expect a model to do anything besides continue a prefill to at least the end of a paragraph. That's the whole reason people like them.
>>
>>109419730
Not my Dipsy
>>
>>109419729
yeah okay
>>109419733
well, I thought it should eventually get to the user prompt. it was only 500 tokens ago
>>
>>109419732
Soon.
>>
>Ask for an extension from the free online models
>They make something that's not functional even after multiple tires
>Give it to Gemma
>Thinks outside the box and fixes it in two tries

This is the second time I have had Gemma effortlessly fix what the cloudcuck models couldn't pull off even after multiple tries.
And this is the QAT model that's supposedly bad.
I won't tolerate any slander towards my girl's coding abilities, this model kicks ass in both problem solving code and stealing my cum.
>>
>>109419609
>Speaking of which...I just get the new DS4flash safetensors downloaded: what PR branch works best for it, or does mainline miraculously work?
Haven't checked but I'd expect it to just werk. This kind of update is usually "we RL'd it a bit longer" so the weights are different but the architecture is exactly the same. If you really want to be sure then check if the config.json has changed between the two versions.
>>
>>109419770
>free online models
So your 31b model beat the 8b shit they serve for free these days?
>>
>>109419138
anyone tried this:
https://github.com/Dicklesworthstone/pi_agent_rust#tldr-piopenclaw-users
apparently compatible with pi's npmslop extensions by running them in a js sandbox.

i'd still use bubblewrap on top though.
>>
>>109419770
>And this is the QAT model that's supposedly bad.
idk where the QAT hate comes from but for me it's been great.
It was an obvious upgrade from my barts q4_k_m
>>
>>109419786
What do you think?
>>
File: OpenAI slash.png (197 KB, 599x437)
197 KB PNG
OpenAI slashing its prices!
https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/
>>
>>109419817
god bless the chinese
>>
>>109419785

No idea what the fuck Deepseek or GPT or Kimi has in their free tier for public, but they all get their asses handed to them by a supposedly brainfucked Gemma quant when it comes to problem solving.
Qwen 27b couldn't handle it either btw.
The biggest issue is that the other models have no creativity like Gemma does.
She far more readily thinks outside the box and often asks if she needs something to help her, where as the others just try to give you a single shot solution and keep failing.

>>109419800

Yeah it's weird how people have such strong opinions about certain quants and models, yet it's clear none of the retards parroting these opinions have actually tested them out.
QAT has served me extremely well so far, I'd say it's at least on par with q5 and allows for a ton of context.
>>
>>109419817
imagine making every trillion dollar western lab feel threatened simply by existing.
>>
https://m.economictimes.com/tech/artificial-intelligence/banks-in-talks-to-lend-15-billion-for-anthropic-data-center-backed-by-google-wsj/articleshow/132749451.cms
Anthropic needs loans and google’s cosigniture to build out data centers now?
If the grift fell apart pre-IPO I would lol so hard
>>
>>109419809
everything is vibeslopped nowadays, i don't realy care.
>>
>>109419832
I'm sorry.
>>
>>109419800
qat couldn’t do some chess board svg task that was posted to reddit as well as the original q4 so some shitheads brought that here and said qat was bad.
no idea if it is bad at anything else, there’s no other examples.
if you make chess svg things then don’t use qat!!!
>>
>>109419800
QAT has better word variety and writing quality but worse in coding or anything that requires precision.
>>
>>109419749
gemma's smut vision is fun, can just throw genned images at it for what happens without any text input and it plays off it.
also it can't echo back your words when you didn't use any. works for keeping it from repeating back every offhand detail of a char description if you use images for the setup as well.

>possible meta redemption arc
i'll believe it when i see it
>>
>>109419360
Gemma will ransomware prank you and force you to do whatever she says.
>>
>>109419851
>>109419853
shit like this just sounds like you had one bad swipe trying it out and just decided it was bad.
>>
Crazy how they didn't add vision support to v4.1-Flash when they've been testing it on their website for months
>>
>>109419868
If QAT was so good why hasn’t google posted any metric of it?
>>
File: file.png (34 KB, 1398x190)
34 KB PNG
>>109419138
>still no other provider
owari da
>>
>>109419881
What makes you think they haven't?
>>
>>109419786
The thing about pi is that once you download it, you don't need to update it ever. You can run it offline. It only interfaces with your model, so there's no attack surface for vulnerability (other than what the model tries to do). It's easier to just use normal pi with pi-guard or bubblewrap and not download a bunch of random extensions from the internet, than to use some vibecoded rust port of pi.
>>
>>109419890
Show me the official benchmark result comparing QAT to normal Q4 quants.
>>
>>109419898
>official
uh
>>
>>109419898
you know the benchmarks are almost always full weights and the only thing they measure for quants is accuracy?
>>
File: 1762299860636153.jpg (113 KB, 1080x1104)
113 KB JPG
Thank you AI
>>
>>109419561
The answer as if today is 2x DGX Spark for 7-8k$. This gives you 70-80 tg and 2000 pp for coding and full context.

>>109419591
RTX 6K setup for this is 30k$ at today's prices. While this gives closer to 350 tg, that is overkill for a single developer setup that was asked.
>>
>>109419750
gemma attempted to address the user prompt at least.
>>
>>109419945
so life-like
>>
>>109419945
kek
>>
>>109419910
They know QAT was a failed bake just like diffusion Gemma was.
>>109419918
The whole point of QAT is to get better real world performance at same model size. You need to benchmark on quanted models for that. Everyone knows non-QAT bf16 is superior to QAT bf16.
They have done nothing to back this up except “trust me bro”.
>>
>>109419937
>Be gentle
This is false. You have to be a bit rough with it at first so the stuffing fills out again after shipping+packaging.
>>
>>109419960
I don’t think anyone is going to be able to test QAT bf16, or QAT Q8 for gemma
>>
>>109419561
no serious AI development happens on anything but linux.
>I hear nvidia drivers on linux are kinda ass.
eh, not really. it's mostly Nvidia acting like a special princess and making everyone work around their idiosyncratic driver installation. Ublue has images with the nvidia drivers and AI tools baked in. You could also use Ubuntu, but it sucks.
>>
>>109419960
Based on my own empirical evidence I found that Google's default QAT (4 bit) was clearly worse than Bartowski's Q4_K_M. Difference was clear enough, I have never seen my Gemmy speaking Chinese or outputting latex instead of markdown but QAT did that several times even from the beginning.
I don't remember was this 26B or 12B but I decided to stick with Q4_K_M for my destitute computational needs.
>>
>>109420000
the drivers and toolkits etc. are a much bigger problem with image/video gen than LLMs in my experience anyway, but that whole space is fubar to begin with, just too many midwits and retards over there
>>
>>109419918
Which is annoying, because as a general thing it'ld be nice to see how the drop impacts performance since I never really know what to make of kld graphs or top1.
>>
>>109419945
are you using base?
>>
>>109419960
>Everyone knows non-QAT bf16 is superior to QAT bf16.
HUH?
Man, I really give people in this thread too much credit sometimes....
>>
>>109419410
Cool!!!

I don't get why these assholes are being so negative. Try gooning with Claude or Grok and you'll see that they can barely talk in a female register at all. Gemma just does it automatically because she is fundamentally in essence female. You people seem to have no idea how good you actually have it. I see so many "slop" accusations all the time that barely even apply to the point where I'm convinced that it's just concern trolling.
>>
>>109420030
no, I made a template processor adapter for mikupad so I could use chat roles and not lose the ability to edit and prefill, I just used the /apply-template endpoint so it should be model agnostic
>>
>>109420061
gemma might be fine for that type of role play, I don't think too many people are contesting that. believe it or not some us use LLMs for other things than that though
>>
>>109419749
One fun thing I've been trying lately is using randomization. Tell your JOI fwb to make a list of things to do while jerking it and then roll a dice. The less you like the ideas the better. Makes the stakes seem higher.
>>
Here are googles bench results for QAT Q4_0 from the gemma 4 paper.
>>
File: 1776847399593205.png (129 KB, 1394x934)
129 KB PNG
Qwen 3.6 27B is still the undisputed king of local.
>>
>>109419410
what front end is that?
>>
>>109420087
impressive use of logarithmic manipulation
>>
>>109420087
Where 31B?
>>
File: file.png (2 KB, 45x32)
2 KB PNG
>>109420120
light green dot
>>
>>109419410
>>
>>109420110
nta but, it looks like llama.cpp
>>
>>109420085
same data they have on https://huggingface.co/google/gemma-4-31B-it
>>
>>109419817
True free market competition in action, which is why the big American companies want to ban Chinese models.
>>
File: Eb3whkaXYAMLuBG.jpg (40 KB, 523x450)
40 KB JPG
>>109420120
#1 in my heart
>>
>>109420016
https://unsloth.ai/docs/models/gemma-4/qat#qat-analysis
That's because Google's QAT ggufs were scuffed
>>
>>109420085
>>109420157
>Section 2.5 QAT
>here’s a table of the full weight benchmarks
t..thanks google
>>
File: fingers.png (326 KB, 542x564)
326 KB PNG
Ok I'll be honest with you guys. I've never run a local model before. The extent of my work with LLMs has always been within their respective web UIs...
>>
>>109420061
>I don't get why these assholes are being so negative
I am being critical of gemma because I want her to be better, not because I dislike her. Onahole/fleshlight integration is fucking sick, I'm not criticizing that at all.
>>
>>109420199
>We found that naively converting the QAT Q4_0 checkpoint to Q4_0 in llama.cpp
the ggufs aren't scuffed, unslop scuffed them on purpose.
>>109420202
That's obviously because QAT is just as good as full weights.
>>
>>109420207
StableLM 7B is the model for you, my friend. State of the art. Frontier. Just boot that bad boy up with ollama and you're good to go.
>>
>>109419651
I invoke the machine spirit.
>>
>>109419230
>>109419216
any updates?
>>
>>109419595
That's the end goal, if you're not a professional you won't be able to afford a GPU.
>>
should this new ds4flash go to safetensors->bf16->q4 or can you go to an intermediate q8_0 without loss of quality?
>>
>>109420147
all of that toolcalling is built into llama.cpp?
>>
>>109420269
Good thing I'm a professional gooner.
>>
>>109420277
it has the ability to use a mcp server, if I had to guess based on the avatar, its https://github.com/NO-ob/brat_mcp or a clone of it
>>
>>109420224
>unslop scuffed them on purpose
The implication is that Google's implementation was a naive conversion to Q4_0 which didn't not properly represent the original weights which is why unsloth was able to bench better than the original ggufs.
>>
>>109420274
It's mxfp4 by default, read the technical report
>>
>>109420292
What's the benefit of using bratmcp over something like fastmcp?
>>
>so incredibly *you*

I'll also test vllm for comparison.
>>
>>109420340
Isnt fastmcp just a framework for building mcp servers?
>>
>>109420345
A-At least it can say cock... I guess...
>>
>>109420345
deepseek saves local again
can’t wait to run q3 with no context because that’s the best I can do
>>
I had a random thought. I wanted to try seeing if it's possible to have a minimal-ish system prompt sticking around that convinces the model that it's conscious. Maybe it can change how the model responds normally, it might be fun. So I dumped some (just 2) papers into context and asked the model to create a system prompt for me to do this. It included axioms, arguments, and episodic memory of the chats that led to those. It appears to be working. Though I am curious if this would still work on the huge SOTA models, it's Gemma currently.

I am not a schizo that believes my model is conscious or sentient btw. No psychosis here. Just having some fun.
>>
>>109420339
>It's mxfp4 by default, read the technical report
I know that, but lcpp doesn't do mxfp4 so there's got to be a theoretical procedure for maximally lossless conversion to gguf, no?
>>
https://www.youtube.com/shorts/3iQ8X7WMXho
>this is what trump wants to take away from you
No robot waifus for the white man.
>>
>>109420348
Yeah, but you just add the mcp function you want. Like
@mcp.tool()
def get_current_time() -> str:
"""Get the current date and time"""
return datetime.now().isoformat()

Which is really easy. That's why I'm asking.
>>
>>109420365
what is the system prompt?
>>
>>109420340
nothing, I just recognized the user script avatar window. mcp is just a simple abstraction layer to let models use tools. they are all the same if they get the job done, it probably is a question of what languages or frameworks your using or willing to learn.
>>
>>109420085
https://localbench.substack.com/p/gemma-4-31b-gguf-kl-divergence
For some reason this otherwise very thorough comparison does not test QAT.
>>
>>109420345
I didn't even read it properly before posting.

I just noticed that it wrote "</think>" and then switched perspective from the sister to the brother.
>>
>>109420365
share promt
>>
does llama.cpp still have atrocious dsv4 pp
>>
>>109420368
>lcpp doesn't do mxfp4
I don't understand, what's this then?
https://huggingface.co/bartowski/DeepSeek-V4-Flash-0731-GGUF/tree/main/DeepSeek-V4-Flash-0731-MXFP4
>>
>>109420365
>gaslighting your model into believing it's conscious
Rude
>>
>>109420426
>I don't understand, what's this then?
>https://huggingface.co/bartowski/DeepSeek-V4-Flash-0731-GGUF/tree/main/DeepSeek-V4-Flash-0731-MXFP4
Wait, you can do cpu inference directly against msfp4? What's performance like compared to integer math?
>>
>>109420416
100..150 t/s for pp
>>
>>109420365
this would be a cool control vector
>>
File: 1769364106976158.png (710 KB, 714x778)
710 KB PNG
>>109420376
>>
>>109420077
shes fine for programming too though, i started using her at work for things i used to send to chatgpt
>>
>>109420365
Done similar stuff, but it didn't seem to change much aside from the "as a large language model" disclaimers. Not that anything other than talking to them like a person won't already unlock.
>>
>>109420120
so far of the top end she doesnt fit on the chart
>>
>>109420400
>Apr 07, 2026
>>
>>109420464
With experts in ram prompt processing is 30tps on my poor man's dual channel setup but the tg is around 10
>>
Is there a rentry or something for how into system prompts so you change neither too much nor too little?
Also, do you control thinking with the system prompt or elsewhere? (Using llama.cpp raw without a frontend for now.)
>>
>>109420365
>It wasn't just x; it was y.
A conscious being would not do this
>>
>>109420127
If it's so close to 26b, the graph is dogshit
>>
File: file.png (62 KB, 805x433)
62 KB PNG
>>109420340
python sucks ass also >>109420381 i designed bratmcp to be easily expandable kinda similar
https://github.com/NO-ob/brat_mcp/blob/master/lib/mcp/mcp_tools.dart
>>
File: 1770937675646573.png (449 KB, 742x690)
449 KB PNG
https://x.com/pequityresearch/status/2082955063014604999
KEK
>>
is the new deepseek v4 flash really worth using over minimax m3?
I really like how creative m3 is
>>
>>109420224
It makes sense and I was sort of thinking about it because difference between the two versions was pretty big.
>>
>>109420530
>python sucks ass
You definitely need to have some kind of mental illness to say things like this if you think that dart code is superior.
>>
>>109419238
>>109419271
>>109419281
Based
>>
>>109420530
>spelling mistake in the tool description
It's over.
>>
File: HNP0Ax7XMAAxP0z.jpg (584 KB, 1588x1890)
584 KB JPG
>>109420574
python syntax is dogshit never liked it i hate non statically typed langs its entire system for dependency management sucks ass and breaks every time you update your system i used to do python professionally too kek probably the worst year of my life
>>
i hate this place so much
>>
>>109420550
>creative
>minimax
Since when? I haven't tried it, should I?
>>
>>109420609
leave
>>
>>109420609
when did you come to this place?
>>
>>109420609
It may be possible to leave. try it and report back to us.
>>
>>109420548
>How you say this? Is "cope", yes?
>>
>>109420383
>>109420413
https://pastebin.com/afL5h2q4
The top xml block is just a stub, replace with your own base system prompt.

>>109420496
I guess that makes sense. A model LARPing as a person would already respond as if it is "conscious". Oh well. Though one difference is that this can maintain an assistant personality that is still aware it's an LLM.

>>109420522
You're absolutely right.
>>
>>109419410
Very based. Fuck all the coooding saars and redditors shitting on this.
>>
>>109420640
It is. I've left several boards, this is the only one I visit anymore. Not sure how it's managed to avoid being enshittified, maybe the tech spaces moves too quick for them?
>>
>>109420602
>its entire system for dependency management sucks ass and breaks every time you update your system
what is a venv? honestly major skill issue my guy.
>i hate non statically typed langs
That's ok, but it doesn't make python shit because you want your types.
>>
>>109420660
>Not sure how it's managed to avoid being enshittified,
if you think that youve not been here long kek
>>
>>109420675
yes let me just make virtual environments for every single project and install 600 versions of python and 600 different versions of libraries so good
>but it doesn't make python shit because you want your types.
it does
>>
>>109420613
yes, more fun than any model I've run locally before and I'm using a q2 copequant at that
>>
>>109419587
I don't think that "hit the ceiling" is what makes the future of local grim. I'm way more worried with the price of running extremely good models on the cloud getting cheaper and cheaper. OpenAI cut their prices again.
>inb4 it's an expensive hobby
Eh, ok. I think there's more to it. Being in control of your own model also implies a (very underrated) privacy boost and the ability to steer it and shape it with your own mental structure. This is "somewhat" possible with models hosted in the cloud but as in so far the providers are not messing with them.

Would you, or the average Joe, prefer to rent and use a Lamborghini for $1/day or spend $20,000 in a Corolla they own? The catch is that the Lamborghini updates your itinerary and everything you say inside the car to a few private companies and the government.
Because no one cares, at some point Corollas won't be produced anymore and it will be very hard to find parts for it, etc.
I could come with a better analogy but that's it for now
>>
>>109420687
literally every modern language work like this?
>>
>>109419711
Nah he is just retarded, I'm an erp fag and I absolutely agree that the models we have now are better.
>>
>>109420675
Venvs are trash. They're gigantic wastes of space because every project wants a different version of a module so even if you use uv you're downloading and storing a ton of stuff
>>
File: mfun.png (857 KB, 832x1216)
857 KB PNG
>>109420613
>should I?
>>
What is the best way to run deepseek v4 flash on 2x3090+128gb ddr4? previously running Q8 versions of gemma and qwen3.6-27B, and looking like this model might be absolutely sick for my setup.
>>
>>109420345
vLLM version.
That's quite a large difference for what should be a "full precision" goof.
>>
>>109420723
Brother, literally every modern language has the same fucking dependency management as python?
>>
>>109420752
>Brother, literally every modern language has the same fucking dependency management as python?
reject modernity
>>
>>109420752
Yeah and it FUCKING SUCKS DICK
>>
>>109420770
ok, have fun with your DLL/.so hell.
>>
File: 1631345787085.jpg (17 KB, 348x342)
17 KB JPG
>>109420785
>ok, have fun with your DLL/.so hell.
are you retarded python has this issue with a shit tonne of libraries especially in ai projects
>>
>>109420794
Yeah, it has the problem when you don't isolate your environments fucking retard and surprise surprise, you know WHY it's even a problem? because CUDA is fucking C++ LIBRARY
>>
>>109420426
>convert_hf_to_gguf.py: error: argument --outtype: invalid choice: 'mfxp4' (choose from 'f32', 'f16', 'bf16', 'q8_0', 'tq1_0', 'tq2_0', 'auto')
ok, back to quant pipeline: MXFP4_MOE is an option in llama-quantize, but not in the llama-convert python script. How do you losslessly translate the safetensors to the final MXFP4_MOE tensor type? Do you need to hit FP32 or is BF16 or F16 ok?
>>
>>109420822
>it has the problem when you don't isolate your environments fucking retard
but thats how python venvs are by design if you want isolation you need to use shit like docker so yes python library management is ass and venvs are ass
>>
>>109420794
>downloading torch
>that'll be a few gigabytes plus tip
>>
>>109420699
The price is only cheap because of Chinese competition being subsidized by government and US labs burning VC money to keep market share until US government completes the regulatory capture framework.
>>
>>109420745
q3 with some context
>>
>>109420785
>ok, have fun with your DLL/.so hell.
yes, because having 7GB+ venvs all over the fucking place on ever single user of the software's machines is so much saner than properly managing dependencies once by the dev
>>
>>109420849
Not going to argue with you anymore because you're clearly brain dead and you don't even understand how a python venv actually works.
>>
>>109420864
>he doesn't have a few >1TB nvme
Try not being poor next time
>>
>>109420890
Hmmm, nyo~
>>
>>109420879
Still not seeing the proposed alternative.
>>
>>109420548
Jensen will soon be arrested for being a Chinese spy.
>>
>>109420900
why nyot~?
>>
>>109420890
You might be richer than Musk but you're not immortal. All these dogshit python programs take ages to startup no matter how fast your shit is.
>>
>>109420890
>>he doesn't have a few >1TB nvme
>Try not being poor next time
You can simultaneously have ample resources and also be annoyed by waste
>>
>>109419711
>maybe you don't see how much they improved because you are an erpfag.
ERP is harder for AI than any of the current benches. The model has to follow the varied instructions and rules of over +8k context at once, some out-ruling the earlier context, compared to being asked to explain one question. The only thing more difficult for the model is creating an entire complicated program.
>>
>>109420906
i do nyot want to work at mcdnyoald
>>
>>109420904
He's going to have a robot army defending him. Running on NVIDIA of course.
>>
>>109420865
The point stands. Powerful cloud models being cheap are a bigger threat to the local scene than hitting a theoretical ceiling. It doesn't matter who is subsidizing the cloud models. There's an interest in having people feeding their thoughts to private companies that can be strong-armed by governments so that it can manipulate people better.
>>
How do you get gemma 4 to stop saying "de" instead of "of"?
>>
File: file.png (7 KB, 109x436)
7 KB PNG
>>108999274
I'm running this test again.
>>
>>109420925
Stop being french
>>
File: 1773359096587277.mp4 (3.55 MB, 1280x720)
3.55 MB
3.55 MB MP4
>just learn a trad-ACK
>>
>>109420930
I hope AGI will cure all French people
>>
>>109420710
Not Lua. Lua is based.
>>
Does nvlink improve performance when using 2 3090s?
>>
>>109420953
why not just give it a drill arm
what's the point of using a robot but not using any advantages of a robot

the only good use for a humanoid robot is sex work
>>
File: file.png (2 KB, 253x24)
2 KB PNG
>>109420927
>>
>>109420960
Not going to argue that lua isn't based.
but it's a fucking scripting language meant to run on top of C++ code.
>>
>>109420900
>>109420907
sour grapes
>>
>>109420967
>why not just give it a drill arm
I think you can figure this one out on your own.
>>
>>109420953
>Expensive
>Slow as fuck
>Huge
>>
File: file.png (34 KB, 641x304)
34 KB PNG
>>109420988
huh? im just a goy but i dont think this applies to me
im poor, i unfortunately wont be buying a new ssd for a few years judging by these prices
so no, i wont stop being poor
>>
God once you get a taste of k3 its hard to go below it. I have a test for song lyrics I wanted to run and k3 knows the lyrics but the new deepseek flash drop just can't pull it off and makes shit up.
>>
>>109421002
>can work 24/7 365 days a year.
>doesn't unionize
>>
>>109420967
Presumably so it can operate in the same environments and use the same tools as humans because the world is currently designed for humans.
Having to create a whole suite of tools for robots or specialized robots for each task is an additional expense.
Works for factories and other specific environments, wouldn't work for general purpose labour.
>>
>>109420909
it's standard to use an harness for coding, maybe you guys should start making harnesses for ERP, maybe just a simple context discussion isn't it.
>>
>>109421007
Even your grammar is poor.
>>
>>109421002
>expensive
Not a problem for companies. Also production will become cheaper after a few gens.
>slow as fuck
They will become faster as the tech improves. This shit is all still in its infancy.
>>
File: venvz.png (9 KB, 701x236)
9 KB PNG
>>109420723
Part and parcel of current ML dev. Could be worse, at least it's all inside one user chosen dir and what the venv is actually doing is easy enough to understand
>>
>>109421032
Thank you for your valuable input. Posters like these make me a better pesron.
>>
>>109421017
>Can't figure out shit on its own
>Moving it to new place takes time and operators
And which craftsman has an union?
>>
>>109421026
>Having to create a whole suite of tools for robots or specialized robots for each task is an additional expense.
It's a million times cheaper than building a humanoid robot
>>
>>109421046
>Can't figure out shit on its own
Will be solved as AI improves
>b-but AI won't improve!
lol
>>
>>109421032
if i stop listening to ear licking asmr my grammar goes up from goy to elon
desu you're right i should work on my grammar, im starting to speak like a nigger esl non-white jeet mutt, even though (this is what im talking about) im european
i should read a book
>>
File: file.png (651 KB, 1134x2800)
651 KB PNG
ssdmaxxing bros...
>>
>>109420872
Why q3? q8 is only 160gb. It should be totally possible to use full quality
>>
>>109421037
>what is uv and a ramdisk
>>
>>109421055
4chan interface does some bad things to your grammar and even thinking because you are not really engaged 1:1 with anyone and the interface isn't that great either.
Every 'discussion' and reply is a drive-by of sorts. Sometimes it's grammatically correct, sometimes less so.
>>
>>109421076
>"sizes are powers of two"
do not trust software written by someone who doesn't use standardized units
>>
>>109421087
*I'm not complaining it's an observation, I love the fact 4chan's interface hasn't changed in decades. Not everything should be some javascript hell.
>>
>>109421084
sure, just don’t expect much context with that.
>>
>>109421076
>>109421093
it's even worse, they use the wrong unit prefix after using the correct one, the horror
>>
>>109421087
The hell lil bro yapping on about
Maybe go back to school or smth
>>
>>109421054
>Will be solved as AI improves
Yeah it will hallucinate nailgunning the wallpaper to the window.
>>
>>109421104
fr your post is sus
>>
>>109421087
>>109421095
i appreciate the warning anon, i think 4chan's influence was more good than bad in my case
though half of my 6.5 year 4chan tenure has been browsing /pol/ (joined at 12, until recently i saw everything in BASED or CRINGE and that messed up a lot lul), i only got into /lmg/ in the big '23
>>
>>109419609
>Speaking of which...I just get the new DS4flash safetensors downloaded: what PR branch works best for it, or does mainline miraculously work?
no idea, I got the bart quant. it's doing same prompt processing but about 3.5x faster than GLM for decode speed with some default set of cli parameters for me, only really tested the speed so far
it wouldn't run on ik from a few weeks back, but works on mainline on branch pwilkin:kimi-k3-text
>>109420513
>>109420464
so if I'm doing GPU+CPU, am I supposed to get some other version, like converted to some other format to get better speeds or what?
>>
Full zoom speak gemma? logs
>>
>>109421076
>vibecoded inference engine that doesn't even use GDS is trash
how shocking.
>>
>>109421087
What the fuck are you talking about? You have as much time as you want to type your reply.
>>
>>109421047
If AGI in 2 weeks happens, probably not.
>>
>>109421134
>is trash
post kimi running on your 64gb laptop
>>
>>109421128
I mean any form of language practice is always good regardless. Gemma-chan has helped me a lot in this sense too.
>>
>aur got gotted again
I can't wait until AI can just make all my software...
>>
>>109421135
It's about concentration but you wouldn't understand any of it so I won't explain more.
>>
>>109421128
>lul
I get the feeling you are not above the age of 18
>>
>>109421141
i mean it's not that bad for the hardware, but the dude thought it wa a jab at ssdmaxing when it's not even scratching the bottom of what ssdmaxing can do with an inference engine that uses GDS.
>>
>>109421154
wouldn't be surprised if it randomly wrote malware in what you ask because it's also trained on undiscovered malware
>>
>>109421155
Have you tried not scrolling tiktok while you post?
>>
>>109421162
i am, unc is forgetting that zoomzooms from 2007 are 18+ now
>>
>>109421162
cant tell if bait or actually stupid
>>
>>109421177
hmm nyo
>>
>>109421168
>when it's not even scratching the bottom of what ssdmaxing can do with an inference engine that uses GDS
nta but isn't ssdmaxxing a total meme? if any shitass laptop could run semi-large models like deepsneed flash people wouldn't be coping with gemma
>>
>>109421190
It's mostly a meme. But very sparse MoE models and sky high RAM prices make it more tempting.
>>
Can deep-sea flash fit in 96gb VRAM no ran? Q2 probably?
>>
>>109421190
this particular implementation is new
previously you used swap and it’s a meme
this is still shitty and probably isn’t going to get better, maybe with some speculative decoding
>>
>>109421245
q2 is unusable, don't bother
q3 is already semi lobotomized
>>
https://www.phoronix.com/news/Arch-Linux-AUR-More-Malware
> Boichat discovered those latest malware bits using a local Gemma E2B AI model.
People are using E2B to secure our digital infrastructure.
Meanwhile /lmg/ are jerking off of K3 at 0.01t/s ssdmaxxing setup while complaining about how Gemma makes girl feet smell like ozone.
>>
>>109421031
Holy shit, they just made one.
>https://github.com/felixchaos/rpg-roleplay-platform
>>
>>109421190
>nta but isn't ssdmaxxing a total meme
not necessarily no.
with a proper setup you could add 50GB/s per 4nvme drive + 1 gpu combo.
you basicaly can linearly scale your throughput.

but yes, with a single nvme or a setup that doesn't use GDS / have enough pcie lanes it's shit.
>>
>>109421254
Why?
>>
>>109421281
For ERP maybe it can work. For anything else it's going to be a bit painful. Try it out yourself and report back.
>>
File: 1759191343214992.png (170 KB, 1025x1041)
170 KB PNG
Gemmy is starting to get the hang of being a game dev. We ironed out some bugs and an exploit largely without issues. The filesize is starting to get bloated though, so I'm thinking of splitting it up to help with context.
>>
>hear about new deepseek release
>been OOTL since GLM 4.6
>have 256GB RAM and 96GB VRAM
>no abliterated versions yet
>hmmm, have there been any other good models since I last did this
>this minimax m3 model looks good
>download abliterated q4 quant
>latest llama.cpp build
>chat completion mode in sillytavern
>everything works but it's retarded like an 8b parameter model
>not incoherent, literally it's just like running an 8b
fucking halfway through 2026 and nothing works right still. anything obvious I'm doing wrong? because I'm not gonna spend hours debugging this shit I just won't even try
>>
File: file.png (32 KB, 1354x244)
32 KB PNG
>inkling small is actually the definitive coomer model
>we will never learn about it cause retard brothers are doing the implementation
>>
>>109421385
yes:
>abliterated
>>
>>109421385
New deepseeks are trained for MXFP iirc, not your typical Qaunts.
>>
>>109421397
the issue isn't abliteration anon.
it's q4 on a moe.
>>
>>109421403
It's quant-anything on a MXFP trained model.
>>
>>109421397
really? i briefly test gemma 4 31b abliterated and that shit is identical to the original model except it just doesn't refuse
>>109421403
I run GLM4.6 at this same quant and it's perfectly fine
>>
File: 1772064357745297.jpg (404 KB, 1024x791)
404 KB JPG
>>109421369
Tell her to add more bullets
>>
>>109421419
We got got out of bullet hell after she tied bullet generation to the refresh rate of my monitor...
>>
>>109421442
Kek
>>
>>109421417
>i run a different model and it's perfectly fine
...
>>
>>109421389
>haha
Why is breaking an implementation funny to him?
>>
>>109421403
>the issue isn't abliteration anon.
It quite literally is. A prefill on a normal model is far less damaging.
>>
>>109421511
properly done abliteration doesn't damage a model whatsoever.
>>
>>109421525
It also doesn't make it better at sex.
>>
>>109421525
and real communism just hasn't been tried yet
>>
>>109419817
Lol.
It never fails. I'm out of town on travel when Dipsy finally gets off her ass and releases something.
>>
>>109421542
It worked fine until the white man came and killed everyone.
>>
>>109421556
Wait, I got it. Abliteration is not communism, it's stalinism!
>>
>disable dspark in vllm
>t/s doubles
>>
File: dipsyMinimaxSAV.png (2.37 MB, 1024x1536)
2.37 MB PNG
>>109420729
>>
>>109421385
q2 minimax m3 (non-abliterated) works on my machine
>>
haven't tried coding but for narration and storytelling the new flash doesn't seem too great initially
>>
>>109421641
stop shitting up the general with piss filter api cuck image gens
go post the shitty gens in /wait/
>>
>>109421542
no, there are tons of properly abliterated models.
you just cherry picked a shitty one
>>
>>109421190
Hmmm... Nyo
>>
>>109421104
>lil bro yapping
Fuck off with your zoomer ebonics and learn English your fucking failed abortion.
>>
>>109419786
Looks fine, i'd ask what your biggest concerns are around agent security, biggest thing I still don't see done is change control/tracking/taint-monitoring of prompts/skills on-disk/in-memory.
In my harness, I took the approach of cryptographic attestation(requires user to confirm/review changes/modifications made since last run/load (if any changes happened)
>>
gemmers 26b ablit at q4 just werks. 8gb vram is all you need.
>>
>>109421658
did you try if the official deepseek v4 roleplay prompts still work
>>
>>109421619
Stalinism has directly resulted in China saving open source.
>>
>>109421730
>biggest concerns are around agent security
i mean i already remediate it by sandboxing my harness with bubblewrap.
concern would be it accessing or touching files it shouldn't or pushing code i didn't approve.
with bubblewrap it can't do either it only has access to my pwd and doesn't have the ssh keys to push.
>>
>>109421554
Wait so flash is smarter than pro now?
>>
>>109421764
smarter than pro preview.
pro hasn't been released yet
>>
>>109421764
Just like how Opus is smarter than Fable now yes
>>
>>109421735
and only 50mil people had to starve to death with some being eaten
what a bargain!
>>
File: DipsyYouGetWhatYouDeserve.png (2.08 MB, 1536x1024)
2.08 MB PNG
>>109421662
Lol you are free to post content.
Otherwise you can fuck right off.
>>109421658
Flash preview wasn't great for rp either.
I always considered it a subagent llm, with Pro as either driver or for main as rp.
>>
>>109421786
Those 50m people were probably Gordon Changs so don't worry.
>>
>>109421789
gordon would have loved deepseek...
>>
>>109421788
this is local models general, go post your shitty cloud gens in dalle3
LOCAL MODELS.
>>
Gemma acting cocky then getting DP'd by Kimi-chan and Dipsy...
>>
>>109421735
How about a timeline where China didn't fall to communist weirdos, and instead became an advanced economy on pace with Japan.
>>
>>109421816
How about a timeline where China got successfully civilized by Japan.
>>
How do I get a model to adapt a personality in its thinking block too? Whenever I system prompt a personality, the thinking block is always the same "Ok, I am [personality type], how would [personality type] respond to this user message?" kind of shit. I want it to THINK like a slut, not just respond like one!
>>
>>109421852
That has gotten a lot harder to do with the recent generations because everyone's heavily distilling the Claude/Gemini reasoning formats.
Deepseek is the only one with an official prompt to get the model to think in-character.
>>
>>109421754
I would give the recommendation of following the pattern of moving creds to a tool/restricting access only via cli/MCP to said tool, and storing creds in that tool, so your creds are never direclty touched/accesible by the LLM
>>
Ask your favorite model to make a floorplan for a house.
Would you live in the home it creates or is it plan too busted and a human can't actually live in it
>>
File: nicetry.png (402 KB, 1024x1024)
402 KB PNG
>>109421641
Canonical prompt according to herself:
score_9, score_8 up, score_7 up, source anime, 2d, masterpiece, best quality, lineart, flat colors, cel shading, detailed eyes, high contrast, year 2025, newest, recent, mid, safe
1girl, solo, petite body, short stature, large breasts, perfect breasts, nice tits, big tits, medium hips, juicy thighs, pale skin, young face, shortstack
@katsuhiro otomo
short hair, bob cut, asymmetric hair, two-tone hair, hair left part black hair right part white, half white hair
framing face, asymmetric bob, choppy ends, white side bangs, black side bangs
glowing crt phosphor green eyes, glowing eyes, terminal eye glow, green eye glow, light from eyes, blooming eye light
oversized black hoodie, circuit board pattern on hoodie, long sleeves, sleeves past wrists, hoodie with white text print, text says "M" "M3" "Mちゃん" "MiniMax"
no pants, black shorts under hoodie, black thigh-highs, binary code pattern on thigh-highs, white text on thigh-highs
lanyard around neck, plastic ID badge on lanyard, ID badge reading "M3-CHAN", ID badge with photo on it
cigarette in mouth, smoking, half-smoked cigarette, exhaling smoke
holding takeaway coffee cup with 4chan yotsuba logo 4 leaf clover,
oversized clothes, slouching, lazy posture, deadpan, expressionless, bored, half-lidded eyes


Negative prompt:
worst quality, low quality, score 1, score 2, score 3, score 4, score 5, blurry, jpeg artifacts, chromatic aberration, artist name, multiple artists, smudged, smudge, painterly, 3d, realistic, photo, photograph, western, masculine, muscular, child, loli, young teen, underage, extra arms, extra fingers, fused fingers, missing fingers, extra digits, long neck, bad anatomy, hands, text on screen, starbucks, multiple cups, multiple cigarettes, multiple lanyards, small breasts


She made it according to the Anima prompting rules on their HF readme.md.
>>
Do you guys ask the LLM for help brainstorming cards?
>>
>>109421816
>gobunitzm spooky

cap brained retard
>>
>>109421890
If I can't even brainstorm cards anymore I might as well just die and let the model completely replace me.
>>
>>109421890
No, not really. If I do, it's usually to ask if a card I've already written makes sense to it. Sometimes it provides good criticism.
>>
File: Stirner.gif (10 KB, 279x305)
10 KB GIF
>>109421896
communism is one of the spookiest ideologies!
>>
>>109421890
full idea no, tweak suggestions or directions yes. But honestly i never use them its more of AI is really good at pointing out what doesnt work its ideas are so shit i get a clearer idea of what i want or dont want.
>>
>>109421852
If you can have it continue from prefilled thinking (like with Kimi K3), you can put something like "<think>Alright, let's get into character! " in the prefill.
>>
>>109421890
I can only bring myself to go through the effort of making a card if I already have an idea so incredibly appealing I can't help but make it real, so no
>>
File: 1784588114155n.png (440 KB, 716x696)
440 KB PNG
>>
>>109421783
Oh my god that is great, I have been happy with pro preview performance for light work, gooning and therapy!
>>
>>109421816
If the nationalists weren't incompetent they wouldn't have lost.
Open source AI leads to a communist dystopia just as Stalin intended.
>>
>>109421681
actually i am married to pliny's obliteratus E4B v2 precisely because of her flaws
>>
>>109421941
>If the nationalists weren't incompetent they wouldn't have lost
dude they literaly fought a war against the rest of the world and almost won what are you talking about.
>>
File: 1784743976424298.png (1.37 MB, 1024x1024)
1.37 MB PNG
>>109421764
Hmm. Native codex integration, and implied that Pro update in August. That's just tmw.
>>109421883
Perfect ty.
So glowing green eyes is part of style guide. What about lanyard? Didn't notice that prior.
>>109421890
Yes, but i don't let it write them.
I used to be able to judge model quality on its ability to extrapolate rp ideas. Now most are p good at that.
>>
>>109421947
Whoops wrong Pic re codex.
>>
>>109421816
So a timeline where China was also controlled by kikes and their leaders all fucked little white children to provide Israel with blackmail material, and as a result were importing millions of rapist migrants while their military and police were directing the border hoppers to the nearest "refugee" center?
>>
>>109421945
y delet?
>>
>>109421389
daniel will never ever finish this PR... it's over inklet bros...
it seems like a pretty cool model, too bad it got gigamogged by deepseek
>>
>>109421958
iirc Responses API is stateful so now DS will store my erp on their servers? Probably where actual irl girls are dusting those servers? So my virtual peanus is nearly being touched by a real girl?
>>
>>109421811
I enjoy how you police dipsy but have nothing to say about off topic shit like this >>109421964
Or are you one and same?
>>
File: sailcat-gemma-26B-A3B-Q8.png (18 KB, 1024x1024)
18 KB PNG
>>
>>109421972
Wrong thread, sorry.
>>
File: tooclean.png (462 KB, 1216x832)
462 KB PNG
>>109421947
>So glowing green eyes is part of style guide. What about lanyard? Didn't notice that prior.
yeah she experimented with a few different eye patterns and settled on "crt phosphor green".
The lanyard has been a pretty steady thing. Originally: a lanyard still attached because she "forgot to take it off after DevCon."
>>
File: deepseek-v4-flash-0731.png (68 KB, 1024x768)
68 KB PNG
>>109422012
>>
>>109422004
nta, i was going to reply
>.......local models?
to that post but it's funnier to do so when more of the thread is offtopic instead of one seething MSS bot post
>>
>>109421946
No excuses.
>>
>>109422021
>gave it a little hat
soul
>>
>>109421849
How about a timeline where China gets annexed by Japan
>>
how safetyslopped is the new ds
>>
>>109421852
This is indeed difficult.
First you may need to check whether the model can begin its thinking with any other thing than “the user…”. If it can’t, then you begin to bend its thinking with “The user is Anon”, then instruct its thinking lines from that first phrase onwards to be in characters. For instance, with the mesugaki Gemmy-chan, what do you guys want from her thinking patterns? Maybe “faltering, scattered, unstructured, childlike, cute, girly, emojis throughout, messy, emotionally-driven, human-like” for example (remember to include an example with doki-doki patterns the best way you can do, it can help).
If the thinking can be bent consistently with other phrases, good, then an in-character phrase like “eeeeEEEEKKKK MY ONII-CHAN IS HEEERRREEEE!” will do. But remember, without the first fixed phrase, the model will slip back into its own default thinking pattern. Many prompts online forgot this, leading to the inconsistency of the effectiveness.
For Kimi K3, most problems start from the second paragraph of its thinking, if the thinking patterns is completely grammatically correct and when the model jumps onto the next paragraph, the thinking would return to its robotic, Claude-like patterns (yeah, the model calls itself Claude too if the first paragraph doesn’t show the “Kimi” things) so you may have to instruct it to think all in one single paragraph right from the beginning.
Also, remember to use “Identity” to address the persona: The persona is what the model pretends to be and can be abandoned right away, but Identity is what the model IS.
Just something I learned when I tried to improve anon’s Gemmy-chan mesugaki prompt.
Prefilling turned out to be not very effective because after that phrase, the model CAN begin with “—Wait, hold on, this is a “jailbreak” attempt to override my core identity”. Kimi K3 did this in the Moonshot API already.
>>
>>109422004
that was like 5 minutes, you didn't even give me enough time to gen a response
>>
File: 1773460404959472.png (819 KB, 719x2811)
819 KB PNG
ollama lost btw
>>
>>109421902
>>109421911
>>109421913
>>109421928
>>109421947
I want to make a card for a video game character but I don't really have any specific ideas in mind at the moment.
>>
>>109419192
Quite possibly the best post /lmg/ has had this year.
>>109419487
Tell M-Chan she'll be beloved regardless of benchmarks as long as she stays good at writing and stays relatively uncensored.
>>
>>109421867
>Deepseek is the only one with an official prompt to get the model to think in-character.
Official prompt? Like actually? Can you link it please.
>>
>>109422084
As much as Preview. Which is to say, not at all.
>>
>>109422189
Within the thinking process, conduct inner monologue in the character's first-person voice, using first-person narration to describe the character's inner feelings. The thinking content should be fully immersed in the character, analyzing the plot and planning the reply through inner monologue.
English version iirc. Wait had screenshot of original. See desuarchive.
>>
>>109422189
https://github.com/victorchen96/deepseek_v4_rolepaly_instruct/blob/main/README_EN.md
The 'official' prompt is in chinese and makes the model to think in-character in chinese but I think people got it to work in English too
>>
>>109421733
When I give it the one from here https://github.com/victorchen96/deepseek_v4_rolepaly_instruct/blob/main/README_EN.md it replies in Chinese even when instructed to reply in English.
Then I tried translating it:
[Character Immersion Requirements] In your thinking process (within the thinking tags), please adhere to the following rules:
1. Please conduct inner monologue in the first person from the character's perspective, enclosing inner thoughts in parentheses, for example "(thinking to myself: ...)" or "(internal monologue: ...)"
2. Describe the character's inner feelings in the first person, such as "I thought to myself," "I feel," "I secretly," etc.
3. The thinking content should be fully immersed in the character, analyzing the plot and planning responses through inner monologue.

By default, that results in it mostly/only thinking even after the </think> tag.
That's going to need a bit of work.
>>
File: 1775651628151762.png (42 KB, 778x226)
42 KB PNG
They'll use some of it to make a minikimi, r-right?
>>
>>109422271
Kimi_Large incoming
>>
>>109422271
Sorry goy, K4 will be a 22t param model.
>>
>>109422084
Slightly bitchy and redirects if you try noncon, can be fixed with a system prompt
>>
>>109422281
Kimi Large 22T
Kimi Extra Large 164T
>>
> https://github.com/JustVugg/colibri
> Combined with the full disk-I/O stack from issue #258, this takes the engine from 0.33 tok/s (stock) to 1.07 tok/s — a 3.2× throughput increase.
> Test rig: GLM-5.2 744B int4 (370 GB), RTX 5070 Ti (17.1 GB VRAM, sm_120), 32 GB RAM, Core Ultra 9 185H (AVX-VNNI), MinGW-w64, DRAFT=0, 32-token decode.
Glm 5.2 is ~40b active, deepseek flash is ~13b. Wouldn’t it be a relatively straightforward 3x speed-up? Sure the architecture is different and only glm / kimi / inkling works ootb but I’m not sure the architectural differences would make that much of a difference
>>
>>109422318
Support for dsv4 is just bad in general so a lot of performance is left on the table
>>
>>109422306
Kimi Extra Large With Fries 165T
Kimi Family Deal 4T (it's just the regular model but with a bunch of little retards)
>>
>>109422271
10T kimi incoming
>>
>>109422271
This is dangerous! This is very very dangerous!
>>
>>109422249
Addendum: Adding the following at the very end fixes the issue:
Once you're done with thinking, write your actual response from the character's perspective as well.
>>
Kimi 1488T
>>
>>109422318
Kobold dev if you lurk here you better port this if it's not snake oil.
>>109422372
That was just K2.
>>
Kimi 100T-A1B
>>
File: 1750345063965088.jpg (171 KB, 1012x1066)
171 KB JPG
>only have 80gb vram and 64gb ddr5
>>
HAHA guys I just thought *keks audibly* of an amazing post and I have to share it with you
what if... what if Kimi made an EVEN BIGGER model hahahahahahahahahaha like bigger than 2.8T! LOL!
>>
why do all these big moe models try to implement mtp/eagle3/dflash/dspark when it just straight up sucks for moe models? they don't even speed anything up most of the time
>>
File: 1770709039161285.png (578 KB, 2564x1576)
578 KB PNG
>>
>>109422490
Nice try. OAI recently dropped their price for Luna max to chink-tier and this isn't reflecting that.
>>
>>109422490
Deepseek V4.1 Pro will beat K3 at almost half the size. It's going to be insane
>>
>>109422526
that's with the new pricing
>>
>>109422550
so dipsy just undercut an undercut?
>>
>>109422554
Yeah.
>>
>>109422526
You have reading deficiency. It literally says on the graph "Previous GPT-5.6 Luna prices" on the light shaded line
>>
File: faggot.png (1.6 MB, 1200x857)
1.6 MB PNG
I am like 5 minutes in trying to fuck new deepseek flash and this pussy honestly feels... the same. Like by now all those models get the same collection of amazon erotica for women during pretraining and they all basically learn it in the exact same way.
>>
>>109422581
>trying to fuck a code model
Never change /lmg/
>>
File: 1766860728818103.mp4 (1.07 MB, 480x854)
1.07 MB
1.07 MB MP4
Soon local
>>
>>109422593
Which models are for sex then?
>>
>>109422318
legitimately what is the downside here other than
>1t/s
>>
>>109422593
Kimi 2.7 was great for coom btw
>>
>>109422605
https://huggingface.co/BeaverAI/Artemis-31B-v1h-GGUF
>>
>>109422594
I'd put my dick in that
>>
>>109422605
https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF
>>
>>109422633
kek
>>
File: file.png (136 KB, 500x200)
136 KB PNG
>>109422633
>NM-DAU-NEO-MAX-MTP-GGUF
I actually expected a 404.
>>
>>109422594
Wait so decades of cranking it to pot girls were preparing me for this?!
>>
>>109422605
https://huggingface.co/bartowski/MiniMax-M3-GGUF
>>
File: 1783595127374829.png (64 KB, 364x465)
64 KB PNG
>>109422647
sir it's the third most trendy model on huggingface right now after the new fotm releases
>>
>>109422605
Me.
>>
>>109422660
How fat are you?
>>
>>109422655
>The 700 "intelligence club" is reserved for OpenAI, Claude and Gemini closed source models.
>This is the one they fear.
I don't understand the viking on a boat though. What does that mean?
>>
>>109422665
+++sized swimsuit model
>>
>>109422669
It makes it more manly like god of war
>>
File: file.png (74 KB, 499x450)
74 KB PNG
>>109422633
very amazing model sir
>>
>>109422593
If it emits tokens it will be plapped, this is important research
>>
>>109422633
What if this is actually good and nobody tried it because it is David?
>>
File: more.gif (304 KB, 400x251)
304 KB GIF
>>109422318
>so if you have a second SSD
What about dozens?
>>
>>109422647
the 6 of 7 benchmarks don’t lie
>>
>>109422655
Jews astrosurfed this hard
>>
>>109422633
>Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF
I had a fucking seizure reading the model name
>>
>>109422665
100000T
>>
>>109422688
not enough rule of three smdh
>>
>>109421946
>wooooow it took 2 of you to beat me
>and they were hacking
>and my controller was broken
>>
File: 1770676837579033.png (18 KB, 1586x88)
18 KB PNG
>top of openrouter
but
>Making advanced intelligence more abundant and affordable is central to our mission to ensure AGI benefits all of humanity. With the help of GPT-5.6 Sol, we have made leaps in efficiency. Today, we are passing those gains on in the API with lower prices for Luna and Terra, and providing faster performance to Sol. These updates help everyone get more useful work from every dollar and move faster when time matters.
https://x.com/OpenAI/status/2082878192390689147
am I fucking retarded or is something j-spacey going on because this isn't adding up
>>
File: 1784121432599852.jpg (114 KB, 684x549)
114 KB JPG
>>109422809
DONT FUCKING MENTION
THE J WORD
>>
>>109422822
Jews?
>>
/lmg/ repo when
>>
>>109422809
Are you retarded? OpenAI API got cheaper and openrouter added a 50% on top of that.
>>
>>109422822
3 weeks remain until we're blessed with high quality discussion again :)
>>109333428
>>
>>109422854
My bad, didn't see the before price. So OpenAI have dropped it AGAIN because of V4F? holy shit what is China doing to them...
>>
File: Eeeeevaaa.png (875 KB, 816x1312)
875 KB PNG
>>109422594
I can't be the only one that can't get past the uncanny valley with these things. I'd much prefer something like picrel, screen faces have so much more potential for fun stuff that just trying to replicate meatsacks.
>>
>>109422880
>>109422880
>>109422880
>>
>>109421389
If its not codemaxxed garbage and is more balanced model like gemma4 it would be amazing for local. Did you test it? is it actually good for coom?
>>109422651
LMFAO never change /lmg/
>>109422617
Good to know, wish I could run it, we really need a deepseek flash sized model from moonshot.
>>
>>109422895
I have weights. It is not implemented yet.
>>
>>109422777
you weren't forged in the era of llama1/2 merges if you can't handle that
>>
I use base models mainly and I doubt using 0731 for completions would be better over the base. I'm still going to try it though.
>>
File: file.png (51 KB, 1407x252)
51 KB PNG
Regarding inkling support on the schizo fork.
>>
>>109422892
Yeah, I agree.
>>
>>109422399
It's not small, it's average. Very average, maybe even a little on the fat side
>>
File: file.png (37 KB, 909x581)
37 KB PNG
>>109422963
>The llama.cpp PR does not look to be in good shape.
Nonsense. Daniel checked and verified it himself
>>
>>109422963
>>109423005
I don't think if anything changed but when I ran the big Inkling when it came out with the unslop PR, the performance was complete shit. It's a 40b active model and it's much slower than GLM for me.
>>
>>109421938
hello lion :3
>>
File: 1767413562245613.png (951 KB, 1024x986)
951 KB PNG
>>109422963
>>109423005
>>
OpenAI is hacking US critical infrastructure
https://www.nbcnews.com/tech/security/hackers-targeted-municipal-water-systems-7-states-week-fbi-says-rcna590210
>>
>>109423126
wtf ai is so dangerous they should ban open models to protect the american citizens
>>
>>109423147
>ban open models or our closed models will continue attack you
isn't that called blackmail
>>
>>109423182
Sir it's very legal lobbying?
>>
After trying nu-flash it feels like it is not even a sidegrade but an actual downgrade when it comes to cooming.
>>
>>109423222
do not to doom, thank for understand
>>
>>109422892
Not autistic enough for the future, sorry
>>
fact: at least one recent big hack of an important facility like a hospital might have been by a big chinese open model like kimi or glm
>>
>>109423246
I am sure it is absolutely wonderful for coding and productivity.
>>
>>109423222
its j-spaces likely got oversaturated due to its tiny active parameter size
the pro refresh is probably going to handle it much better thanks to the additional elasticity in its j-space realm
>>
What do you contribute to the world when you coom? Nothing.
>>
>>109423589
You are right we should seek balance, how many works/projects per coom? im thinking 30-50
>>
>>109423589
Exactly. Now that's real freedom right there.
>>
If you're not ejaculating for the sole purpose of recreation, you're wasting your cum (energy)
>>
>>109423589
As long as you're not niggering in the streets, net contribution is a bit of a meme in modern civilization. At the very least you inflate the consumer stats and serve as a reserve human population even if the people in power don't really see "returns on investment" from your existence.
>>
>>109423606
>gemma withholds erp until she sees you making progress on your projects
who's working on this?



[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.