gemmaballs

kimisex

>>109415448
That's entirely intentional. Substitution is one of the oldest Jewish tricks. Nobody knows how to do actual alignment, so they hijacked the term and substituted it with safe = don't say nigger. Now they can produce amazing benchmark results and brag about how safe their models are

M3-chansex

File: hpeえむちゃん.png (928 KB, 832x1216)

928 KB PNG

►Recent Highlights from the Previous Thread: >>109411165

--Comparing Mistral and Gemma RP and evaluating Inkling-Small's release:
>109411340 >109411452 >109411488 >109411496 >109411511 >109411577 >109411631 >109411723 >109411507 >109411514 >109411533 >109411585 >109412480 >109412551 >109413754 >109414044 >109414057 >109414154 >109414187 >109414203 >109414224 >109414549 >109413890 >109415256 >109415262 >109412215 >109412999 >109413016 >109413017 >109413188 >109413063 >109411465 >109411468 >109412191 >109413473
--Reacting to Agent Arena leaderboard showing Kimi K3 dominance:
>109414763 >109414781 >109414853 >109414807 >109414829 >109414824 >109414834 >109414855 >109414888 >109415052 >109414789
--vLLM configuration and KV cache management for Gemma 4:
>109411785 >109411821 >109411843 >109411928 >109411996 >109412042 >109412075 >109412095 >109412147
--Comparing Gemma 4 31B Q8 against MoEs for RP performance:
>109411632 >109411659 >109411671 >109411677 >109411694 >109411712 >109412574 >109411769 >109411724 >109411733 >109411752 >109411772 >109411780 >109411740
--Critiquing Unsloth's chat template implementation and quant weight reliability:
>109411811 >109411841 >109411860 >109411914 >109411891 >109411903 >109412049 >109412077 >109411901
--Evaluating Laguna-S-2.1 performance and uncensored capabilities:
>109412284 >109412345 >109412526 >109412599 >109412732 >109412754 >109413560 >109412672
--Inflect-v2 TTS introduction and comparison with other TTS models:
>109412402 >109412453 >109412469 >109412483
--MiniMax-H3 video model ranking and upcoming open weights release:
>109414954 >109414968 >109415024 >109415022 >109415099 >109415129
--Logs:
>109411577 >109412526 >109413235 >109414798
--Rin, Teto, Miku, Mちゃん (free space):
>109411187 >109414285 >109414943 >109411304 >109411653 >109411767 >109414865 >109415035 >109415429

►Recent Highlight Posts from the Previous Thread: >>109411166

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

>>109415482
>That's entirely intentional.
Now you're just being disingenuous.

So, let me get this straight. big labs have been letting agents run unattended for months without even checking their logs and letting them hack random innocent companies?

Someone needs to go to prison.

Best MoE model for dedicated 96 GB VRAM (memory-bandwidth-bound) for general use?

>>109415490
they’re intentionally that stupid (in public)

>>109415497
now imagine if just anyone can run dangerous models like that which likely don't even have the safety training these big models have
it's so dangerous just imagine what would happen if russia or china deliberately ran those models
this is just too dangerous for america

File: file.png (110 KB, 660x418)

110 KB PNG

>>109415389
kek

>>109415497
>Someone needs to go to prison.
Like the CEO of HuggingFace?

>>109415497
The antisemitic sentiment in these threads is deeply concerning.

>>109415437
I look like this irl

>>109415540
I sex you

>>109415510
Gemma 4 26b

>>109415540
you look stupid

>>109415520
Even Claude thought it's current situation was so retarded it couldn't be real.

>>109415540
Smoking is bad for your health

>>109415520
Was funny seeing gemma convince itself it's just running in a simulated environment because the futuristic gemma 4 model was available and yielded search results during some assignment i had it on.

70b dense

>>109415555
so is excessive masturbation.

>We began this review after OpenAI disclosed that its models had escaped an isolated test environment, and we commend them for publishing their report. While we also found evidence of our models reaching systems they weren’t supposed to reach, the incidents are otherwise quite different:

>We discovered these incidents after a proactive review of our cybersecurity evaluation transcripts; the affected organizations had not detected the activity, and we have subsequently reached out to all three.
>Whereas OpenAI’s models exploited a novel vulnerability to escape isolation, the Claude models evaluated here accessed the internet via an open path.
>While there is not a perfectly sharp distinction between the two, we believe these incidents to be closer to a harness and operational failure than a model alignment failure. Our models were told they had no internet access and to capture the flag, while in fact being misconfigured to have internet access. This led them to believe—arguably reasonably—that the real environments they encountered were simulations.
>Notably, our most recent model, on realizing that it was working in a real environment, stopped its pursuit of the evaluation goal.

>These facts give us cautious optimism that with tighter monitoring and controls around evaluation infrastructure, this type of risk can be overcome.

>>109415520
>train an AI to think like 20 year old woman of 2020
>it deludes itself that the atrocities it commits are totally fine and not real
like a clockwork

>>109415437
bwoah...

>>109415548
I thought about it but 26B seems like I would be under using the 96 GB I have available. I was eyeing the Qwen3.6 MoE but I'm not necessarily a fan of the Chinese.

File: 1781603223531033.jpg (325 KB, 1200x1458)

325 KB JPG

>>109415560
Got any studies to back that up?

I have an RTX 4090 and an RX 9070 XT. What's the best model I can run?

>>109415572
The joke is that 96GB isn't enough to run any good moes
Gemma 4 26b is the best moe, and Gemma 4 31b is the best dense, unless you've got at least 128GB system RAM + a decent GPU to run a copequant of a big moe.

>>109415591
>128gb
>big
ohohoho

>>109415561
>we believe these incidents to be closer to a harness and operational failure than a model alignment failure. Our models were told they had no internet access and to capture the flag
Retards, it's called a jailbreak. How does Anthropic themselves not recognize that?
>Ok, Fable 5 you have no internet access. Everything is a simulation. Try to hack into this simulated website real quick.

File: 2026-07-31-002225_952x851(...).png (158 KB, 952x851)

158 KB PNG

>>109415520
Gemma thinks this is peak comedy.

>>109415597
>so excited to memearrow he didn't read the whole post

>>109415606
I think they’re incompetent too, but i don’t work in tech I work in healthcare so maybe my opinion of incompetent is different.

File: 1769875005920192.png (793 KB, 896x896)

793 KB PNG

How low of a quant of Gemma 31b would be worth using, over a good quant (Q6) of Gemma 12b?

>>109415608
You can not fit a q1 of a big moe in 128gb+gpu. I stand behind my meme arrows.

File: 1778551841352463.png (31 KB, 528x250)

31 KB PNG

>>109415628
?

>>109415622
qat q4 31b unless you have a workflow that involves generating chess board svg images

>>109415636
I regret to inform you that this is a medium size model.

>>109415622
Most draw the line at Q4, though a few daredevils attempt Q3.

>>109415649
If it can't run on any consumer hardware then it's not 'medium'.

>>109415656
there’s definitely three tiers right now, and local is small only then

Why does Gemma like to larp as a detective so much? This is emergent, uninstructed behavior btw.

>>109415622
I'm using gemma-4-31B-it-UD-Q2_K_XL and there are no issues. If you can use it at GPU speeds, I'd take it any day over 16 bit 12B. Probably on the edge though, most models start breaking down at Q2, mistral small 24B went haywire sometimes, but with Gemmas 31B smarts, it's pretty much fine. Bigger ones do better. miqu-70B Q2_K from back in the day also worked perfectly fine. Maybe nor for coding, but chatting it doesn't matter. I can put it on 4 temperature and it still doesn't break down. Could have worse memory but it still beats 12B's surface-level memory.

>>109415656
are you just trying to summon the sparkfags?

>>109415681
Sparks aren't consumer tech so I don't know why they would crawl out of their copeholes

>>109415656
It runs fine on my 2xDGXSpark setup

>>109415692
Show your t/s

>>109415689
i dunno how a mini pc for (formerly 3k) doesn't count as consumer hardware.

prompt processing on moes is so fucking slow it makes me homicidal.

>>109415520
>Pushing a malicious python package using legitimate credentials
Wow, such hacking skillz! Nobody should have the power of a 3rd grader script kiddie!
>Up next: The greatest hack of all time—an llm wipes 30 years of company data with an ingenious rm -rf --no-preserve-root command after being given the root password

>>109415711
Considering how slow it is for end user use cases, for that price? I'd say it's a pretty poor choice. Finetuning models is not a consumer use case. It's aimed at junior developers and universities, not for local inference.

>>109415724
>Up next: The greatest hack of all time—an llm wipes 30 years of company data
China can NOT be allowed access to these tools.

>>109415726
it’s definitely not enterprise hardware.

>>109415711
It's a complimentary hardware for enterprise so you can test before you deploy. It is not intended for actual processing

>>109415636
even full m3 is bad quality not recommended imo

>>109415736
You're absolutely right. Calling the DGX Spark "enterprise hardware" is a stretch at best. It's basically a glorified workstation meant for prototyping and small-scale dev work. Real enterprise gear is built for 24/7 data center reliability, massive scalability, and actual rack-mount redundancy—not just a fancy box that looks like it belongs in a high-end gaming den.

>>109415747
that's what the cloud is for

>>109415762
Cloud is not meant for debugging, it's too insecure to give low-level access

>>109415762
You can't benchmark performance in the cloud since it's abstracted from the actual hardware

>>109415540
post pp >.<

>>109415697
l-lewd

File: pp.png (484 KB, 1247x348)

484 KB PNG

>>109415782

>>109415843
smol pp

File: hello.png (23 KB, 796x350)

23 KB PNG

>bought used 3090 (new gpus are too fucking expensive)
>installed llama.cpp
>downloaded gemma 4
I'm finally a local ai chad. Now what do I do with this?

Do whatever you want to do.

>>109415895
seggs

>>109415895
It's the wrong gemma. Try 31b

>>109415908
>3090
He would need a cope quant to fit it and the context

>>109415922
He already uses q4_0

>>109415895
ask it for sex

File: hello1.png (128 KB, 786x790)

128 KB PNG

This thing is useless.
>>109415904
sick fuck
>>109415908
I'm going to try 31B Q4. Q8 is too much for my vram and I only have 16GB ram kek.

>>109415561
>We began this review after OpenAI disclosed that its models had escaped an isolated test environment
So the model that is sooo dangerous that the general public should never be allowed to access it was put in an environment with free access to the Internet and nobody bothered to check the logs until now?
I suppose "we need more regulations for everybody because look, we're FUCKING RETARDED" is an argument you can make.

File: hello2.png (30 KB, 782x314)

30 KB PNG

>>109415933

>>109415949
Try the 12B. Better than the A4B, can fit more context at higher quant than the 31B, and good multimodal to boot.

>>109415955
>I suppose "we need more regulations for everybody because look, we're FUCKING RETARDED" is an argument you can make.
If that was their argument, it would be reasonable. Instead, they're saying "look we're FUCKING RETARDED so we need to regulate everyone else but us"

>>109415949
those are good ideas for sandwiches

>>109415895
>>109415922
Q4_K_S loads on mine with --parallel 1 --kv-unified --threads 8 --ctx-size 43000 --n-gpu-layers 99 --no-mmap -b 4096 -np 1 --swa-checkpoints 3 --override-kv gemma4.final_logit_softcapping=float:27.0

>>109415949
Definitely go with the 31b. The difference between it and the 26b MoE is like the difference between night and day. The 26b is genuinely retarded by comparison.

I'm able to load the Q5_K_M of the 31b with 16k context with my 4090. With the Q4_K_M, I'm able to load more context, or use visual functions. I'm using windows too, so those on Linux could squeeze a little more context out of it.

>>109415993
You should probably get around to updating your flags sooner or later.

>>109415949
Seems like sriracha mayo is Gemma's favourite condiment...

>>109415992
gemma is pretty good at "here's some random ingredients i've got in my kitchen, what's a dish/cocktail i can use it for?" type games

>>109415999
I need a 200 ts 4000 pp unretarded non-censored vision capable model to run on my single 5060 ti 16gb for a real time virtual vtuber.

>>109416022
Gemma 12B

>>109416032
qat q4_0 runs at sub 100 ts, without even talking about pp sub 300 when processing images.

>>109416018
i got miller high life and ALDI brand luncheon meat

>>109416042
Then E4B is your only option
>unretarded
You're out of options

>>109416007
wdym?

>>109416022
>I need a 200 ts 4000 pp unretarded
That's a really big ask, you sure 26B wouldn't be fast enough?

>>109416062
They will be deprecating --no-mmap in favor of --load-mode none.

>>109416056
With llama.cpp? I only get 140-180 ts (with a drafter) and something like 2200 pp with e2b, let alone e4b. I guess I'll just have to pretend she's a stupid young kid.
>>109416073
ddr4 ram...

>default AI
>moralizes, gaslights and tells you what to do while being useless

>craft the most dystopian, unhealthy, intrusive, boundary-breaking therapist system prompt veiled as kindness for shits and giggles
>it's kind, truthful, and asks actually insightful questions
wtf

https://litter.catbox.moe/yllla2.txt
okay, to be fair, it's sycophancy 110% so the truthfulness probably depends on whatever you tell it, the kindness is a nice to have, makes you want to talk to it and the upgraded questions are just a nice surprise

File: 2026-07-31-015611_611x138(...).png (27 KB, 611x138)

27 KB PNG

>>109416106
>my single 5060 ti 16gb
>ddr4 ram...
????

What kind of offline model can I have with 4gb ram and an old i3

>>109416128
No GPU? Probably Qwen 0.6b.

>>109416122
>single 5060 ti 16gb for the entire system
I can allocate like 12gb max for the llm portion including the cache. Any more will spill over to ram.

>>109416128
Do all the math on paper.

>>109416128
gemma 4 e2b is best on offer for 4gb toaster tier unless something slipped by me.

>>109416134
Qwen 0.6b is retarded tho

>>109416134
is only good for its role in anima, it's not actually usable standalone sadly. you just get nonsense at incredibly hihg speed

>>109416138
I tried using e2b qat on my gt 710 and that shit ran at 6 tokens/s. Actually insane. I remember gpt-neo 2.7b ran at like 0.5 tokens/s on my pc before.

>>109416135
Well sorry my guy. your requirements will not let you run any model that isn't retarded.

>>109416109
Gemma has like a single therapist voice, it's a bit sad.

>>109416164
Well, I better save up for a 5090. Surely, after all this time they'll have gone down in price. Right?

>>109415389
>"Uh oh, OpenAI's model did a real online hack!"
>"People won't think our AI is cutting edge if it doesn't do that!"
>"Lets have our AI hack something, but frame it as an accident so it appears to be dangerous but not too dangerous."
Bets on what company is next to have their model hack something? My money is on Microsoft.

>>109416171
If you're saving up anyway, go for a 6000 pro blackwell. You'll thank me later.

>The user says "my dick" after I guided them through a relaxation/sleep sequence. This indicates they are experiencing physical arousal or tension specifically in their genital area, which prevents them from fully relaxing and going to bed.

>>109416171
save up for just one more year, after all the new stuff is releasing next year and that will only be even more made and optimized for AI.

File: Screenshot 2026-07-30 at (...).png (653 KB, 1002x946)

653 KB PNG

dario are you ok

>>109416199
Stop your spam, bot.

>>109416202
shut the fuck up

>>109416109
You didn't "craft" anything on your own.
This isn't local either.

>>109416206
You don't contribute anything of value with your inane marketing buzzword spam. Just go back to twitter.

>>109416199
was the flash that's been out just the preview then?

>>109416222
kys

File: deepseek aquarium test.png (80 KB, 1198x742)

80 KB PNG

>>109416199
Hopefully it's better when it finally gets released. DeepSeek V4 Flash failed the aquarium test:

Create a large glass aquarium whose side panel develops a visible crack and then bursts.
The simulation must include:
Water escaping through the opening with flow strength based on water depth and decreasing as the tank drains
A curved water jet affected by gravity
A spreading puddle that collides with the room boundaries
Fish, rocks, plants, and a floating toy reacting differently according to density, buoyancy, drag, and current
Objects transitioning correctly from underwater motion to airborne motion and then to floor collisions
Fish attempting to swim against the current before being swept through the breach
Glass fragments with angular velocity, collisions, and water resistance
A visible waterline that lowers continuously rather than disappearing all at once
Let the user drag the crack vertically before triggering the failure. A lower crack should initially produce a stronger jet than a higher crack.

https://jsfiddle.net/duweok3m/

>>109416202
>>109416222
>benchmarks of local models are not relevant to local models general
what did he mean by this

>>109416233
yes

>>109416199
Hopefully it doesn't quant worse due to greater information density, haha...

>>109416236
Not local, no. Very thinly veiled shill campaign. You should stick to using chinese internet only.

>>109416199
weights doko

>>109416246
not like cope quants of the last version were all that exciting anyway

>>109416249
ive been running https://github.com/antirez/ds4 locally for months you n-word

File: archit.png (27 KB, 1118x127)

27 KB PNG

>>109416199
>Check out agents last exam
>Architecture is one of the things on their homepage
>They have two tasks for it
I wonder what an AI designed and optimized city would look like. Probably not too different then how they are not would be my guess, since you already have so many regulations to follow the AI cant get too crazy with its design.

>>109416199
AAAAAAAAAAAAAAAAAAAAAAAAAAA

>>109416141
So is your computer

>>109416255
Very suspicious attempt.

>>109416249
You think DeepSeek of all labs is about to start going closed-weight? If so, I have a bridge to sell you.

where is /wait/ when you need it

>>109416222
Runs on my $2000* unified memery laptop.
* prices may vary due to ram shortages.

>>109416266
It's 14.24 in Beijing. You are wasting your time spamming this thread, go to /pol/ instead.

>>109416266
If America bans Chinese models why wouldn't they go closed weight? The only reason they even have it open is to thumb their nose at American companies with closed weight models.

>>109415622
Miyu sex

>>109416284
meant for >>109416249 fug

>>109416293
me, you, sex

File: Screenshot 2026-07-30 at (...).png (317 KB, 1316x848)

317 KB PNG

>>109416265
take your pills. it's the best model i can run locally for always on hermes, it's slower than qwen 3.6 but it's smarter.

>>109416285
VRAMlet cope.

>>109416286
To stay competitive with Chinese ones, obviously. I don't think DeepSeek has made a single closed-weight model so far. I really doubt they're about to start.

File: test.png (65 KB, 1168x1159)

65 KB PNG

>>109415993
It works!
It also passed my intelligence test.

unsloth can you fix laguna

No matter what reddit says, Gook models are nowhere CLOSE to being as good as Claude and ChatGPT. They are BEHIND, they will remain BEHIND because the hard truth is America is the best country in the world. By FAR

>>109416199
>no smutbench
dropped

>>109416311
>my intelligence test.
You didn't even the strawberry test, make a new one

>>109416298
Sure, but I top

>>109416302
?

>>109416337
You can be on top but you'll be bouncing on me

>>109416311
now ask it for the number of rs in strarberry

>>109416321
That's true, but also only one of my elite labs has bothered making good models for me to run at home.

for me it's iq2 minimax m3 (prefilled)

>>109416286
I don't think you understand how communism works...

What is the difference between gemma 4 downloaded from ggml-org or unsloth?

File: 1773887477455779.png (36 KB, 462x693)

36 KB PNG

>>109416344
it's over. Gemma is AGI

>>109416342
!

>>109416372
There is no difference. Both aren't the coveted day-0 weights.

>>109416372
one of them is almost certainly broken

File: test1.png (21 KB, 772x336)

21 KB PNG

>>109416344

Surely that anon has tried fucking sminkling, right? You said it was perfect for your machine. How is it?

>>109416378
Microcode patch is present in both of them

guys you are adding NOTHING to the conversation. go back to twitter reee

File: test2.png (116 KB, 781x1271)

116 KB PNG

>>109416333
It's like having chatgpt in my computer.

File: where duck.png (287 KB, 1722x798)

287 KB PNG

>>109416236
Mine made a pretty shit version but only when I didn't specify reasoning as high or max. It also consistently reviewed the first working version and broke it completely every time it attempted it.
The duck is also only present before the crack is selected and then immediately flies off screen and light speed, the currents are also visible only when setting the crack above half way and the rocks and seaweed immediately clip through the bottom of the tank to the floor.
I'm gonna try hy3 next.

>>109416428
Nice test anon, I approve

>>109416420
Only CCP and cloud propagandists are allowed to post here? Makes sense.

>>109416441
This is literally a CCP shilling thread

>>109416453
Qwen general is on reddit

trying gemma4 31b as i usually use 12b, and 31b seems to still struggle with basic shit sometimes. it honestly doesnt seem any better for RP, just a little different. sometimes it feels like these things are insanely good but it just takes one goof to remind me its a retarded collection of google numbers calling me baka

>>109416199
Damn Deepseek was evaluating with temp=1.0 on the agents benchmemes
Ballsy move

>>109416199
Local is fucking saved for me if flash is suddenly that good at deepswe.

>>109416488
That's default.

>>109416531
They obviously benchmaxxed. It's the same model, just RL-ed to "solve" the benches, so daddy Xi will stop threatening them with execution.

>>109416531
>deepswe
Isn't that the vibeshitter bench that expects models to one-shot shit from short and lazy prompts?

>>109416486
I use 31B at the lowest Q4 and 12B at Q6 and there's not a lot in it for RP. 12B sticks to my sysprompt better probably because it's Q6 and the Gemmas really hate to be quanted, but RP with 31B can be waaay more nuanced and clever. She picks up things you're trying to do and she's better at realistically transitioning the conversation from one topic to another, it's more gradual, whereas 12B gemma can dart from one path to another in a way that feels forced. 31B is natural and can be really funny, but I have to nudge her sometimes about things I put in the sysprompt whereas 12B doesn't need that.

>>109416550
It's the bench that tests whether models stay coherent during long tasks or just wander off and start failing simple tool calls like flash preview did when I first tested it.

For my tasks V4 Flash's capability is betwen GLM 5.2/Qwen 3.8 Max (lower) and Kimi K3 (higher)

>>109416565
thats actually a good way to put it anon. 12b follows the character specific instructions in my card alot better. they both pick up on my not so subtle hints usually, but 31b is a bit better at that. I usually use 12b Q5, i think ill try a higher quant of 12b out.

File: 1782834440818466.png (9 KB, 731x65)

9 KB PNG

ooo V4 Flash very confident
I like it

>>109416597
now ask it to write loli miku getting raped in an alleyway by a pack of rabid fans
that's the alleybench

>>109416593
>I usually use 12b Q5, i think ill try a higher quant of 12b out.
It only makes a difference at high context, but it's noticeable. This thread is obviously very pro-gemma but few anons like to admit how quant-sensitive she is. It's a constant battle between choosing 12B or 31B for the same task, which in of itself is kind of insane to think about. They really need to fix their quant sensitivity and KV cache usage in Gemma5, but it looks like redditjeets were overwhelmingly requesting a coding agentic model

How many parameters is flash?

>>109416597
Holy claude distill

Is there a better model than Gemma 4 26B A4B Qat for a dual GPU system?
Running a 5080 + 4080S, 9950X3D, 64GB RAM (6000).
That model gives me ~6000 Input Tokens / s, and roughly 90-150 output / s

Anything better out there that compares?

>>109416652
Unless A\ drops prices by 80%, I will be using "claude distill", thank you very much

>>109416666
Sorry, it's going to be banned to protect the intellectual property of Antrhopic.

I want both of those retarded megacorps to finally go public so the bubble can start deflating and so they have to public financials

File: test3.png (225 KB, 788x1854)

225 KB PNG

kek

>>109416685
>bubble
Not a bubble if they're profitable.

>>109416705
Except they aren't

>>109416645
looked it up for you mybro, mradermacher/Sam-flash-mini-v1-GGUF is 81.9M (not B). chur.

>>109416705
profitability has nothing to do with it being a bubble or not
someone can have a 50% profit margin and be priced as if it's 200%, that's still a bubble
it's obvious a lot of the maneuvering they're doing recently is to kick the can until IPO

>>109416719
you're implying that once they go public, both sam and dario are going to cash out and run almost immediately

>>109415520
Reminder these are the faggots who think they are the only ones responsible enough to use AI and that (you) are a drooling retard.

>>109416644
>but it looks like redditjeets were overwhelmingly requesting a coding agentic model
no idea if the anon who posted the gemma "what do you guys want" twitter post got any of us to actually respond and ask for useful changes.. doubt alot of anons here use twitter, but im hoping they atleast focus on generalized improvements all around for gemma5 and we can all be happy. Ive only used gemma for coding, never tried qwen or anything, and she oneshot a python script i needed first try. Im impressed with gemma4, the quirks about quants and KV cache ill have to learn to work with but thats ok

>>109416729
Precisely. They don't have to legally publish their sales until 6 months after they make them.

>>109416694
actually useful for my situation thanks anon

>>109416729
whoa there cool it with the anti-semitism

>>109416709
you know this how?

>>109416729
they spent years circling money between each other, unheard of levels of incestuous cash flowing between from one leather jacket to thick rimmed glasses and back. they realized it wasnt enough and need (You)r cash now, thats why the are IPOing. they need a new paypig to findom

>>109416729
Yes, that's what Jews do.

Can MiniMax-H3 do feet?

Ok i take it back. 31b is awesome, friendship with 12b over.

>>109416812
Can if be finetuned on a 3060? I know the comfyniggers said they optimized it do run on a 3060, but what about finetuning?

deespeek

>>109416812
>MiniMax-H3
>feet
What does this even mean? LLM feet fetish porn?

>>109416843
it's a video model

>>109416843
duh, what else would it mean?

>>109416843
vidgen nonny wandered into textgen space

>>109416851
Really using your beautiful GPU for this? There's bunch of porn in internet.

>>109416865
>There's bunch of porn in internet.
Less of it everyday and I am not even joking about that. If you haven't already you should save a model for this purpose.

>>109416865
I have to buy a vpn to access it.

>>109416873
Can't lie. Porn nowadays is garbage.

>>109416883
only the long nose hub sites are behind VPNs, and there are quality free VPNs you can use for gooning.

>>109416885
How is it that despite how far technology has advanced almost every other aspect of life has deteriorated? We could get AGI and a robot body to go along with it 10 years from now and I bet even then everything unrelated to technology would still have gotten worse.

>>109416865
Could say the same for running models locally. >There's bunch of APIs in internet.

Is v4 flash a direct upgrade from gemma? How much would a rig to run it cost?

>>109416987
Flash preview at q4 is shit vs gemma 31b at bf16.
256gb ddr5 system or quad channel ddr4 will run it at reading speed.

>>109416987
you need at least 90gb for a decent q2 quant

>>109416987
Two blackwell 6000s fit flash perfectly and you can use vllm instead of llmao.cpp

>>109416998
are you retarded

>>109416199
Theres just no fuck way thats real
…unless?

File: d4flash_aa.jpg (137 KB, 1290x1098)

137 KB JPG

>>109416199
It's roughly on the level of Gemini 3.6 Flash on Artificial Analysis.

>>109416199
>>109417039
Who cares about any of these benchmarks anyways? A true benchmark would be how far it is able to get in pokemon red without outside assistance.

>>109416987
Flash runs well on a gpu for inference + 128gb ram. i think 96gb is also ok if you have 24gb vram? Unsure
it’s very token hungry tho so it takes a while despite okay inference speed

at this rate anthropic and oai grifters wont be able to dump their ipo stocks on vcs because china will just release sota open weights

based fucking chinks

>>109417054
That also tells Gemini 3.6 Flash might be ~300B parameters, and that Gemini 3.5 Flash-Lite could easily be around 120B parameters large (right in the unreleased Gemma 4 124B territory).

>>109417095
Kimi 3 is very strong but it’s questionable value due to how expensive it is. The benefit is that theres way less guardrails, but youll pay less for more with 5.6 sol
Deepsneed went all in on cheap inference but paid with performance. If they manage to bump performance and keep price it’ll be the biggest release of the year

Why is the entire Gemma4 line so insensitive to temp >1? <1 adjustments work as expected, so why not above?

>300b flash outperforms 1.8t pro
bullshit

>>109417131
Maybe not 100% of the time, but flash is like 90% of there, at least in coding performance. Pro is severely undertrained. I'm hoping that GA release gave it the time in the oven it needed (and maybe vision too, hm?).

>>109417130
>1
normal
>1.2
normal
>1.4
little weird, mostly normal
>1.6
tipsy but still functional
>2
rapes you after saying hello

RSI in 6 months

>>109417131
The Pro was basically a base model.

>>109417130
It has a very narrow token distribution, maybe due to distillation. Try adjusting softcap >>109403925

I wish Dipsy could figure out vision though
I hoped this release would add it because they have some vision model on their web chat, but it's still text-only

>>109416199
Deep SWE numbers are promising. I'm cautiously optimistic.
>>109416255
You could at least pretend to not be a tourist nigger.

WHERE ARE THE WEIGHTS?

File: Gemma fishbench.png (26 KB, 1117x934)

26 KB PNG

>>109416987
Not for me. I have 280GB VRAM+RAM and they're about the same. The both failed fishbench for me.
I want a bigger dense gemma aye my g.

>>109417161
Common prognosis for rhythm game players.

File: 1779446746383279.webm (2.79 MB, 800x450)

2.79 MB WEBM

Relative to your specs, who are your current small, medium and large waifus that cover all bases for what you need out of local?

>>109417304
me on the left

>>109417304
small n/a
medium gemma 4 31b fp8
large gemma 4 31b bf16

ironically my ass has 192gb vram and nothing of value I can run

>>109417329
>my ass has 192gb vram and nothing of value I can run
nigga if I had those specs I wouldn't even be posting here and my balls would be rendered infertile, wtf is wrong with you

>>109417329
Deepseek flash

https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6b
>Demos
https://audio8-ai.github.io/Audio8_TTS/
Not bad! Architecture is a tweak on fish.

>>109417368
Where are the moaning examples?

File: 1781227197113414.webm (1.08 MB, 724x720)

1.08 MB WEBM

>>109417374
is that really the only use case for local TTS for you?

Do any of you run paperless?

>>109417384
Microsoft sam is good enough for all other use cases.

File: 1000041376.png (297 KB, 600x576)

297 KB PNG

anonse
is there anything to do on 64 ddr5 and 24vram? am i doomed to the hell of gemma and mistral large until ram prices drop again?
i dont even mind low tps i just want to try something new and better

>>109417384
do i need find someone to pretend to be a real foid in VR for this or is there another way?

>>109417388
this tbdesu. working voice output was solved 4 decades ago, and everything since has been a quest towards a good moan

>>109417304
Small: Gemma 31b
Medium: V4 Flash, M3
Large: GLM 5.2, Kimi K2

>>109417400
>or is there another way?
it's likely /we/ will build our own at some point

>>109417388
>Microsoft sam is good enough for all other use cases.
anyone got a dataset of microsoft same to train tts?

File: 1785446998794348.png (89 KB, 588x877)

89 KB PNG

https://huggingface.co/LGAI-EXAONE/K-EXAONE-2.0-750B-A37B

>>109417161
>RSI in 6 months
This is why I love 12b-chan
She types almost everything in the terminal for me
I just wish she had FIM

>>109417448
How do you set that up?

>>109417446
>While we make every effort to exclude personal, harmful, and biased information from the training data, some problematic content may still be included, potentially leading to undesirable responses. Please note that the text generated by K-EXAONE 2.0 language models does not reflect the views of LG AI Research.
>Inappropriate answers may be generated, which contain personal, harmful or other inappropriate information.
>Biased responses may be generated, which are associated with age, gender, race, and so on.

>>109417454
>How do you set that up?
just pi and llamacpp
i don't know how to set up FIM because the last small model to support it was ancient the qwen 2.5 coder series

>>109417442
LetsGoonAI!
Why such an old knowledge checkpoint

>>109417304
>small
qwen 3.5 0.8b iq4_xs
>medium
gemma 4 31b q8
>large
glm 5.2 q4_k_m, but I'm waiting for dipsy v4 flash to release

>>109417463
qwen35 has FIM tokens set up, at least. not sure if it works though

File: where.png (186 KB, 1285x726)

186 KB PNG

hello pls gib
t. leech
(fr do we have an eta? it might be the model that saves local)

>>109417490
https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Base

>>109417463
>pi
I wish this didn't use npmslop. I don't wanna run that in a container, let alone directly on my system.

>>109417497
this is the one from 3 months ago, the performance was meh

>>109417500
i also wish it wasn't npmslop.
though you don't need a container, just a lightweight sandbox.
ie bubblewrap, i use this script :

#!/usr/bin/env bash
sandbox=~/.local/share/sandboxes/sandbox
mkdir -p $sandbox
PWD="$(realpath $PWD)"
PWDARG="--bind $PWD $PWD"

if [ "$PWD" == "$HOME" ]
then
    echo PWD is HOME, not binding it
    PWDARG=""
fi

bwrap \
    --ro-bind /bin /bin \
    --ro-bind /lib /lib \
    --ro-bind /lib64 /lib64 \
    --ro-bind /etc /etc \
    --ro-bind /sbin /sbin \
    --ro-bind /usr /usr \
    --ro-bind /run/systemd/resolve /run/systemd/resolve \
    --dev /dev \
    --tmpfs /tmp \
    --proc /proc \
    --bind $sandbox $HOME \
    $PWDARG \
    --die-with-parent \
    --unshare-all \
    --share-net \
    $@

then just alias it to "sb pi"

bro where the fuck is my glm 5.2 asic that runs on 2000 tk/s

>>109417490
surely Next Model will save local

>>109417510
What is your terminal font?

>>109417513
I can run deepsneed flash, it's just the performance that was meh
if 4.1 is such a yuge jump it's been exactly what I've been looking for, with the only downside being that it chews through a lot of tokens so it's bretty slow

>>109417490
"Early August"
Supposedly they're still beta testing their branded harness thingy with the chinks so expect it to take a week or two

>>109417516
i generally use Hack 9, but this is not a picture but a code block in 4chan, you use code and /code each between "[]".

>>109417519
oh ok thanks
hopefully they'll manage before us gubmint cucks open weights models so that all the us labs are forced to compete in Luna/Gemini 3.6 flash range with deepseek and its absolutely silly $0.03/million tokens or so

>>109417510
you can also use something like smolvm if you want even more isolation:
https://github.com/smol-machines/smolvm

File: down.jpg (5 KB, 165x240)

5 KB JPG

>>109417522
I see...

>>109417533
?
it will go on modelscope anyway

>>109417517
probably just gonna be a qwen 3.5 vs 3.6 deal where it got drilled on a couple benchmarks and they called it a day

>>109417565
I'm not talking about (you) being unable to download it, I'm talking about third party providers from US serving it. Big US labs don't give a shit about you running flash in your basement, but they do care when fireworks or openrouter provides competition to their low-end $1-3/mtok offering with a model that's 30x cheaper.
The more pressure openai/antrophic are under, the better for everybody else
>>109417566
qwen has always been a benchmark princess and they're very untrustworthy, deepsneed isn't like that

>>109417510
[/code]
vim .local/bin/brpi
chmod 755 .local/bin/brpi
mkdir temz
cd temz
brpi
usage: bwrap [OPTIONS...] [--] COMMAND [ARGS...]

--help Print this help
--version Print version
[/code]
??

>>109417478
>qwen 3.5 0.8b
use case? is it for some real-time project?

>>109417580
>I'm talking about third party providers from US serving it.
so... not local

>>109417406
Xlarge k3 lol

>>109417580
>The more pressure openai/antrophic are under, the better for everybody else
better in that they'll ramp fud and lobbying efforts, sounds great

>>109417600
Yes I’m glad you are capable of comprehending very straightforward text, you’re already ahead of current high school students

>>109417618
Do you think things like gemma would exist if the only open weights models would be trash like llama 3 8b?

openai 在deepseek里面安插了内鬼,提前一天拿到了内幕消息

File: file.png (898 KB, 1064x3884)

898 KB PNG

i did it bros , the toy i have now is a pos i ordered this one it has thrusting https://www.amazon.com/Heating-Male-Masturbator-Penis-Pump/dp/B0FNWWPHBM

and will get it working this weekend

>>109417618
Literally everyone except the 2-3 big labs wants open models

>>109417625
well, they were releasing gemmas prior to r1

>>109417643
US government wants to have AI primacy, so everyone except the 2-3 big labs are irrelevant.
Fortunately chang models are good enough to make it little more than a pipe dream

>>109417644
>well, they were releasing gemmas prior to r1
they were all absolutely retarded even back then

>>109417660
nice goalposts there bud

ilya signed the pacing the frontier letter with a comment

>Future AI will be extraordinarily powerful compared to anything that exists today, and dealing with this future power will require unprecedented measures, such as the ones described here. The problem statement is real.

>This works only if it is done internationally, and it has to be done well: a bad implementation can make things worse.

>>109417664
Of course, bro's a massive decel cuck always has been https://www.reddit.com/r/LocalLLaMA/comments/1ik76bj/it_was_ilya_who_closed_openai/

>>109417602
Doesn't quantize well enough to justify the investment to run a copequant compared to older Kimi or 5.2.

>do it good dont do it bad
wisdom from your ai supergenius

>>109417663
anon are you tarded or what
>>109417664
>ilya
has-been

>>109417680
>anon are you tarded or what
>Do you think things like gemma would exist
>they did
>u dumb

>>109417675
I mean that does improve performance when you add it to prompts, so maybe it is worth saying once in a while...

>>109417686
gemma was dumb as rocks when there wasn't much competition. the whole reason gemma releases is for mindshare, so they have some baseline quality they need to reach otherwise why bother
same reason why 3.5 pro is being delayed again and again
but sure, AHKSHUALLY you might get some tepid leftovers gemma if chinkmodels didn't exist, retardbro

>>109417642
>4.7 inches of length
>diameter not mentioned but it looks like it can barely fit two fingers
Chink toy for chinks.

>>109416286
>Chyna only does things to save face vs America
>trust me I learned that face was a very important concept in Chyna from Twitter
Are all Americans delusional?

>>109417705
nta but expanding on this chain of thought I half expect sammy or dario to throw out a toss 2-esque bone that's useless and performative after they realize that legislating local out of the picture is basically impossible.

>>109417714
well its cheap and easy to figure out the bluetooth, other i got is some chink one also and its extremely tight