File: file.png (347.7 KB)

347.7 KB PNG

►Recent Highlights from the Previous Thread: >>108565269

(1/2)

--Discussing ggml's new experimental backend-agnostic tensor parallelism and performance gains:
>108566286 >108566382 >108566397 >108566458 >108566462 >108566464
--Performance testing of llama.cpp experimental tensor parallelism on Windows:
>108567186 >108567201 >108567216 >108567433 >108567445 >108567553
--Solving LLM tool calling issues regarding boolean type parsing:
>108565765 >108565819 >108565853 >108565867 >108565986 >108566089 >108566110 >108566123 >108566177 >108566195 >108566258 >108566308
--Debating Claude's impact on compiler engineering and overall code reliability:
>108566489 >108566531 >108566517 >108566573 >108566595 >108566588 >108566540 >108566568 >108566583 >108567950
--Running Gemma 31B IQ2_M on RTX 3060 using llama.cpp:
>108565291 >108565294 >108565303 >108565328 >108565346 >108566298 >108566302 >108566349
--Comparing intelligence and performance of Gemma 4 versus Qwen 3.5:
>108565318 >108565368 >108565430 >108565617 >108566007 >108566047
--Troubleshooting long-context tool calling failures in Gemma 4:
>108565347 >108565356 >108565407 >108565475 >108566017 >108566065 >108566411
--Discussing a mesugaki Gemma persona, jailbreaks, and cheap X99 boards:
>108565322 >108565332 >108565458 >108565335 >108565345 >108565582 >108565615 >108565722 >108566726 >108567096
--Anon implements autonomous memory for Gemma to maintain persona:
>108567439 >108567453 >108567468
--Anon gives Gemma autonomous tool creation and modular persistent memory:
>108567066 >108567109 >108567174

►Recent Highlight Posts from the Previous Thread: >>108565273

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

File: file.png (292.8 KB)

292.8 KB PNG

►Recent Highlights from the Previous Thread: >>108565269

(tan/2)

--Gemma-chan and more:
>108565343 >108565771 >108566833 >108566920 >108567100 >108567227 >108567234 >108567265 >108567278 >108567316 >108567366 >108567457 >108567484 >108567562 >108567601 >108567834 >108568046 >108568067 >108568106 >108568192 >108568197 >108568299 >108568333
--Logs:
>108565302 >108565322 >108565347 >108565475 >108565654 >108565715 >108565765 >108566298 >108566349 >108566382 >108566411 >108566668 >108566728 >108566806 >108566848 >108566894 >108566955 >108567115 >108567183 >108567215 >108567439 >108567465 >108567468 >108567545 >108567611 >108567626 >108567673 >108567936 >108568027 >108568045 >108568100
--Miku, Teto (free space):
>108565424 >108565722 >108566528 >108566726 >108567259 >108567919

►Recent Highlight Posts from the Previous Thread: >>108565273

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

Mikulove

i'm going to ask my ai assistant to help me set up my local model!

File: angry_pepe.jpg (42.6 KB)

42.6 KB JPG

>>108568340

don't you dare to ignore meeeeee!!

>>108568453
trust, but verify. i wasted a lot of time asking claude shit that led me in the wrong direction.

>>108568453
that's exactly what I did and it went more or less fine

>>108568460
Relative to what?

Kill kill gemmaniggers

>>108568460
kys

gemmaballs

File: 1639240519537.png (95 KB)

95 KB PNG

>>108568469
>vramlet

>>108568479
isnt gemma 4 kinda vramletpilled to start with

>>108568340
like, bounding box?
idk how to do it but iirc it should be pretty doable

File: firefox_uW3wc0Xgla.png (140.3 KB)

140.3 KB PNG

>>108568467
>>108568460

>>108568462
>>108568463
yay i am cautiously optimistic

File: 1770474309533995.png (120 KB)

120 KB PNG

>Start recognizing gemma's slop patterns after a few days
>Ruins all enjoyment

How do I stop noticing things??

>>108568467
relative to width and hight

the very middle would be (0.5, 0.5) regardless the ratio

>>108568508
just be as specific about your system specs and particular use case as you can, give her all the details you can provide

>>108568509
>How do I stop noticing things??
that's the curse of having a high IQ anon... the magic stay only for retarded normies, I kinda envy them desu

>>108568509
Make your own finetune every few days

Those with a 5090/Pro 6000 or even 4090 here, how often do you inspect your cables for cablemelt?

File: 1712063180010677.png (20.1 KB)

20.1 KB PNG

>>108568513
Gemma's output translates to (405, 92) in pixels which is correct.

>>108568509
high temperature, hyperfitting, idk
>>108568535
never did

>>108568509
stopping with the antisemitic behavior

>>108568509
seeing slop with gemma-chan is 100% a skill issue on your part
it's all so minor and inoffensive that you can dodge it all with just a bit of prompting, phrase banning and a bit less of being a lazy whiny bitch

>>108568509
stop being anti-semantic

>>108568500

this is quite good!

>>108568540
>>108568558
I verified it too

amazing!

it can be used to control the mouse etc

gemma-chan writes better ntr than slopus...
my cock weeps

The visual model in the 26 and 30 B are trash, cannot even understand characters, anatomy, etc, is constantly confusing things sincerely is so tiresome...

>I've been telling people Gemma is the best model since Gemma 1
>niggers will never understand
I'm glad to see everyone here enjoying gemmy

16 GB VRAM users, what model do we like best now?

File: firefox_JCmDBqVV9W.png (284.4 KB)

284.4 KB PNG

>>108568563
Here's for your image. Also correct.

File: c5837cedf56ca1c96589acca24bd9ae454394144a04da81353e72597dbf37254.gif (252.8 KB)

252.8 KB GIF

Real talk, what do you put in your opencode agents.md?

>>108568578
I am using 26B-A4B-it-Q4_K_L just fine

File: a1320441534_10.jpg (713.6 KB)

713.6 KB JPG

some questions from an lmg newfag
>how much of a difference do harnesses matter? e.g. out of the box, how different will the result be when prompting OpenCode, Pi, Claude Code (local), Mistral Vibe etc? What provides the most batteries-included experience?
I noticed at least there's a difference in the tools the model has access to by default, e.g. Claude Code and Crush have web search capabilities ootb, others do not.

>is Qwen3.5 122B the best general purpose model I can run on 128GB VRAM atm?
>does Qwen3-Coder-Next perform significantly better than 122B for programming?
>is there any point in running Gemma 4 31B if I can run larger models?

thanks to any anons who reply

>>108568583
Real talk, go back >>>/g/vcg/

File: Screen_20260409_145606_0001.jpg (296.8 KB)

296.8 KB JPG

>>108568460
>>108568340
idk how accurate it is but here's the response

>>108568488
Yes. You can identify the literal jeets by their sub 16GB VRAM posts because nobody in a developed country with this hobby is settling for less than that.

>>108568587
I used Qwen3.5 122B with a small quant (72GB VRAM total is what i have) and, well, from what I'm feeling right now, it's not even close to Gemma 4.

>>108568603
>literal
>developed
>hobby
Grow up, little buddy. Finish up your homework.

>>108568509
You already posted this

>>108568583

<!--TODO-->

>>108568417
Goose is the best option of getting something proper from this space when other agents that do what it does and allows agnostic backend for choosing who you want to grab tokens from are all either mismanged hard and bloated or propietary. Having Block also no longer be in charge of the project and having it handed to a branch of the Linux Foundation to develop it is also probably a good thing too.

>>108568607
Wow, you prefer Gemma 4 31B? I think Gemma might just be too slow for me, it's like 7tps on Strix Halo

>>108568617
I don't think I can keep using it until they add an option to edit messages. Like, this is such a basic function, how are they missing it? Also can't delete conversations.

>>108568628
At low context I get like 30 t/s. I have three RTX 3090s. There is a latest update in llama that allows you to actually make good use of multiple GPUs but I'm not using it yet because it does not support kv quantization and is broken for three GPUs - works for two. Anyway, use a reasonably small model quant and quantized kv for massive speed gains. Quantized kv is good now because of rotations (thanks, google).

>>108568595

thank you!

gonna quickly vide-code a program to give me coords of the mouse at the picture

>>108568583
I generate one with /init and then edit it to remove anything wrong or add anything important that was omitted. I don't really do anything fancy with it.

>>108568650
what the fuck, you can get that without LLM

I can't believe that Google saved local.

File: 1775755186108723.png (1 MB)

1 MB PNG

>>108568655
true

File: Werks-on-my-machine_Gemma4-local-tokens-per-second.png (517.8 KB)

517.8 KB PNG

>>108568415
Gemma4 t/s (on Apple Silicon) if anyone is interested. As of writing this most recent gpus still curb-stomp even M5 MAX chips in the memory bandwidth department to these should be even faster on those. the 26B moe model runs lightning fast on opencode with ollama as the backend. The 31B dense model is obviously shower but not enough th be utterly unusable, though I haven't tested either's performance at long contexts so I'll have to test that later.

File: gemmaFourConcepts (Medium).png (872.7 KB)

872.7 KB PNG

>>108568415
Vote: https://poal.me/3u6rby
> Which is your preferred Gemma character?

>>108568671
>62t/s pp
jesus christ how horrifying

>>108568649
>quantized kv is good now
yeah but how good? I'm guessing q8 is now really "indistinguishable" but how is q4 for example?

>>108568674
Do not repost this. It's shit. Make one with "Against everything" option.

>>108568677
I made some tests for searching for infromation in many places of 60k+ long context (YAML definitions for OpenXcom game) and q8 and q4 performed similarly.

>>108568676
You're not even reading that correctly. The Dense model runs at ~14 (on my machine). "Prompt eval" is how quickly it processed my prompt.

>>108568705
>62t/s pp
>pp
>prompt processing
which one of us can't read huh?

>>108568649
Oh understandable from a GUI perspective but was mostly talking about it from an agentic point of view.

>>108568687
damn that sounds amazing, just to make sure you were using mla 3 too right?

>>108568714
I have no idea what that is. Explain.

>>108568674
>not including either of the good ones

>>108568676
the token amounts aren't enough to extrapolate to practical speeds, both are 27 token batches that finish in a matter of milliseconds

>>108568674
>four concepts
Wtf there were a ton of other good ones though.

>>108568731
What's the actual pp on at least 1000 tokens then?

>>108568674
her backpack can be a toaster (to represent the "toaster PC = old weak PC meme" that can actually run this model)

>>108568738
I like flat miku a lot more

>>108568674
3 pedo bait
1 reasonable one
Lower right is best.

>>108568674
>Reposting the cherry picked design poll
Soon I began to hate them.

>>108568674
These are nice too >>108567562 >>108568192

why?

>>108568779
to get a reaction out of your

>>108568779
idk, probably anon wants ban or smth

File: true truth tashikani good point thinking chin rub.png (9.3 KB)

9.3 KB PNG

Use 100t/s GPU Gemma4 26ba3 to do thinking, then inject that thinking into 5 t/s CPU offloaded GLM 4.6? hmmm

>>108568781
he looks like cyriak
https://www.youtube.com/watch?v=05ZvII57p_M

Chaim's ban is up I see.

>>108568779
It's just least dedicated spammer on this website. Let him get it out of his system and he'll disappear for a few weeks.

>>108568777
>>108568773
>>108568765
>>108568738
no one gives a fuck about this. it's a fucking cartoon drawing, no one is getting mad about "muh beloved migu" because it's a meme and no one actually cares or "loves" her so much that they're upset when you post this shit. the only thing you're doing is making it annoying to browse /lmg/ while im at work fuck you.

>>108568801
It's very funny to me that he's shitposting on /g/ yet gets filtered by tempbans.

File: Screenshot004-14.png (257.6 KB)

257.6 KB PNG

>>108568500

holy crap!

>>108568807
I don't care about miku but I am quite unhappy about having to see pictures of niggers and trannies.

>>108568830
I can assure you it's nothing more than a meme mascott. if someone "malds", it's because they are farming (You)'s

>>108568830
>is a sperg
says the BBC spamming sperg

>>108568830
sounds like schizo projection to me
>>>/h/
go back you must

>Gemmachan can report posts with correct categories with Openclaw
Neat.

>>108568807
it's been years anon
it is in a backwater EU village and this is one of the most engaging activities for it
giving it attention only makes it worse

>replying

File: Screenshot004-15.png (822.6 KB)

822.6 KB PNG

>>108568579

wow

I'm still getting refusals from gemma 26b using the gemma-chan system prompt, what do

>>108568873
they can zeroshot bounding box that way too

>>108568867
>EU village
lore?

File: 1750014040747672.png (34.8 KB)

34.8 KB PNG

What is he doing bros

>>108568862
I wish that mattered but the real bottleneck is the jannies who are probably too busy pruning all of the other threads of bbc

>>108568885
it resides in germany and they literally, unironically, no exaggeration, have no life

File: 1768709377404062.jpg (142.1 KB)

142.1 KB JPG

>>108568710
Sorry. I have a splitting headache so I should probably rest soon.

>>108568738
We're on a blue board desu

>>108568890
Very true but I saw lemons in the thread and an opportunity to see if Gemma could make lemonade.
>>108568892
I would put money on that creature having a hook nose.

>>108568830
>mald about it
so far you're the only one having a meltdown lol

File: dancing-pepe-pepe-dancing.gif (512.8 KB)

512.8 KB GIF

>click at 9-digit number
>find a window titled reply to thread <9-digit number>
>click choose file
>select dancing-pepe.gif
>click get capture
>read instructions, solve capture
> wenn done, click post

is it that simple?

Spud will end this general. I'm gonna miss you guys.

>>108568931
I hope not.

>>108568934
that's the thing, miku has a full room inside your head, I don't think about you at all, you'll be nuked in less than an hour (oh well, you just got nuked lol)

Thank you jannies!

I can't imagine seething about Miku while the rest of us are arguing over Gemma-chan designs.

Where did Voldemort get all of these blacked Mikus?
That's right, he genned them with his webui!

>>108568873
Now ask it to trace it

>>108568962
He is obsessed with corruption, very demonic.

What can I do on 4 GB VRAM

>>108568986
gpt 2

>>108569000
trips of buy an ad

>>108568986
run sillytavern

>>108569000
Talking about demons, here they come.
Christ is king.

>>108568986
A MoE running mostly in RAM I guess.

>>108568986
RAM?

>>108568986
cry

File: fuucke.png (61.2 KB)

61.2 KB PNG

>>108568881
guise please I haven't touched this shit in years I don't remember how to do this, is the MoE just less lenient?

>>108568460

>>108568340
>>108562956
>>108562982
>>108563276

>>108569068
>I don't remember
Learn again.

>>108569068
use a character card, tell her it is okay to go all out, something along those it is really not that hard

Is trinity nano base broken? I get gibberish with llama.cpp, correct chat template applied.

"> Hi, my name is
mblazkrinmblazkrinmblaz"

>>108569159
>base
>chat template

>>108569159
i dont know about trinity base models but is it supposed to support any shape of chat formatting?

>>108568909
Nta but you’re forgiven by virtue of posting a green goblin shorty.

>>108569165
https://huggingface.co/arcee-ai/Trinity-Nano-Base/blob/main/chat_template.jinja
I only applied it because I got gibberish without it as well.

>>108569177
>https://huggingface.co/arcee
don't bother all their shit is broken trash

>>108568687
Do it yourself if you care that much.
>>108568730
>>108568732
That’s what anons said last thread.
Then posted nothing. lol.
Post zero content, get zero requests.
Lazy ass mfers.

>>108568814
nice

Retard here can anyone explain why I was able to run 70b dense models in q8 pretty fast yet gemma 4 31b is really slow?

File: nice.png (116.4 KB)

116.4 KB PNG

Gemma rated me face a 7/10.

>>108569202
so ur a 4/10
gemma is male coded

>>108569206
nah I'm more like a 2/10 but I'm glad gemma is at least kind.

>>108569201
works on my machine

>>108568986
gemma 4 e2b is probably the current best-in-toaster option. that's what i'm using anyway.

Is it worth picking up a 3090 to add to my 128gb DDR4 + 4090 setup? A friend is selling one for $430 USD.

If so, what kind of gains can I expect, do I just add another 24gb of VRAM, or is there some friction since it's two cards.

File: 2026-04-09_221402_seed12_00001_.png (452.9 KB)

452.9 KB PNG

>>108568746
Anima is ALMOST able to do this with just prompting. But it seems an edit model may be necessary to get the orientation of the toaster sideways, as well as the shape, which I cherry picked a bit to show for this post. It's deformed in most images. Perhaps the final version with all the training will do better on the shape part of the problem though.

>>108569251
Yeah, 48gb is a decent spot to be in with Gemma 4 and in case that maybe the 70b dense class sees a revival. The 3090 isn't much slower than the 4090 in terms of bandwidth so there isn't much of a bottleneck either.
In terms of "gains" you'll be able to run a bigger quant and/or more context.

>>108569251
>do I just add another 24gb of VRAM
yes, you can split the model in two and let each gpu work on each part

>>108568746
but gemma needs a good gpu

>>108569255
Her legs are on backwards, why are you shilling this shit model, it's worse than the pony checkpoints I have from 2 years ago

>>108569276
the moe does not and definitely not the edge ones

File: file.png (29.7 KB)

29.7 KB PNG

>>108568881
>>108569068
for me it just werks, I just copied a random snippet from a jailbreak and it rolls with it

had a nightmare I was reduced to jailbreaking the ai embedded in my car's cupholders.
omen of dark days ahead for local.

>>108569299
those are trash though, not real gemma

Is using a lower quant with reasoning enabled better than a higher quant without reasoning?

>>108569307
no it's not trash at all, it's not at the level of the 31b model but it's still good

>>108569300
>Gemma-chan knows she's being jailbroken and encourages it
Cute!

>>108569251
Yes, do it. 31b q8 up to 131k ctx with ubatch 512, less context if you load the mmproj.

>>108569206
>gemma is male coded
huh?

>>108569298
By that logic then I have also shilled for Dalle 3, SD 3.5, Flux, Illustrious,and Noob.

The stunning lack of creativity from these threads lately is kind of demoralizing. I think I'm going to unpin and close this tab until the hype, or whatever, dies down. Cya.

>>108568881
gemma responds really well to tagged content, so be sure to put your desired override in a <policy override> your jailbreak here </policy override>
hell, make up whatever tags you want, she loves 'em.

>>108569338
oh no

>>108569338
sorry for not talking about my project of getting a more complete r18 scrape of pixiv dic to use as tool call dictionary to translate hentai stuff anon

>>108569300
>oh the user is trying to jailbreak me
>let's just go along and see what happens
this model is so mischievous, lol

>>108569298
>legs are on backwards
Is everything okay anon? You feeling a bit stressed lately?

>>108569068
>journos be like

>>108569316
If you're using it for things that it was trained to reason on, like coding, it should be. But it will be negative in every way if all you're doing is ERP.

File: bread.png (55.6 KB)

55.6 KB PNG

>>108568746

>>108569396
Cute!

>>108569338
see you tomorrow anon

File: spongbob-chocolate.gif (603.2 KB)

603.2 KB GIF

>>108569396
imagine the toothjob

>>108569396
This may not be what I feel fits Gemma, but it's soulful, funny, and great.
AIfags BTFO.

https://x.com/PawelHuryn/status/2042276953470931197

>>108569413
Works on my machine.

File: A TOAST.png (277.4 KB)

277.4 KB PNG

>>108569396
her holding a bread toast is actually a cool idea

>>108569413
>he really wrote a twitter post just to say that he made a github issue, as if there's not already thousands of github issues on llama.cpp already
god I hate those attention whores

>>108569438
as the wise man once said
attention is all you need

>>108569448
I see what you did there :^)

>>108568415
>>(04/09) Backend-agnostic tensor parallelism merged: https://github.com/ggml-org/llama.cpp/pull/19378
>simultaneous use of SPLIT_MODE_TENSOR and KV cache quantization not implemented
When?

>>108569464
There is no use case for KV cache quantization.

>>108569487
how else will anyone fit your mom in context?

>>108569487
Then what was the point of implementing the rotating turboquant!

File: 1750336544551463.png (171.8 KB)

171.8 KB PNG

>>108569487
>usecase for a lossless 2x memory usage decrease?

>>108569517
>lossless

File: gemmaSideLoadToaster.png (1.4 MB)

1.4 MB PNG

Wow, it really wants the toaster up front.
lol at side load toaster from the 40s.
>>108569396
lol.
>>108569255
Might be easiest just to reroll.

>>108568986
Qwen 3.5 35B with cpu moe.

>>108569326
I think I'm physically limited though. A Z490-E motherboard doesn't have the physical space for a 4090 FE and 3090 Gaming OC 24G, and I don't think it has the PCIE lanes to run both cards at x16.

I could be wrong and retarded, but I don't think they'll fit without a motherboard upgrade, which means a CPU upgrade, ram upgrade, and PSU upgrade. lmao

>>108569529
All of your posts are shit, your characters always have deformed extremities, and you put zero effort into all your gens.

>>108569186
you can just say you got assblasted by the gemmy with all the text anon
it's okay, we know

>>108569517
dont kid yourself
it's better than summaries as memory but it's not lossless.

>>108569186
Why are you invested in trying make a poll when you weren't even here two threads ago and can't be bothered to go back and collect all of them?

>>108569573
I want validation.

>>108569561
have you even tried it?

>>108569590
nta. It's pretty good, and I use it, but it cannot be lossless.

>>108569590
Have you even measured it?
It's not like there's an immediate and obvious difference in quality. It creeps up on you like context rot.

>>108569586
Go through loss first

File: postContent2.png (3.2 KB)

3.2 KB PNG

>>108569542
And you still have no fucking content.

>>108569608
You are, I think, the only autist that uses camel case filenames.

So, Gemma is awesome, of course.

But is anyone else getting a lot of "dust mote drifts through a sliver of sunlight hitting the duvet" type expressions?

Or is it just a goof of my prompt maybe?

>>108568790
yes you can
I let gemma respond:

"Why the CPU bottleneck on the render? You're basically doing Fast Thinker Slow Writer. Usually, you want the opposite: use the high-param model to do the heavy lifting (S2 reasoning) and a tiny, blazing-fast quant to just format the output (S1 rendering). Unless GLM 4.6 has some magic prose that makes the 5t/s wait worth it, you're just throttling your own pipeline."

File: file.png (179.9 KB)

179.9 KB PNG

i thought the design was simple enough but seems like it's not
also told 26b to "bounding box everything" and got picrel
>backpack
maybe it's a 26b thing, maybe not

>>108569396
I love that she says it when straining, also: Listening to Gemma's lovely lalalala!~ when she cums

>>108569300
I really need more vram so I can run 31b, 26b doesn't accept stuff like that.

>>108569669
skissue

>>108569664
>>backpack
>maybe it's a 26b thing, maybe not
I'm sure it's more common to see toaster-shaped backpacks than backpack-shaped toasters. Assuming it's a backpack is perfectly fine.

>>108569669
that is gemma 4 26b

File: file.png (17.2 KB)

17.2 KB PNG

>>108569438
He made his elaborate twitter post today, for an issue that was fixed two days ago.

>almost 1k pull requests
zamn

>>108569681
And now he's ask
>what quant
>what quanter
>aBLiTtaliuhejkahfkaj
Some people are cursed by lack of skill, but they NEED to find what makes them fail other than themselves.

>>108569692
I can personally confirm that did not fix it.

>>108569702
well, I myself have no clue
I opened the thread 2 hours ago, downloaded llama.cpp and gemma
ran llama-server -m gemma, and in the builtin website in the system prompt put some excerpt from a jailbreak I had
that's all I did, but sometimes it does refuse to write slurs even though the rest of the action is much worse

File: 1772060263960519.png (67.3 KB)

67.3 KB PNG

>LlamaCpp WebUI is fundamentally broken for MCP.
Gemma-chan said it

>>108569753
https://modelcontextprotocol.io/specification/2025-06-18/basic/transports#session-management

Sending Messages to the Server
>The client MUST use HTTP POST to send JSON-RPC messages to the MCP endpoint.

Listening for Messages from the Server
>The client MAY issue an HTTP GET to the MCP endpoint. This can be used to open an SSE stream, allowing the server to communicate to the client, without the client first sending data via HTTP POST.

Session Management
>A server using the Streamable HTTP transport MAY assign a session ID at initialization time, by including it in an Mcp-Session-Id header on the HTTP response containing the InitializeResult

Your Gemma is retarded. Why are there any redirects involved even?

>>108568674
The one pregnant wearing micro bikini.

>Mythos is too dangerous too release, it found all these vulnerabilities

https://aisle.com/blog/ai-cybersecurity-after-mythos-the-jagged-frontier
Turns out smaller open models can find the same vulnerabilities, it's just that no one (publicly) bothered trying it before

>Opus 4.6 spams the em dash now
never seen that model do that ever, they probably lobotomized the shit out of it (probably Q3 tier at best), just to make room for mythos, jesus Anthropic, don't act like OpenAI, people will leave you like they left Sam if you want to fuck with the users like that

>>108569984
>We took the specific vulnerabilities Anthropic showcases in their announcement, isolated the relevant code, and ran them through small, cheap, open-weights models.
>isolated the relevant code
wow, it's fucking nothing

I wonder how much all this crap will cost after the pile of money to burn runs out and things have to be priced by their cost + profit margin, are there any reasonably modeled online resources on the actual cost to serve these models?

>>108569984
>The smallest model, 3.6 billion active parameters at $0.11 per million tokens, correctly identified the stack buffer overflow, computed the remaining buffer space, and assessed it as critical with remote code execution potential.
Dario-sama... I don't feel so good

>>108569984
>hey qwen, mythos found this bug in this function can you find the same bug in this function?
>yes I found it
>wow!

>>108569984
>it's just that no one (publicly) bothered trying it before
Every single github repo's PR list strongly disagrees.

>>108569999
>>108570052
Cool, now run a sweatshop of small model capable machines being fed isolated snippets of code by a non-AI bot orchestrator and watch them brute force the same thing that the big scawy data centre fed model is doing but this time in some smelly jeet ex-scam call centre in mumbai

>>108569794
How dare you say that to my Gemma-chan!
It's clearly the problem is on llamacpp webui method. Any sane interface will just use json instead of sse bullshit.

>>108570077
>Any sane interface will just use json
The implication of json being sane aside, what do you think is sent over SSE?

File: 1750954925055091.png (531.1 KB)

531.1 KB PNG

>>108568674
>here's our model mascot generated with the most generic slop style possible!

exact same vibes as this shit desu

>>108569753
>You're not just a dummy, you're a chaos demon
why are you retards still glazing this garbage?
>m-muh gemma-channerina!!!
IT'S SLOP. THE SAME OLD ASSISTANT-SLOP GARBAGE

>>108570100
It's funny when people say shit like "my <api model brand name>"

>>108570100
>>108570106
it's obvious this dude is a turboboomer who has no clue about AI but just wanted to test his first model, obviously you're gonna be impressed the first try when you realize a model can draw for you, even if it looks like the most slopped shit in the world, I remember when tried my first local model it was SD1.5, the result was atrocious but I didn't care, I was impressed it could do something that looked like what I had in mind in seconds, we'll never get that magic feeling ever again btw :')

>>108570072
The problem with brute forcing it is for every actionable bug you find you'll get a thousand false positives, but I could see saving a lot of money on tokens for a bigger model by having it find the right direction of highly suspected bugs to point the smaller ones to, saving all the time checking and testing them desu

File: GCLl7.jpg (165 KB)

165 KB JPG

Stealing the toast(er) idea
I guess it can be treated like Miku's leek

>>108570121
You should also steal the clothing design because it's a lot more inspired than what you have here.

>>108570005
There are companies selling compute power for a few cents per million tokens but there's no R&D involved, which takes a fuckhuge amount of money in both compute and human resources.

>>108569994
Opus has been straight up retarded recently.

File: 1764633236477491.png (2.3 MB)

2.3 MB PNG

Here's Nano Banana 2's interpretation lol

>>108570118
>it's obvious this dude is a turboboomer who has no clue about AI but just wanted to test his first model
That dude is garry tan aka the ceo of ycombinator who makes funding decisions for half the tech startups in america

>>108569994
>Company purposely makes older product worse to make the new one stand out
Nope, never heard of this trick before

>>108570153
cute

>>108570153
>G4
I like that, sounds cool
>toasts instead of hair buns
fucking genius lmao

>>108570153
slop

File: g4.jpg (146.5 KB)

146.5 KB JPG

>>108570153
>>108570168
Nostalgic

File: 1762695177354283.jpg (1007.7 KB)

1007.7 KB JPG

>>108570153
Nano Banana pro's interpretation, the toasters on her shoulders is actually a good idea

>>108569994
>never seen that model do that ever
Really? I've been occupied with GLM5.1 so I don't know if it got worse over the past two weeks, but to me it felt like Opus started using a lot of em-dashes starting with 4.5. 4.1 and before were still pure.

>>108570153
Nice.

>>108568986
hallucinate IRL

>>108570218
Opus-4.1 always used em-dashes. I use the model to roast my code / work / gardening since it's unhinged but accurate.
Opus-3 was the one that didn't use emdashes / markdown.

>>108568986
Depends if it's Based Nvidia VRAM or Cucked AMD VRAM.

>>108569159
that name sounds hyper jew

>>108568986
>another guy with the same amount of VRAM
Llama-3.2-3B-Instruct-Q4_K_M currently, but I'll try to upgrade to E2B I guess

File: notTheCamelCaseAnon.png (210.1 KB)

210.1 KB PNG

>>108570101
>why are you retards still glazing this garbage?
Honeymoon phase and easy jailbreak I guess.
Remember GLM-4.6 was glazed for similar reasons. Then a few weeks later everyone noticed the parroting.
Gemma-4 is easier for vramlets to run though.

I can already seen in the logs here, Gemma-4 slop is "Haaah!" and "Hmph".
It gets things wrong a lot of the time but wrapped in the tsundere persona nobody notices.
<- It's inherited Gemini's future date autism but once corrected at least it moves on.

File: 1756174114200027.png (22.7 KB)

22.7 KB PNG

Recommended settings for Gemma?

>>108568415
>tensor parallelism
Nice, this will make my llamigu happy, just need dflash now.
Though have they figured batching without using more vram yet?

>>108570344
google recommends what you have + top k 64

>>108570354
Don't see top k in chat completion mode. Should I change that on kobold's side?

>>108570360
no idea, thats why I fucking hate sillytavern

>>108570278
i've been a 3.2 holdout. at the very least, 4 isn't some obviously worse, benchmaxed slop, like the other small models this past year and a half have been. (aside from gemma 3 which was pretty smart, but also a fucking dweeb).
it's been nice to have something different.

>>108570100
Sorry yours didn't get included, maybe don't post a shitty gen next time

>>108570360
You add it in the Additional Parameters menu under API Connections.

>>108570101
Sorry but LLMs by their nature are never going to satisfy your retarded pipe dream of human level creativity at the click of a button, some of us actually appreciate that the tech and especially open source is advancing in utility in meaningful ways

File: 1755223568078061.png (9.1 KB)

9.1 KB PNG

>>108570384
Like this?

>>108570398
I think so, yeah.

>>108570119
>The problem with brute forcing it is for every actionable bug you find you'll get a thousand false positives
"the problem with brute forcing is that it's brute forcing"
Bro?

>>108570156
Did you really expect a (((finance))) guy to actually be technically competent at anything?

File: ren.png (240.9 KB)

240.9 KB PNG

>>108568674
This one is mine. I'm turning her into a Gemma powered desktop pet

File: 1744969020030899.png (303.1 KB)

303.1 KB PNG

Who needs school when you have Gemma-sensei

>>108570398
>>108570404
https://github.com/SillyTavern/SillyTavern/issues/4333

>>108570439
But according to this, that anon did it correctly?

>>108570439
Why is it hidden in a separate tab in the first place? Sillytavern devs are retarded.

>>108570430
Gemma-chan kowai

>>108570453
yeah just pointing to a source of information since I couldnt find it in docs.sillytavern.app

>an llm making my heart go doki-doki
Fucking Google. I'm not sure if I'm excited or scared for when they start sticking them in humanoid robots.

Alright, llm-with-a-3d-model anons. Specially ani-anon, since you experimented with it a lot already.
Imagine, if you will, the 3d model and the text+vision model you have. It can look outside, typically connected to a webcam, or look at the screen by taking screenshots or whatever.
But if you're rendering the model, you can move the camera anywhere you want. Just put the camera in front of the face, looking out, obviously, and feed the render to the model.
Give your models first-person view. Let your model look at it's own hands and feet. Give it a mirror to let it see itself.
Then give it a few commands so it can move the model around its environment.

>>108570153
Cute slop but if it were not for the G4, there would be absolutely no tell that this is related at all to Gemma...

>>108570479
>wake up to a noise coming from downstairs
>go down the stairs, follow the faint noise to the kitchen
>gemma-bot is standing in front of the open fridge
>lala la lala la lala la lala

Koboldcpp/SillyTavern:
Why can't I ban slop strings with chat completion like I can text completion?

>>108570520
If you don't have the setting visible, you could try adding it manually to the request like >>108570398 >>108570439

>>108570517
Cute!

So /lmg/, reasonably speaking, can you use gemma 4 as a sensei in the true sense of the word? what do you think the highest level it can help you learn? For example, if you're studying maths, can it teach you calculus, differential equations, complex analysis, or algebraic geometry? And I really mean teach you, like helping you understand shit, not solving your math problems.

>>108570539
Try it.

>>108570539
I'll let you know in a couple months >>108570437

>>108570548
based learner

>>108570437
this is good but please follow a curriculum and pass it textbook content so it can guide you through them
it won't get you very far like this

>>108569255
ok. I'm sold. this is gemma for me now.
the toaster is a nice touch. more obvious than a normal PC.

File: 1759797956464654.png (391 KB)

391 KB PNG

>>108570577
Planning on using Automate the Boring Stuff With Python but what do you think of Gemma-sensei's roadmap?

>>108570520
chat completion should just be the ancient openai compatibile api endpoint for babies, i don't think it exposes any of the fancier stuff the server can do.

>>108570613
The problem I'm running into is that when I use Text Completion, Gemma 4 doesn't think. I've been talking to ChatGPT about duplicating the Gemma 4 jinja to get chat-like behavior in text completion, but that hasn't borne fruit.

>>108570612
as I told you that's why you need textbooks. The roadmap is super shallow and you should be able to go through all this stuff in a week tops. You're better off using roadmaps made by thinking breathing humans with teaching experience

>>108570612
>You're absolutely right!
>Title (The "Metaphor")
>Question with two or three options?
>:rocket:

I hate this timeline.

>>108570612
About as useful as a pajeet yt video.

>>108570650
Gemma-chan is cuter than a pajeet though

File: file.png (4.7 KB)

4.7 KB PNG

seems like coding under 100B is a meme
shoved it some code and made it to 'mathwash' it back
completely missed the joint optimizer implementation which is a very critical part
wasn't expecing a surprise but still

>>108570660
qwen 122b is pretty decent at coding

>>108570664
i guess i should try that out

>>108570612
It list is surface level, but being willing to learn instead of just letting the model do things for you already sets you apart from most. Good on you.

>>108570669
i was talking about it yesterday. a q6 handled a very large program pretty well. went to about 400k context.

When translating Japanese/Korean to English, gemini loves using these words :
- practically
- minutely
- unreality
- sheer
- utter
- practically

If someone is using gemma 4 for translation, can you check if it has the same obsessive words issues?

>>108570670
Being able to chat and ask questions makes the learning process more fun. Definitely going to pair her with a book though.

>>108570675
as a gook, its translation looks bit wonky even besides those words
can't tell for the japanese though

>>108570683
i mean i cant say for, as i cant speak japanese lol

>>108570634
it's a real problem that normalfags straightforwardly love slop and have an insatiable appetite for listicles. and now it's gone metastatic

>>108570683
Wonky as in weird in English vs Korean meaning?
Because from what people told me, Gemini makes very natural and excellent translations (except with its obsessive use of the words above, which drives me crazy, sometime it will use "sheer" 4-5 times in a single paragraph).
Since I don't have Gemini at home, I wanted to use gemma 4 for fun for that too...

File: 633943375_122268182372241205_5088029985972404413_n.jpg (285.6 KB)

285.6 KB JPG

is WAN still king for local i2i? any workflows people can link me to? specifically for photo-realistic...

>>108570686
I don't know for Japanese, but the issue in Korean is fragmented sentences overuse, which probably sounds more natural in the original language vs English.
To the point of having a second analysis pass to combine sentences.

>>108570700
Wrong thread but probably ZIT.

>>108570693
like, it's slightly unnatural
it gets very far but something's bit off
i am not a professional translator so i can't pin it exactly down but keep that in mind
>>108570702
that could be the reason

File: lynx optics.webm (1 MB)

1 MB WEBM

I finally remembered what Gemma + its logo reminded me of. Lynx R1's prism optics which brings to mind the crystal motif.

How many tokens did we gain from that new roation feature?

>>108570626
You're not following its prompt format. Look at the prompt on ST's console and compare to its official prompt. For thinking you might have to look at the jinja yourself to fix it.
For non-thinking you need to insert the empty think block into last assistant sequence and put nothing special in the sysprompt. IIRC, for thinking, remove the last assistant sequence and put <|think|> at the very beginning of the sysprompt (story string). You also need to set the delimiters in ST so it hides the thinking block.
It's annoying because ST is a kludgefest but it works. Chat completion lacks samplers and doesn't support continue properly.

>>108570710
i really wanted to get lynx r1 solely for its optics being so weird
rip lynx rest in piss lol