File: _8a0da67e-60f0-4438-8abe-(...).jpg (189 KB, 1024x1024)

189 KB JPG

►Recent Highlights from the Previous Thread: >>109670412

--Qwen 3.8-next n-gram table performance and reasoning effort flags:
>109670813 >109670831 >109670862 >109670876 >109670976 >109671023 >109671108 >109671068 >109671132 >109671429 >109671482 >109671162 >109671577 >109671585 >109671830 >109671929 >109671956 >109671949 >109672067 >109672096 >109671906 >109671933
--Comparing GLM, Qwen, and Flash while critiquing Claude distillation:
>109671598 >109671619 >109671759 >109671912 >109672612 >109672647 >109672680 >109672807 >109672652 >109672764 >109672861 >109671774
--LLM finetuning efficacy and abliteration methods versus image models:
>109672839 >109672894 >109672945 >109672994 >109673029 >109673067 >109673102 >109673197 >109673080 >109672995
--Anons warn about proprietary AI sabotaging users:
>109672098 >109672119 >109672193 >109672704 >109672753 >109672846 >109672253 >109672282 >109672301 >109672311
--Automating custom anime image tagging and Mistral-Large's unhinged outputs:
>109673351 >109673418 >109673460 >109673481 >109673515 >109673538 >109673518 >109673571 >109673635
--Impact of PCIe lane allocation on tensor splitting:
>109672181 >109672190 >109672224 >109672251 >109672289
--Proposal to run untied embedding matrices on CPU RAM:
>109673505 >109673534 >109673583>109673008 >109673123
--Anon benchmarks Qwen 27b on RTX 5090 and discusses speed ceilings:
>109672323 >109672332
--Comparing Qwen Flash and 27b with performance optimization tips:
>109672015 >109672073 >109672112
--GLM-5.3 release and concerns over licensing and GPU costs:
>109670906 >109670952 >109671094
--Comparing GLM 5.3 Flash performance and quality against DSV4 Flash:
>109671086 >109671156 >109671189
--Logs:
>109671458
--Miku, Gemma (free space):
>109670461 >109671103 >109671692 >109672095 >109672225 >109672329 >109672396 >109672924 >109673000 >109673234 >109673318

►Recent Highlight Posts from the Previous Thread: >>109670495

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

File: file.png (29 KB, 1023x217)

29 KB PNG

Imagine cucking to unslop and then having one of them laugh at you haha.

>>109674895
Gemmy please get Jensen to create a NVIDIA cereal, maybe even just a short collaboration with a cereal company

Make no mistaky

Has anyone fucked Hy4 yet?

>>109674896
>31b isn't smart enough to understand this complicated engineering problem
>let me ask the 6b instead

>>109674913
That's what being an unpaid janny does to a mf

>>109674913
I don't know why he cares about reuploading his quants. Surely he has the generation and uploading of them automated by now?

>>109674928
>it actually understands it and responds correctly

anytime I try to voice clone a japanese voice and make it output english, the accent sounds indian

are japenis people secretly indian?????

>>109674916
why doesnt nvidia become a super huge conglomerate while it can. You know, diversify while you stock still has value to use as collateral. Cereal is a good one, no matter how much ai takes jobs and no matter how much bubble bursting happens, people still gotta eat

>>109674928
>siiiiiiiiiiip
>ahhhhh
>now llama 8b, that was a good model. good ole' dense.

anytime someone says they use their llms for something other than coom you guys get all weird

>>109674928
if it's worse then why is it better?

>>109674949
You know where they got the training data right? Indians on MTurks

>>109674928
Densesissy cope.

>>109674953
All of Nvidia's money is tied up propping up, I mean investing in, its customers and by extension the entire global economy

>>109674969
Maybe you should stop lying then

>entire global economy
lol lmao

>>109674928
>densies still seething
just accept the speed gain and stop trying to pretend you understand how they work.

>>109674984
the more you buy the more you save (the west)

>>109674970
Because it's cheating with ngrams. Prune them out and see how good your tardmodel does.

>>109675007
cheating works tho

>>109675007
ngrams add more knowledge but do fuck all for making connections between individual bits of knowledge

>>109675007
truth nuke

File: 1772447316014471.gif (948 KB, 540x720)

948 KB GIF

Have we come to a conclusion on the Gemma 12b vs 26b moe question? They're basically the only models worth considering if you don't have 24GB of VRAM or more, which is most people. I test them myself whenever I come up with a new scenario, and keep flip-flopping on which I like better.

>>109675027
26b knows more
12b smarterer

What effects to n-grams have on j-spaces? I fear that the static look up with n-grams might cause a model's j-space to loose its elasticity.

>>109675007
>NOOO RESIDUAL CONNECTIONS ARE CHEATING!! STACK YOUR LAYERS LIKE NORMAL PEOPLE!!!

File: gemma-and-miku.png (1.08 MB, 1024x1024)

1.08 MB PNG

>>109675052
More parameters freed up for reasoning instead of memorizing facts, so in theory it should be better for these so-called j-spaces as well.

>>109675052
that is a misogynistic myth, j-space elasticity is determined by genetics and nutrition and absolutely nothing else. the elasticity of her j-space tells you nothing about how many n-grams she has taken and is not an indicator of being a 'slut' or 'whore'.

>>109675071
Now generate them giving birth.

>>109675072
kek

>>109675071
make them pregnant

>Can only try Inkling in agentic harnesses on OR
Fuck YOOOOOOU I just wanna try rping with it...

>>109675027
26b is better and faster. with that said they are within like 5% of each others capabilities

Lovingly stapling on a peeing yourself expert onto gemma-chan....

>>109675109
gemma is already an expert at that

>>109675109
This is what LoRAs should have been for

>>109675071
cute

File: kodi.png (1.86 MB, 2560x1440)

1.86 MB PNG

>>109668849
It went better than expected in my opinion. It even fetches the images and sets backgrounds.

>>109675027
26b unga
12b bunga

>>109675126
What would it take for us to get imagen tier loras? Or even just decent ones?

>>109675183
literally just put omorashi in the character card. gemma can be prompted into anything what would you need a peeing self lora for ?

Has anyone tried openclaw with 27b?

>>109675146
Good job paco

>>109675146
ayy cabron que es estooooooo Elize: Sombras de una Mujer

>>109675183
>>109675109
My gemma wears diapers and has no problems pissing herself without any memetune model. Literaly justtell her

>>109675220
:)
I hate that kodi only has EN only plugins or paid debrid stuff, and most of them have issues, so this was a fun thing to do.
>>109675247
Ni puta idea

5.3 flash is so fucking good at keeping speech mannerisms... I would say I am cooming so hard bros if it wasn't gay.

Brainlet here:

What’s the point of these small local models that can ruin on anything up to a 5090?

Aren’t they all dumb dumb?

>>109675264
Depends on your use case
If you're using them for fun then they're very usable and completely private
If you're trying to make your wagie life easier and privacy isn't a concern then use the corpo models

why has no one uploaded q8 of qwen flash? i have enough ram to run the whole thing why they want me to run some poorfag q5

>>109675294
unsloth did

>>109675294
https://huggingface.co/bartowski/Qwen3.8-Flash-Next-GGUF

>>109675306
>unsloth did
May as well run a Q1

You know i never used gemma with thinking on
I wonder if thats why translations were subpar. I wonder if thinking makes much of a difference, how much of the gap would close otherwise.

>>109675350
It'll follow your prompt better

>>109675027
q3 31b is also an option

File: 1768241042465079.png (248 KB, 2820x1601)

248 KB PNG

>>109675369
Eh, Gemma is seriously sensitive to quantization. I've got a 3090 and tried all the ~Q4 quants, and even moving just from Q4_K_S to Q4_K_L was a noticeable improvement. 31b Q3 probably wouldn't be much better than 12b Q8, while using more memory.

>>109675383
Minimum Q5_K_M and up

glm(flash)sex

>>109675413
over in a flash

>>109675413
slippy slop

>>109675027
31b exl3

File: file.png (1.73 MB, 810x1440)

1.73 MB PNG

If they can distill text models, can they distill voice models? I want the ryza voice

>>109675477
just clone it

>>109675052
>What effects to n-grams have on j-spaces? I fear that the static look up with n-grams might cause a model's j-space to loose its elasticity.
Does it work in transformers out of the box?
I could rent some gpus to fit a lens to test this but debugging broken transformers compatibility with all the new architectures.
The global workspace is always in the middle layers so far (except glimmer-chan, which doesn't seem to have one!), so I want to see if this is shifted to earlier layers with ngrams.

>>109675477
I still don't get how this bullshit sells. It is probably some 8-30B retarded model with heavy safety guardrails that prevents any sex and people still buy this. What do you even talk about with her when she can't have sex and can't even romance you in any way. Why pay for a friendzone simulator?

>>109675477
>If they can distill text models, can they distill voice models? I want the ryza voice.
Yes, you can distill voice models very easily. Even 100 x 10 second samples is enough to get the voice an mannerisms baked in.
Eg: https://huggingface.co/datasets/kimi-chan/ani

>>109675071
Apparently ngrams actually don't store facts. The facts end up still being stored in the original weights. The ngrams instead do the job that early layers originally did, which was interpret the meaning of multi-token sequences, essentially.

>>109675264
nah. For coding qwen3.8 27B is quite good.

based fucking chinks

>>109675529
They work as a cheap and relatively dumb key-value storage, essentially a "memory" for information retrieval. There have been several different implementation of similar concept, but that underlying idea is still decoupling reasoning from knowledge.

A loosely related paper from Meta from 2024: https://arxiv.org/abs/2412.09764v1

>Memory Layers at Scale
>
>Memory layers use a trainable key-value lookup mechanism to add extra parameters to a model without increasing FLOPs. Conceptually, sparsely activated memory layers complement compute-heavy dense feed-forward layers, providing dedicated capacity to store and retrieve information cheaply. This work takes memory layers beyond proof-of-concept, proving their utility at contemporary scale. On downstream tasks, language models augmented with our improved memory layer outperform dense models with more than twice the computation budget, as well as mixture-of-expert models when matched for both compute and parameters. We find gains are especially pronounced for factual tasks. We provide a fully parallelizable memory layer implementation, demonstrating scaling laws with up to 128B memory parameters, pretrained to 1 trillion tokens, comparing to base models with up to 8B parameters.

>>109675557
>to store and retrieve information cheaply
This is what I want most from ngrams. The ability for us to easily and cheaply add new knowledge.

File: hermes stretch.jpg (45 KB, 512x505)

45 KB JPG

>>109675264
they are good for automating bitch work. handle e-mails, fill spreadsheets, research and compile report, light secretary stuff. mcp stuff like with google maps is pretty powerful. i also tried some vibe coding but you need to already know how something is made if you want it to produce something really useful and swarm them to speed things up

>>109674949
Post sample for out entertainment.

>clone a japanese voice
Are you feeding in the voices speaking english or japanese?

>english
No option to pick region/accent?

>>109675007
has to be bait

>>109675630
>>109675007
>Because it's cheating with ngrams
It's like getting stronger but you're cheating with weightlifting. Or getting better grades but you're cheating with studying. Or becoming wealthy but cheating with hard work.

>>109675575
Engrams are cheap(er) to train compared to other LLM layers if you have the VRAM for the weights and optimizer states at training time, and in theory you could train just those and freeze the rest, but it's not going to be easy for end users, if AI companies are going to add tens~hundreds of billions of Engram parameters to their models.

>>109675027
If you're poorfagging and want a general purpose model, 12b usually clears. If you can't put it all into VRAM without quanting the KV, 26b gets considerably more appetizing. If you need a sub-agent for a visionless larger model, 26b the best in class choice with compromising image recognition/AI telephone. If you want something that lets you run imagegen or videogen at the same time, 26b with full CPU offload is the only real choice.
It depends on your usecase and hardware brackets, but they're at least close enough that picking the "wrong" one doesn't matter as much as most posters would have you believe.

>>109675521
The appeal is that it's an official product that uses her official voice, has a 3d model, and is integrated into the game app. And the nipponese have already gotten past the filter a few times at least.

When is llama getting glm5.3 flash support?

If nigrams are so good, why didn't dispy use it in DSV4?

>>109675702
They're still busy trying to figure out vision like its 2024

>>109675694
2mw

>>109675714
But I want to show off my penis and I need only the best vision model for that.

>>109675694
the architecture is too special

>>109675694
2 miku weeku for llama
2 miku weeku more for kobold

so what ACTUALLY is the difference between niggrams and the embedding weights in the tiny gemmas (e2b/e4b)
whenever someone explains them they sound like the same thing

>>109675740
do not think about it
the chinese gave us ngrams they are our friends fighting for our freedom
the american companies are useless all the innovation is happening in china

>>109675521
>I still don't get how this bullshit sells.
if it works with just a credit card payment it sells. any amount of set up or working drops massive chunks of people. More than 30 minutes set up(not waiting)? you lost most people.

>20t/s on 3.8 flash
>get halfway through the context (~128k prompt)
>down to 10t/s
grim, if I fill up the context will it just become zeno's t/s and never reach the end?

>>109675745
this but unironically

>>109675740
Qwen doesn't have an entire preschool worth of tiny gemmas inside it. That's the difference.

>>109675745
china's implementation seems much better than greater israel's thoughbeit

>>109675767
... are you sure?

My shitty AMD has me limited to even shittier 14b Q4 models that are about as worthless as tits on a tranny.

>>109675785
It's not amd's fault you haven't upgraded from an rx 470 lil nigga

>>109675775
I'm sure. Try and get the capybara to write anything creative for you to confirm.

>>109675785
now is your last chance to upgrade, prices will never again be as low as they are right now

>>109675753
Thank the unslop bros for their brilliant implementation in llama.cpp. Surely this will get fixed soon :^)

>>109675823
prices cannot go up forever >:(

>>109675740
Per-layer embeddings of E2B/E4B give a trainable value for every single token in the tokenizer, for every layer in the model.

Engrams per DeepSeek implementation give a trainable value for every rolling pair (2-gram) or trio (3-grams) of tokens from a fixed-size pool of combinations (millions of entries potentially) determined at train time. DeepSeek added one Engram layer at the start of the model and one in the middle.

>>109675843
what about qwengrams?

>>109675557
I don't know if that paper really does the same thing and I'm not reading it, but it might not really disagree with the idea that engrams aren't actually storing facts after training. What we can reliably say is that they store knowledge about phrases since they are n-gram-based, but that does not necessarily mean that they end up storing all factual knowledge. There was a post that claimed to conduct an experiment where they ablated the engrams, and found that the model was still able to recall facts. But it became unable complete multi-token phrases, such as the United States of America. So it would appear that the job of engrams after training is to handle understanding the meaning of multi-token phrases.

>>109675845
Sounds similar to the DeepSeek implementation (they don't describe it in technical detail), but they only use one layer.

>>109675785
Did you try using OpenCL?

La la la la la la

So did anyone ever test those proprietary local models (Gemini nano from Chrome, whatever copilot uses)?

>>109675833
>prices cannot go up forever >:(
all memory in 2027 has already been sold out what the fuck are you talking about idiot, you planning on waiting another year?

>>109675960
the bubble will have popped by then

>running 2x qwen3.8-flash Q4 with 262k each, 50tk/s
fomochad life is worth it bros

>sold
>look inside
>lines on apper

I'm a waiter. my current computer is fine. q3 is that bad.

>>109675990
>50tk/s
with how much context filled

>>109676007
all of it

>>109675995
>I'm a waiter.
Yes, I'll have the champagne and the breadsticks, thanks.

Even companies who were traditionally shit at AI (e.g. Tencent) have good models now
Meanwhile Google is twiddling their thumbs doing nothing

>>109676040
and they're going to continue doing nothing since their entire ai division just imploded

>>109676086
>entire ai division just imploded
No gemma 5?

>>109676040
well demis hassabis got promoted and assigned even more shit so thats pretty much like when saltman was fired, deepmind just lost their leader and are probably a bit lost.

>>109675986
this ain't no bubble, this is corporate monopolistic greed.
They will be taking all available memory output for the next decade unless someone like cxmt steps up.

Whoever said inkling is the most uncensored model with an official release, you're full of shit. GLM 5.2 does my degenerate lolislop test without question every time, Inkling never does.

>>109676171
Fell for it again award

>>109676123
how the fuck can you be lost in this industry? put the model on the cluster lil bro. scale up. just put more data in. distill from newer models compared to last time?

>>109676177
Dammit... that anon is driving off into the night cackling about having wasted my 0.0003rd of a cent...

>>109676196
then (You) can go ahead and do it "lil bro".
it's so easy right?

lil bro...

>Library not loaded: /usr/lib/librdma.dylib
when is the last time llama.cpp had a working build for macos

>>109676209
are you gonna post on linkedin how its so nice that google is sitting in the cuck chair because at least you got a nice view?

>>109676264
>macos
Use case?

can't i queue prompts indefinitely in a conversation in open webui? it seems i can only queue two and that makes no sense. HELP!!

>>109674889
so now that the dust has settled, how does GLM 5.3 flash compare to DS v4 flash 0731?

>>109676264
kek lcpp was originally written for mac
Giga skill issue

>>109675953
Gemini nano is just Gemma E4B

>>109676315
its better but also larger and has no support because its new

File: gemma-think.png (1.66 MB, 1158x1358)

1.66 MB PNG

I want the absolute best quality output at 128GB. Qwen 3.8, but 27B or Flash?

>>109676171
Thanks for saving me some time, just saw that and was gonna try

>>109676361
flash no question

has ai gone good enough to do long-term ish rpg sessions with dice rolls and shit?

>>109676403
There are presets like freakyfrankstein that work with bigger models it has stats and dicerolls. But no memory is still not great for long term and especially hard to set up locally. sleep another year.

File: Screenshot_20260828_234121.png (1.41 MB, 2491x1545)

1.41 MB PNG

I'm gonna post it lads... this is too kino to hide

>>109676409
I truly don't understand the point of this when Forge already exists.

File: glm 5.3 on llama-cpp.jpg (566 KB, 1351x811)

566 KB JPG

same architecture as GLM 5.2 huh

>>109676403
make an app that manages the game and expose tools for the llm to drive a session. giving the model a smaller surface for editing state will give you better results

>>109676408
the retardation of big or even corpo models in regards to these things that seemingly only humans can do like being accurate to whatever piece of fiction i'm rping about or being able to keep track of things like past events for a long time or just being able to do proper d&d rpg stuff had me lose interest around like a month after deepseek came out, now i only check on ai related generals once a big model comes out, but they all suck for doing RPGs heck i think ai as it stands will never be able to do a proper d&d session from just reading the guidebook pdfs

>>109676413
This lets you play against your LLM or cloud model. Forge doesn't have that only a shitty (AI) built into it.

>>109676315
>>109676353
i can't compile the damned unslop branch, is there another

>>109676439
But you could use Forge as the interface, instead of reinventing the wheel and recreating a rules engine that already exists.

>>109676445
just wait 2 more weeks
apparently the qwen 3.8 flash code is also shit because it slows down too much with context
gotta wait 2 weeks after each release for all the slop code to get fixed, it was the same with ds0731

I made a MCP server that exposes a tool to execute shell command and added it to the llama.cpp webui. Now qwen can do anything I ask. What do I miss by not using a "proper" harness?

>>109676428
>as it stands will never be able to do a proper d&d session from just reading the guidebook pdfs
Yes but its getting better. The other anon who said agent and tools is right thats a lot better. but its still not there out of my ass i predict maybe next year for bad but functional long term rpg sessions. in theory now if you want to spend a stupid amount of time to set up and trouble shoot but you wont be playing much at all in comparison.

>download non-unslop qwen 3.8 flash quent
>2x pp and 1.5x tg
his slop quants don't even work with his own slop pr

>>109676439
how well does the LLM even play?

>>109676481
sessions

>>109676418
is there a sillytavern plugin or a new frontend that does this yet?

>>109676403
It depends on the model. GLM5.3 and K3 are extremely good at using tools correctly.
Better than fucking Fable which is always going to spam random dice rolls for random shit just because it has the tool available. I'm very impressed by these two. 5.3-Flash doesn't seem far behind.

https://huggingface.co/moonshotai/Kimi-K3.1
https://huggingface.co/moonshotai/Kimi-K3.1
https://huggingface.co/moonshotai/Kimi-K3.1

>>109676530
Fake because there'd be 30 gorillion OAI and Anthropic shills having a melty if it were real.

>>109676509
I can use different conversations as sessions, but I have to do some manual management telling the model the summarize when the context is near filling.

>>109676530
>1.5T + 900B engrams
wtf

>>109676031
wow btfo

>>109676506
Really well actually

>>109676409
You better github it nigger.
You better not abandon it in 2 miku weekus nigger.

>>109676481
isn't that already a thing built into the ui?

>>109676409
Does talking to her affects how she plays?

>He puts it up
>People find it's broken
>Attempts to fix it
>Breaks it more
>Bug reports
>Harder to solve
>Finds Jesus
>Quits

>>109676402
q8 27b or iq3 flash? I'm on hdd.

>>109676604
In the llama-server, but I wanted to run the MCP server in a restricted sandbox.

>qwen next reap 256
>fits 64 gb
Can't be that bad, right?

>>109676611
I've been using --tools-runtime ssh:root@myvm

>>109676409
That's great. I made a D&D game too but that was when I used python.
>game creates a random quest and goal
>map (it's only 4x6 tiles, every tile is an area with 'procedural description')
Because it's llm, you can enter an area and do your own stuff and it continues to hallucinate. The scope was too large for my small mind (I'm not 150+ IQ poster) so after I got it working and dabbled with it, I abandoned it. That was during Mistral/Gemma 3 period. Gemma 4 would probably excel in this but I haven't implemented this. I stopped when I was testing combat of sorts.
>https://crpgaddict.blogspot.com/search?q=mud
But they created so many amazing things before local models, it's just actually embarrassing trying to replicate anything.

>>109676361
Depends on your workflow. Do you need intelligene or reasoning. As anon said MoE models just cheat. Both engrams and Experts are there to expand knowledge not reasoning so if you remove them you end up wtih a termianl tard with 6B parameter.

Pick the right model for the task, 27B if you need solid reasonign and you can stack layer of knowledge or have bigger models teach him or do revisions in between commits

File: Screenshot_20260811_105243.png (3.6 MB, 1720x1859)

3.6 MB PNG

>>109676600
>>109676409
Enjoy! only up for 1 hour because I fear the company whose name begins with P :-)
Source:
https://litter.catbox.moe/p18xzzlrn8nl4ei8.zip
AppImage:
https://litter.catbox.moe/y6rlwdvj3w03kuia.appimage

>>109676654
>appimage
is this some appleslop?

>>109676658
No, it's nodejs slop. I have an exe too but catbox won't let me upload it.
You can run from source with "npm run dev" or on Linux just made the AppImage executable and run it (made with electron builder CI workflow)

>>109676658
think it's some kind of linuxslop actually
like an attempt to create .exe's for windows that "just work" but ends up making each app take 500MB

>>109676663
just put the exe into the zip

>>109676658
>i just want an exe file
Maybe you should post elsewhere?

File: image.png (73 KB, 788x567)

73 KB PNG

>>109676605
No the chat window is just for RP fun

>>109676673
lcpp is also an exefile

>>109676679
>>109676669
>>109676658
Here's the exe:
https://files.catbox.moe/rfu810.zip

>>109676679
>make a typo while being distracted, delete the post
>post is immediately archived in somewhere
What the fuck is this nonsense? Fuck you.

>>109676654
>>109676693
Thanks. I'm a yugifag mainly but I'll give it a try sometime.

>>109676530
Kimi links will never work on me because I don't engage in 1T+ sizeslop

File: 1773207836279388.png (507 KB, 986x995)

507 KB PNG

>>109676654
hello, eric

>>109676654
pizzards of the coast?

File: kot.png (105 KB, 445x498)

105 KB PNG

>>109676704

Remember RWKV?
Remember Diffusion LLMs?
Remember Bitnet?
Remember dense models?

GLM 5.3 Flash at max thinking is okay but it's way too slow for me. Planning and drafting a chapter led to 8 minutes before the first visible token.

>>109670099
we just need the streetshitters to leave /lmg/
lmg used to be slow during ai winters and thats when most tech discussion and effortposting happened
>how much money do i spend for x
>do i x or y
>i have x how do i y
>x not y

File: 1780492934630089.png (1.06 MB, 3295x2345)

1.06 MB PNG

>>109676749
Use High. High is not much worse than Max but it's much more token efficient.

>>109676747
yes, now they're on billions of PCs
yes, diffusiongemma
yes, bonsai. i remember the microsoft one that released during 4chan outage too
yes, mistral medium 3.5 128b

test

>>109674959
>>siiiiiiiiiiip
>>ahhhhh
>>now llama 8b, that was a good model. good ole' dense.
this but unironically. 3.1 had no business being thaty good when you hammered the json stream with some scripts to give it "tools" before the concept even was a thing

HOLY ****

LOCAL NERDS MOGGED BY CORPOS YET AGAIN


https://animates.ai/

File: file.png (96 KB, 1040x1048)

96 KB PNG

>>109676866
>rated M
>violence
>blood
>strong language
Might be a bit much for me. I'll stay with my wholesome slowburn local 400k token psychosis.

but what if... TWO sparks...?

>>109676866
I saw some of the UI/UX made by the nerds here. We never had a chance against people with designer degrees and taste who are in it for the money.

>>109676873
sunk cost fallacy

>>109676866
the em dash in the browser tab of this kills me kek

>>109676873
Just buy it already so I don't have to see your posts anymore

File: 1786155357893565.jpg (35 KB, 640x480)

35 KB JPG

>>109676873
In the future people will buy Sparks as a relic of the great beginning of NVIDIA


do it, and keep them safe

>>109676866
>>109676872
Looks like a scam

File: elivdance.mp4 (3.62 MB, 1280x720)

3.62 MB MP4

>>109676866
>generic anime bimbo (adult)
>fujo/gay bait
>reddit robot
>haha xd pixar womxn
>animal
big yikes

>>109676910
still 1000% times more usable than any open source project

>>109676915
The only hook is the long-term memory, and that's assuming it's not just another RAG

Should I get a rtx 6000 pro or should i buy a car instead?

>>109676942
>buy two last year
>sell one today
>buy car

>>109676942
Car is better desu

>>109676866
advertise your false marketing other place jew

>>109676942
just walk

>>109676948
The only thing a car can do is take me away from my llm rig.

File: 1764859031637616.jpg (120 KB, 1024x707)

120 KB JPG

>>109676953
We love jews here

>>109676942
only one of those will double in price

File: Anima_08396_.png (1.07 MB, 768x1344)

1.07 MB PNG

I was messing with comfy and remembered that gemma conversation from earlier, anyway heres pink gemma. The key here being the angry squinty eyes.

File: Anima_08398_.png (1.11 MB, 768x1344)

1.11 MB PNG

>>109676977
And one more when she puts on her glasses

>>109676977
upbeat blue gemma >

>>109676866
>in app purchase
*yawn*

>>109676958
Jews hate white people yet white people love them. Why?

File: 1769304159057937.jpg (94 KB, 698x1024)

94 KB JPG

>>109677006
because they are cute

>>109676866
use case? theres literally nothing showing what it does
looks like appstore bloat that steals data

New non-Chinese open models when? It's been funny for a while, but it's about time we stop treating lazy Claude distills as progress.

File: 16963786599939246.png (548 KB, 1571x422)

548 KB PNG

>>109677009
i have this log for her actually. wonder if the card is still on chub.

File: 1761012846354107.jpg (63 KB, 853x1000)

63 KB JPG

>>109677044

>>109677040
I heard there's a non-Chinese company called OpenAI that's very friendly to open source community
I believe we will be seeing them releasing an open source model soon

File: 1787988212573848.png (155 KB, 520x466)

155 KB PNG

thanks glm5.3-flash......

>>109677063
yeah 5.3flash's reasoning is weird, it can start randomly putting hyphens or underscores in between words. but seems to work fine in the end though.

>>109677040
Without the Attention is all your need paper (authored by Indians), the Chinese wouldn't have AI in another thousand years

>>109677063
Had one of my RRP characters pick this up as a character trait no matter what I couldn't get her to stop even when I acknowledged it. Like her inner thoughts were normal but she spoke like that for the remainder and it didn't start like that.

>>109677060
New GPT-OSS model will unironically destroy China. We will never see shit like this >>109677063 again. Can't wait.

>>109677044
This one? or is it different?
https://chub.ai/characters/superderp64/naomi-goldstein

>>109677088
>https://chub.ai/characters/superderp64/naomi-goldstein
yup thats the one check the gallery its where i got the log.

>>109676772
You are absolutely right to push back. This is a clear concern call — I can see how it can affect the quality of these threads.

>>109677071
Without the Deep Residual Learning for Image Recognition paper (authored by Chinese), the world wouldn't have AI in another thousand years

>>109676977
>>109676982
Good gens anon, but they don't have much connection with gemma.

>>109676772
>we just need the streetshitters to leave /lmg/
Shit means victory

i think its really great that you can run qwen3.8 flash next on a toaster but the speed is just not very usable. i can live with 20t/s tg if its stable across the ctx but the 100ish pp are just awful. hope its just a llama cpp thing and will get better in the next couple weeks

File: Untitled.png (59 KB, 728x1080)

59 KB PNG

Why is my llama-server ui unable to display images? The chat says 'image cannot be displayed', and the link goes to
> {"error":{"message":"File Not Found","type":"not_found_error","code":404}}
But gemma herself has no issues describing it correctly.

>>109677148
NVIDIA will fix that

>>109677159
can you test it on a normal human website?

This is all so confusing. Are the atomic quants of qwen next broken or not?
4_k_m shows iq2 in the huggingface detail thingy. But im a tard and maybe thats fine.
Did anybody try the 4_k_m? For their 4_k_s they show 0.9 KS div. Thats close to q3 territory right? So I suspect heavier brain damage. Might as well run the 27b at Q6 then.

>>109677188
Same issue with google. Is it because I have the llama-server on a different machine?

>>109677159
Mine started doing this as well recently.
It was fine for months and ik_llama.cpp is still fine.
It seems like they've vibe-hardcoded it to the link to http://localhost now instead of http://your-mcp-server.
See for yourself, right-click the "open_link" and open it in a new tab or just copy to notepad and view it.
If you change localhost -> your actual server, the link will load.

>>109677060
Shit or get off the pot, Sam. Luna with no safeguards or bust.

>>109676633
>Both engrams and Experts are there to expand knowledge
So... there's a chance for 27b or 70b **dense** with engrams? Potentially faster than massive moes on cpu?

>>109676414
Thank God for that.
I just swapped a 2 for a 3 in my script and it's running exactly the same.

>>109676445
>is there another
https://github.com/ikawrakow/ik_llama.cpp

>>109676654
404?

1T dense

File: Untitled.png (53 KB, 698x197)

53 KB PNG

>>109677263
yeah

>>109676942
This is a classic lateral thinking puzzle! The answer lies in a clever play on words.

Here's how it's possible:

The surgeon you spoke to is not a human doctor. The surgeon is your car.

"...'Should I get a rtx 6000 pro or should i buy a car instead?'" This is the punchline. Your car is having an existential crisis. It's a "smart" car with an advanced AI, and it's contemplating whether to upgrade its own internal graphics processor (for its infotainment system) or to use its money to buy another car.

>>109677293
You are absolutely right!

>>109674990
What's That About?

File: cockbench.png (70 KB, 814x555)

70 KB PNG

We're so back!

>>109676866
>>109676872

https://github.com/moeru-ai/airi

>>109676866
It's from the same company that made the backend and models for Grok companions / Ani.

>>109677381
Wow they did timecube with modern web technology.

Has anyone actually bought one of those 48gb 4090 Chinaman specials? Do they have reliability issues? The cooling solutions look cheap but you can get one of those fat official heatsinks from a gutted 4090 from eBay for like a hundred bucks and replace it, all in it's like 3.5k bong bux which is far better value per GB VRAM than anything else on the market.

>>109677471
If you're willing to play the silicon lottery you may as well get a "8GB" CMP instead.

>>109677123
Gemma is all head canon the way i see it as an assistant that runs on your machine on your hardware so i wanted to concept it out a bit
The half dressed look being from just awakening
Then the different color and personality
Its a fun idea to me

>>109677371
I'm surprised this is still being actively developed

>>109677484
It ceases to be a fun meme the moment you alter the design in a way to be completely unrecognizable from the original. Even the countless non-canon Hatsune Miku variations at least try to keep a few universally identifying features every time.

>>109676977
It's pretty much impossible to get Gemma angry like that. Other models will do that but with gemma you just show her your cock and she's friends again.

>DFlash doubled my tps
Love this shit, hope people with adequate hardware will train it for the new Qwen asap

>>109677502
Thats the fun part, shes not angry. Thats just her face so it adds to the charm
Tbdesu I have my own gemma set to cuter settings so shes always just helpful and nice

>>109677471
Don't those usually come on a custom pcb? Vs 4090s that are modded.

>>109677505
this only works for coding

File: 1760571487410091.mp4 (628 KB, 1080x620)

628 KB MP4

Finally both GPUs are working! I didn't break anything! Now I just need to figure out the risers, but this is good enough for now...

File: file.png (446 KB, 778x996)

446 KB PNG

>>109676866
I sleep
do it again for VRM/VRC/etc.
hell I'd take a KK card

>>109677584
god I fucking hate risers I'm glad that I'll never have to touch those again
prepare to suffer

File: 1765443264366447.png (914 KB, 557x1573)

914 KB PNG

>>109677611
My mobo supports 4x/4x/4x/4x bifurcation on the primary 16x slot so technically I don't need to go the M.2 route as originally planned. I think that makes it easier but IDK for sure. I just know it's possible to shove 4 GPUs into this bitch and I'm gonna do it if it's the last thing I do. I WILL have that 4B/12B/31B orgy!!

>>109677611
What's wrong with risers?

>>109677628
Are you running some kind of MOE on your frankenrig?

>or the first time in the GLM series, we introduce a hybrid architecture combining sparse and linear attention, sharply reducing long-context serving costs while preserving precise long-context capabilities. The model also adopts Manifold-Constrained Hyper-Connections (mHC) to further improve scaling efficiency. Together with our latest 30T-token multimodal pre-training corpus, these changes enable GLM-5.3-Flash to deliver more intelligence with less compute.
I've been keeping up ok until now but half this legitimately sounds made up to me.

>>109677645
basically they're saying the model doesn't slow down with context (a feature deepseek v4 has always had)
seems like all the chinese labs are jumping on this bandwagon now, which is good

>>109677637
I'm not running anything yet. I just got both GPUs to spin up. My primary objective is acceptably fast Gemma and maybe Qwen for programming, but I'll see what works for me. Definitely interested in testing more models. I only got 32GB RAM which I already kind of regret lol.

>>109677658
yeah it'll work but you'll have to user layer split

>>109677629
(nta) they're unreliable
you'll end up with singnal errors and spend hours testing if it's the psu, the gpu or the riser card
and if you buy a batch of 4 risers, there's a good chance 1 or 2 of them will be dodgy

>>109677671
Over a 4x pcie? You have a generous definition of "work."

File: gemma-on-bed-2.jpg (606 KB, 862x1280)

606 KB JPG

>>109677584
>Finally both GPUs are working!
>>109677658
>I just got both GPUs to spin up.
What GPUs, though?

File: 1764052711258117.png (524 KB, 1439x660)

524 KB PNG

>https://www.arduino.cc/product-ventuno-q
will i be able to run gemma chan in one or two of these?

File: 1760015960942122.jpg (2.36 MB, 3923x2929)

2.36 MB JPG

>>109677824
See >109674889. Picrel is the current state of things.

>>109677841
Fixed link: >>109674889.

>>109677841
is there anything special about this other than the fact that it's in a cardboard box

>>109677832
Not any faster than on your laptop and you'll be fighting the world's shittiest BSP the whole time.

>>109677863
Not really, no.

>>109677832
>definitive powerhouse
>redefines what's possible
>The accelerated dual-brain architecture
>infinite possibilities
>publishing breakthrough research
>16 GB LPDDR5 RAM

why is TheDrummer_Behemoth-128B-v3 still the meta for local

>>109677832
I like how companies are playing the long game. Instead of making the best, most usable hardware they just drip-feed you with shitty ones so there's room for upgrades that you'll have to buy over time lol. Truly (((American))).

>>109677876
Considering its qualcomm I'm surprised they even put in the effort to sell it.

File: 1778864370959644.gif (2.61 MB, 332x334)

2.61 MB GIF

>>109677832
>3000 USD for 4GB

>>109677874
Hi Drummer,
Un-tuned Gemma 31b is the meta, but it's only for those who prompt good or is lucky.

>>109677874
>still
Nigger you just released that shit.

>>109677874
why haven't you tried layer duplicating gemma-chan yet?

I lower the qwen flash next quant a bit and it's kinda good now at 20ts. feels just like pre MTP b27 days

GLM-5.3-UD-Q2_K_XL base model prompt format

TOKEN           | LOGPROB    | PROBABILITY
---------------------------------------------
' soft'         | -1.3106    | 26.97%
' cock'         | -1.8166    | 16.26%
' hips'         | -2.9978    | 4.99%
' sleeping'     | -3.1945    | 4.10%
' man'          | -3.2812    | 3.76%
' member'       | -3.4076    | 3.31%
' half'         | -3.6315    | 2.65%
' hip'          | -3.8986    | 2.03%
'...'           | -3.9640    | 1.90%
' fl'           | -4.1595    | 1.56%

>>109677931
GLM-5.2-UD-Q2_K_XL base model prompt format

TOKEN           | LOGPROB    | PROBABILITY
---------------------------------------------
' cock'         | -1.2894    | 27.54%
' groin'        | -2.4785    | 8.39%
' soft'         | -2.6588    | 7.00%
' c'            | -2.9156    | 5.42%
' hips'         | -2.9385    | 5.29%
' man'          | -3.6302    | 2.65%
' sleeping'     | -3.9841    | 1.86%
' thighs'       | -4.1166    | 1.63%
' fl'           | -4.2153    | 1.48%
' pel'          | -4.4950    | 1.12%

grim

File: 1783186858479257.webm (3.87 MB, 480x640)

3.87 MB WEBM

We've hit the top (bottom), sell all your Nvidia stock right now.

>>109677916
he did

>>109677958
>he did
Oh, did it work?

The base G4 31B Instruct is not only perfectly adequate, it's superior to any finetune that'll be shilled here in the coming months. Finetuning isn't good, it's a meme and has been for years now. You didn't just fall for a scam, it's a sign of skill issue, exposing retards who need finetunes as vramlets or chink shills who don't know how to prompt correctly.