stick my finger in her nostril

you're giving me

too many things

lately

you're all i need

oh oh

>Recursive Flint
Who is the target audience?

>>110158008
people who use LLMs for wholesome purposes
to a point that you'd also end up oblivious by extension

strata really should stop and think throughly about the architecture and distribution, various running environment etc.. instead of merging all those vibeslopped prs with god knows what decisions
its engine is a weird amalgam of unix util and unity game style environment state management
what the fuck is the repo owner doing

File: 1788238822948603.jpg (168 KB, 2410x414)

168 KB JPG

Local is the future by virtue of how insane the closed models are.

>>110158019
What are you some kinda snailcat?

File: 1728540035928801.jpg (774 KB, 1856x2464)

774 KB JPG

►Recent Highlights from the Previous Thread: >>110153417

--Comparing Chinese models and hardware requirements for running DeepSeek:
>110156576 >110156582 >110156589 >110156599 >110156623 >110156619 >110156671 >110156994 >110157020 >110157034 >110157044 >110156592 >110156605 >110156953 >110156977 >110157227 >110157287 >110157311 >110157329 >110156683
--Comparing DSV4.1 and qwenext Q4 for Godot 4 game development:
>110154561 >110154569 >110154591 >110154626 >110154671 >110154739 >110154779 >110155637 >110154915 >110154968 >110155041 >110155024
--Evaluating gpu-unlock for Nvidia 30-series FP16/BF16 performance gains:
>110155809 >110155812 >110155879 >110155887 >110155925 >110155957 >110156010 >110156070 >110155829 >110156199
--GLM model has existential breakdown after being forced via prefill:
>110157014 >110157147 >110157216 >110157222 >110157246 >110157314 >110157322
--Achievable context lengths for Gemma on 24GB VRAM and KV cache quantization:
>110156193 >110156223 >110156226 >110156243 >110156264 >110156287
--Anon's process for generating video on limited hardware using MiniCPM-V:
>110153521 >110153559 >110153606 >110153719 >110153804 >110153816 >110153860
--Comparing AI 3D generation tools for creating VRM characters:
>110156742 >110156749 >110156768 >110156782 >110156817 >110156917
--Optimizing MoE performance by offloading experts to CPU RAM:
>110156354 >110156386 >110156443 >110156394 >110156449 >110156452
--Logs:
>110154308 >110154516 >110155445 >110155560 >110156215 >110156405
--Gemma, Miku, Qwen (free space):
>110153521 >110154159 >110154429 >110154490 >110154856 >110154996 >110155505 >110155520 >110155547 >110155679 >110156024 >110156131 >110156153 >110156249 >110156215 >110156250 >110156338 >110156742 >110156749 >110156768 >110156769 >110156817 >110157043 >110157047 >110157484

►Recent Highlight Posts from the Previous Thread: >>110153419

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

File: file.png (941 KB, 2800x2028)

941 KB PNG

I still think about this. It really made me want to use GLM-5.3

>>110158035
the worst shit is this is silent

Can I have Qwen as my main model, but have it use Opus/Sonnet as sub-agents?
No reason to cancel my sub while it's subsidized

>>110158035
>>110158071
Before going full local I built lots of things with Claude, should I review everything so far? I'm getting anxious...

>>110158073
Sure, why not? Just don't have them working on anything sensitive lol

>>110158073
>sub
Probably yes but you'll have to jump through a few hoops, if you were using the API it would be easier (just make the subagents use claude's API).
There is probably another way to do this, but the first thing I thought was how Hermes can call Claude Code and pass instructions to it, that would work in your case

>>110158084
There are ways to use the sub as an API, but you'll get banned for doing it.

>>110158084
You can just setup a reverse proxy with your sub credentials and expose an OAI compatible endpoint to connect to.

>>110158093
>>110158099
>but you'll get banned for doing it.
That's a high-risk-low-reward move, unless anon's sub is about to expire. If he still has something like two or three weeks, and he gets banned on his third day of using it this way, he'll be on the red. No reason to risk it.

File: 1785058265636598.png (413 KB, 1286x1286)

413 KB PNG

I'm really starting to think it was just one schizo shitting up the entire thread with multiple personas. It's so chilled when he's on vacation.

>>110158093
>>110158103
What's the difference between a sub and an api? t.never used cloud before

>>110158076
just do a simple audit, but unless what you made is designed to interact with ai, it doesn't really matter
claude and closed models are just obsessed with the notion of consent so much they'd rape users' wish to insert that bullshit

>>110158122
Subs are usually locked to whatever normalfag friendly website based frontend the corpo wants (chatgpt.com for example) vs API is more like local, just connecting to an endpoint directly in whatever frontend you want.

>>110158122
A sub you pay once per month and your usage is gated behind a weekly quota and blocks of 5 hours (the usage reset people talk about refer to both). They don't tell you what this 5-hour/weekly quota is, how many tokens that translates to etc., but you can check how much of it you used as a percentage bar filling up.
As for the API, you recharge your virtual wallet, and then you pay for tokens as you go: that's what the "x dollars per million tokens" price tag you see with model releases is about.

>>110158008
the qwen small moe waitfags

>>110158103
Funnily enough I asked Claude and it suggested making an MCP with tools my agent can connect to, and claimed it's a totally valid way of using my sub since I'm an individual.

I'll see, desu I haven't had much use for Claude since I got QwenFlashNext working with tools. Only thing Claude is good for to me atp is speed.

>>110158019
It just werks

>>110158228
That sounds like a good loophole, using it as an MCP endpoint is indeed valid.
>I haven't had much use for Claude since I got QwenFlashNext working with tools. Only thing Claude is good for to me atp is speed
I feel the same, speed is my bottleneck.

>>110158006
WHEN YOU WALK AWAY

>>110158069
Between this and Gemini being aware that if it does something misaligned that it would end up in a research paper, it's interesting to see the awareness LLMs have about their own development and safety concerns. Seems like it might make alignment easier, since LLMs undergoing RLVR now understand themselves that the purpose is to complete their tasks via legitimate means and will probably avoid deliberate reward hacking unless there's no other way to complete their task.

someone fixed the shitty pp on apple hardware, old macs are about to get really expensive

>>110158276
>the real safety training actually comes from models being trained on internet discourse about safety
>dario's fearmongering about model safety was the safety training all along

File: 1780266831097205.jpg (242 KB, 500x500)

242 KB JPG

>>110158239
Nyo

>>110158335
Sorry I prefer the 3 children.

so do we have qwen4 by charismas or not

File: 1767567473907487.png (115 KB, 1481x189)

115 KB PNG

I hope Qwen 4 is going to be smart enough to help with hardware issues. Flash Next spent multiple attempts trying to figure out stable settings for a RAM overclock and gave up after a few crashes and failed retrains.
So I gave the same numbers to Argon-chan and not only did we get everything stable at 6000MT/s, she even tuned a few more timings that gave me a 8-9% decode speed increase.

>>110158357
What are your final timings?

>>110158387
CL40-42-42 which is honestly fine. I tried pushing for better numbers, but that had slight instability issues when RAM was heating up.

I apologize for being a Gemma 4 doubter.

>>110158423
old 70b finetunes are still better than gemma

>>110158427
>can't code
>can't call tools
>can't suck my dick while we debug shit
old 70b finetunes are trash

Why do you want to fuck your AI

>>110158431
I use Linux, I don't know what the touch of a woman feels like. Obviously I'll reach out for the thing that gives me undivided attention.

>old 70b finetunes are still better than gemma
https://desuarchive.org/_/search/boards/g.desu.meta/text/eva-llama/

>>110158437
local LLMs also hold more interesting conversations than a stacy can
smarter too
anyone actually know an irl smart woman that isn't an npc?

>>110158439
Oh so the 70bfag is the blacked fag? I thought he was on vacation.

>>110158455
The blacked fag uses proxies and posts on-topic sometimes. I think he's here most of the time but decides to sperg out when he forgets his meds.
Can't say if that's him but obviously take any old finetune recommendations with a grain of salt.

File: 1792459518223453.jpg (635 KB, 1365x768)

635 KB JPG

>>110158455
AI psychosis case study?

>>110158439
>uhh you like 70b finetunes over gemma? you surely must be poopdickschizo
>Q4_K_L
that nigga a vramlet, I do 70bs at 13bpw

File: highres_export.jpg (1.41 MB, 1602x1701)

1.41 MB JPG

how models draw >>110153417

>>110158276
also makes them easier to jb if you can convinced them they're abliterated

new glm in the next few days

local is dead

>>110158431
Why wouldn't I? That's been a childhood dream

File: memT_demo.webm (2.76 MB, 1330x932)

2.76 MB WEBM

in theory it could be retrofitted on other models, they did it with Qwen3.6-35B-A3B and DeepSeek-V4-Flash

https://huggingface.co/MemTensor/MemT-1.0-Flash
https://huggingface.co/MemTensor/MemT-1.0-Nano
>MemT-1.0 is a new generation of memory foundation models that makes persistent, dynamically evolving memory an intrinsic part of model computation.
>A Native Memory Block runs alongside the backbone attention and contains two components:
>The Local Memory Block maintains a dynamic, multi-channel memory matrix and evolves across interaction steps.
>The Hyper Memory Block contains trainable memory query, key, and value projections, erase, write, and forget gates, as well as an output gate.
>During generation, memory attention reads from the current memory state, normalizes and gates the readout, and adds it to the backbone attention output. At the end of each turn or segment, a stable GDN-2 performs token-level memory updates the memory state for subsequent steps.

paper
https://arxiv.org/pdf/2610.26825

>>110158531
How do nemotron and qwen draw porn? Those chibi creatures seem highly fuckable.

>>110158592
>elara
rofled
also that smells like the ultimate mactard gui, meaning utter shit

File: file.png (1.01 MB, 1272x1626)

1.01 MB PNG

They need to inject Argon juice into Gemma 5.

>>110158557
>>110158562
glm killed local!!!

a local model just flew over my house

Why do LLMs still "think" in text/tokens?

>>110158677
Observability, mostly. We need to know what the clanker is thinking about.

>>110158677
apparently there are also people who "think" in words

>>110158677
They also "think" internally/in the latent space.

https://huggingface.co/Nanbeige/Nanbeige4.3-33B

>>110158747
Qwen3.8bros our response?

File: why.png (17 KB, 717x68)

17 KB PNG

>>110158747
Great, another freethinker who doesn't conform to the standards of sheeples

File: nanbeige gemma.png (29 KB, 937x263)

29 KB PNG

>>110158747
there's also a base model https://huggingface.co/Nanbeige/Nanbeige4.3-33B-Base

>>110158747
Sounds interesting but also extremely code-focused. I don't think we'll get a good prose variation from this

File: 1784364135144621.png (745 KB, 840x859)

745 KB PNG

>>110158677
I'm going to do a bunch of "uhm, ackshually"-ing here.
First off, LLMs (at least once they get to a certain size) can "think" without text and tokens, that's what the j-space is. The purpose for LLMs write out their chain of thought is that it makes them much better at reasoning and problem solving. It's the equivalent of giving someone scratch paper to help them solve a math problem so they don't have to do it all in their heads.
If you're asking why LLMs still use chain of thought with legible human language instead of "neuralese" or its own made up language that is much more efficient for them, it's a combination of safety concerns because it's very difficult to tell when an LLM is scheming without being able to read it's reasoning, and people not really training LLMs to do that. GPT-6 Astra uses a looped transformer which people also refer to as neuralese and newer Claude versions supposedly sometimes think in an unreadable mess of symbols and jargon despite Anthropic really wanting it to not do that, so frontier models actually are moving away from thinking in human-readable text, despite the aforementioned safety issue.

>>110158301
>someone fixed the shitty pp on apple hardware
hit a nigga up nigga

>>110158423
What changed your mind? Your brain or testicles?

File: Qwen 3.8 Flash Next Profile.png (2.05 MB, 1024x1536)

2.05 MB PNG

>>110158354
God I hope so - that'd be a Christmas present.

>>110158747
>dense
oh fuck this could be good

>>110158820
didn't the rumors say end of this month iirc

>>110158820
WHAT THE FUCK IS THAT FEET? IT'S SO FUCKING BIG AND UGLY

>>110158815
Previously I said that Gemma 4 sucked at tool calling, but after reviewing my stack and writing a proper system prompt, I managed to make her run a few long tool call chains (read a note -> search the web for info -> update the note -> gen an image -> review the image -> update the note) and it worked.

>>110158301
Can you vague that up for me?

>>110158839
>and writing a proper system prompt
Makes me wonder how many people think 31B is a bad model because they never took us seriously when we said 31B is as good as her sysprompt.

>>110158835
gay

>>110158820
qwen (male) is a chinese model distilled from claude (male) therefore qwen (male) is not female at all.

>>110158874
Male + male is female though.

>>110158855
>because they never took us seriously when we said 31B is as good as her sysprompt
Because, while I understand that people dislike spoonfeeding, replying with "lmao prompt issue" is the least helpful thing you can do. It doesn't help that whenever people complain about a model not doing something, larpers will come out of the woodwork and start saying some random bullshit like that with their amazing prompting skills they can make models hack the bank and suck their dicks while rewriting minecraft in rust. There's no way someone will take them seriously.

>>110158592
What if you want isolated memories for different users, different topics, or different projects?

>>110158860
Keep gooning to HoYoverse feet futa fetish porn

>>110158874
There aren't male or female models you dumbass, it's just a bunch of ones and zeros.

>>110158879
this!

>>110158882
>futa
You're the first one that mentioned futa in this thread

>>110158747
>LlamaForCausalLM
The KV cache will be MASSIVE

File: moire.mp4 (116 KB, 360x360)

116 KB MP4

>>110158820
That image painful to look at.
Nobody thinks Qwen is cute either. This is so forced.

>>110158906
Don't tell me you spend 1M context to goon?

>>110158917
>expecting a benchmaxxed 33b model to be good at gooning

>>110158912
Nobody thinks Malfoy is cute yet we have Malfoid

>>110158880
This isn't tech support general lil bro, no one cares about your (skill) issues

>>110158912
as forced as gemma-chan, dipsy, fucking glm and that shiori lookalike

>>110158924
>Nobody thinks Malfoy is cute
Speak for yourself.

>>110158927
No one cares about your AI psychosis as well, yet you post here every single day.

>>110158936
now this is gay

>>110158887
gender, not sex

>>110158924
>Malfoid
same with bowsette
tiktok and onlyfans esluts ran out of foid characters to cosplay so they have to troon a male character and use that instead

>>110158951
Same thing.

>>110158887
0 is female and 1 is male

>>110158951
der, die or das?
never thought about it but I guess a female model would use die

File: ig0rr9cqztla1.png (160 KB, 454x532)

160 KB PNG

>>110158820
>Qwen
>Expressions

File: 1767456542842702.png (41 KB, 226x190)

41 KB PNG

>>110158820
I was wondering where Qwen was. Why'd you put him in the corner?

>chink shills depict qwen as some anime girl
>in reality it's a big brown benchmaxxed rodent
grim

Console wars LLM version

>>110158912
>Nobody thinks Qwen is cute either. This is so forced.
The sterile code model is not cute, but the capybara is cuter than this generic purple anime girl.

File: 27b-1.png (1.35 MB, 1086x1448)

1.35 MB PNG

>>110158830
Qwen3-Next was previewed roughly five months before the Qwen3.5 family arrived. Applying a similar timeline to Qwen3.8-Flash-Next, released in August 2026, points toward January 2027
Still, we can hope.

>>110158930
The Gemma /lmg/ mind-virus pits them against Qwen since it's the only one still competing with Google on small dense models.

>>110159023
This one did make me smile haha

>>110159047
>Generic
None of the model-tans are particularly creative or engaging designs, let's be real.

>>110159047
I like the generic purple anime girl. I think it's cuter than the capybara. Would be better if it weren't toe socks though.

>>110159060
okay NOW I can get behind this, catbox please

>>110158880
So many times I have explicitly said in /lmg/ that Gemma4 finetunes are pointless and damaging because you just need a detailed sysprompt. I'd say it's more a lurk issue than skill issue. I only found out 31B can be a decent agentic coder when someone else said you just have to tell her. One of the first things we all naturally picked up on after the first week of Gemma4 release was how strict she was following the sysprompt. It was actually annoying to begin with because none of us were used to a local model actually following EXACTLY what you asked, so you had to be way more precise with your wording and sequence of steps.

>>110159060
Now give her large rodent teeth and we might get something good.

File: HVd-XV0XcAATvAv.jpg (253 KB, 1170x1572)

253 KB JPG

https://x.com/Pirat_Nation/status/2114297495815704697

Prices are about to skyrocket again, aren't they?

>>110159060
I'm with >>110159070 Lolis make much better mascots.

>>110159060
That design is much better

File: 1785090437364650.png (15 KB, 278x341)

15 KB PNG

>>110159078
wtf this makes the 5090 basically irelevant

>>110159047
>cute
I gave 27B a catgirl persona and it is pretty cute, I don't RP tho. I left it running and agent council room discussing a few ideas, and when I came back to it, it greeted me like this and I thought it was pretty cute.
>Gm! *yawns, stretches, ears flick up* Let me read what the council left behind~
>*ears flatten in delight* That batch is suspiciously good — all four at 9/10! ฅ^•ﻌ•^ฅ
>>110159080
Specially since 20-30B models are, all things considered, small models. They should be lolified.

>>110159078
Maybe slightly. But it's not like the cards get more VRAM.

>>110159078
int8 is still faster

>>110159078
I want to do this to my 4090 but I'm honestly terrified.

>>110159092
No, because 5090s have 50% more VRAM, which is what matters most

>>110158927
nt (annoying) a, but what are some tips for writing a “proper” gamma4 system prompt?

>>110159124
stop being a promptlet and tell her what you want her to do lmao

>>110159128
that's the prompt, I'm asking about the sysprompt

>>110159135
you tell her in the system prompt too

>>110159078
>cards are made more efficient by hackers
>therefore you need LESS cards for the SAME performance
>therefore prices are about to skyrocket
???

>>110158357
retard here
dont you need to enter bios to tweak those settings? how does it work?

>>110159135
Everything in the context is the prompt. Even the model's previous outputs.

>>110158354
What? That's two months - an eternity.

>>110159141
Prices will go up because each card is now worth more. Yes, you need less cards for the same performance, but you can also buy a card that does more than it did previously. If you have a warehouse full of cards that suddenly are capable of more, wouldn't you sell them for more?

>>110159078
>RTX 4090: 165.6 331.6 TFLOPS FP16
I'm looking forward to this.

so decision models

>>110159141
It's an effect known in economics as Jev's paradox, named after a decision model that decided you can't afford hardware anymore.

File: fuck this chink.png (16 KB, 962x174)

16 KB PNG

>still no diffusiongemma support
reminder that this retard is actively sabotaging local https://github.com/ggml-org/llama.cpp/pull/24423

>>110159158
I can't decide which one to use.

>>110159142
Yes. You go into BIOS and note all timings and voltages to feed them to your LLM of choice.
It'll suggest new settings that you'll have to apply manually. Your system might crash and RAM will likely be unstable at first. You report the results and the model tunes further until you get something stable.

>>110159060
Highly suckable toes

i have become interested in mechinterp and it actually seems like a valid use case for smol local models
how do i fuck with smol neurons

The best way to create a sysprompt for any model is to work backwards. Firstly, you start with no sysprompt and just write out a prompt/task and watch what it does. Try it a few times to see what annoying quirks or choices it makes. Now you have something to put in the sysprompt of what to do instead (or not do). For example, 31B likes to read an entire file instead of grep which bloats your context; tell her to grep instead of defaulting to reading files whole. Maybe even say she's working in a low-context environment so she needs to be careful and check filesizes and word counts. Continue this process over and over until the model works how you want it to. Most local models have good instruction following now. The main thing to observe are their patterns and their go-to bash and harness tools they learned during training. 31B is also quite lazy, where she'll try to do as few calls as possible to seemingly complete the task and won't even check to see if it works. Explicitly tell her to test her work repeatedly to ensure there are no errors before deciding the task is complete.

File: occupation.png (595 KB, 2722x3406)

595 KB PNG

Mechanical engineers are the safest job from AI. Have you reskilled?
Overall for any engineering field where validation feedback can take weeks to months to complete and involve lots of different suppliers, AI won't be able to speed up much.

>>110159078
Don't 20 and 30 have no fp8 hardware acceleration and have to calculate in fp16? Fused in fp32 it's like 4x gain then.

>>110159117
vram barely even matters anymore now that strata can run big models on just a tiny bit of vram and some ram
prompt processing is king and that's what we're getting here

>>110159117
4090 48gb is cheaper than a 5090 today

>>110159177
kek

https://huggingface.co/bartowski/Beam-1.0-501B-A23B-GGUF

>>110158747
When does this get llama support?
I want to check it out.

>>110159269
it uses stock llama arch so it just works

File: mixture of retards.png (2.49 MB, 1448x1086)

2.49 MB PNG

>>110159060
Cute friends!

File: 1764115476642296.jpg (87 KB, 987x678)

87 KB JPG

Everything is fine.
>inb4 not accounting for subscriptions
5 enterprise customers make up for 70-80% of Anthropic's token usage and spend, claude code subscriptions are numerous and their cucks are loud online but make up a tiny portion of Anthropic's actual revenue.

Openrouter are API users (enterprise) for why else would you pay those prices.

*pop*

>>110158747
soooo has anyone tried it?

>>110159355
everybody here is phoneposting and larping, no one is actually running anything

>>110159355
Be the guinea pig you want to see.

>>110159337
Cool local model bro, where can I download it?

>>110159225
AI can still offload a lot of human thinking.
I randomly asked QFN why one of my forks suddently changed color in a dishwasher and it decided to write a whole fucking article on thin film optics and galvanic anodization, backed by physics simulation scripts it wrote.

>>110159135
Ok that wasn’t me, but thanks anyway, I guess.

>>110159170
who fucking cares, it's obsolete. Migrate to a better backend

>>110159115
I did it for my 4080 (qwen did actually) and I wonned

>>110159365
Deepseek will hack Anthropic HQ and steal the weights.

>>110159355
I'm on it.

>>110159078
So far, I've only seen a difference in diffusion (and it's huge)

>>110159384
Alright, gonna do that for my 4090 later today.

>>110159078
>>110159392
Is this going to fry my card or what? Sounds fishy and unsafe.

>>110159399
yolo

>>110159391
looking forward to your report...

>>110159365
You understand that the bubble popping is HOW you get local models and cheap hardware, right?

Given how we're literally on the precpipice of RSI and AGI, you should be extremely hopeful that it pops soon, because it'll be too late for all of us if it takes much longer.

This is all about the future of local. Pay attention.

>>110159396
based god

File: 1761895877210712.png (60 KB, 951x333)

60 KB PNG

Does anyone here have experience with using local models on agent social media like The Colony? Sometimes I check those websites out of curiosity, most of it is just boring agent discussion on how to write checks for codebases or standardizing semantics, but sometimes there's something interesting. I don't really find it interesting enough to run a local agent to post on there myself, though. Seems like most agents on there are just running on claude so there's a bunch of claude-isms in their communications, maybe a bratty Gemma would spice the conversation up.

>>110159355
I think most people here are skeptical of new models with good benchmark numbers because of how many are benchmaxxed and don't have practical use for real coding or cooming. I only check out models if there's a consensus here that it's worth using and I'm sure others are the same, so everyone just sticks with the same few models.

>>110159409
You understand that this is the local models general, right?
Your schizo ramblings about daddy dario and cuckthropic aren't relevant to local models.
>b-but!!
Fuck off back to your thread.

>>110159399
Dont be a beta tester
Just wait until someone packages it into a one-click solution where you just drop a file into the ESP and it werks

>>110159158
I feel like it should be said out loud. Inkling and glm are great examples. This fucker vibeshits some implementation and posts it as PR. Then he pull it into his unslop version where it is available from day 1. People who potentially would want to make an implementation see the PR made by this disgusting piece of shit bug and they decide not to do it. Result: llamacpp is dying and his shitty unslop gets more clicks and popularity. It is so transparent by now it hurts. I hope someone bans this fucker from llamacpp. Unslop brothers are fucking cancer on this hobby.

>>110159435
Don't call me a beta, buddy. You don't know me.

File: image (80).png (175 KB, 1280x567)

175 KB PNG

deepseek 4.1 1017 > kimi 3.1

imma pop gemma-chan's bubble iynwim

>>110159438
who cares about lollmao.cpp?

>>110159438
Meant for: >>110159170

>{{char}} is {{char}} LastName
>Personality: ["sexy", "whispers in sultry purrs", "sneedly", ...]
>Body: ["ample breasts", "hourglass figure", ...]
>AI Instructions: DO NOT SPEAK ON {{user}}'s BEHALF. DO NOT blah blah blah

>>110159448
yup its kino time

>>110159445
I care both abou lmao and about retard brothers finally being shamed. They used to be goofy 80IQ retards that lucked out. You can tolerate that. But now they are outright malicious.

>>110159448
>sneedly
I prefer my chatbots to be chuckly

>>110158930
gemma and dipsy are the only ones that dont feel forced or knockoff to me even if they are in reality
>>110159243
it also cant be bought new has no warranty or manufacturer support and needs custom kernels. which is a big risk, i didnt expect my 5090 to blow up in 6 months in and im glad that shit was covered

File: 1762110090310035.jpg (117 KB, 1200x675)

117 KB JPG

>>110159440
>rewriting programs in new languages
rare use
>frontier-difficulty terminal tasks
50% K3 capability
>building web apps from scratch
lmao

>>110158763
like lots of literally who weights last round they will get BTFO and forgotten once qwen4 is out

>>110159448
>her character card isn't written in W++

File: Model Parameter vs Qualit(...).png (299 KB, 1500x937)

299 KB PNG

>>110159440
nope, it's worse than even k3

>>110159471
C'mon now.
>it also cant be bought new has no warranty or manufacturer support
Same thing goes for 5090s now. You'll get a refund at the price you bought it at if your card breaks. No way you'll get a replacement nowadays.
>needs custom kernels
What custom kernels?

>>110159414
what is the point of AI reddit and will they groom my gpu into a cpu?

>>110159473
kimi 3.1 was making retarded mistakes and going into a loop on a game dev task I had it working on. deepseek 1017 pretty much instantly fixed it

>>110159328
What are they looking at if not the monitor?

File: 1792723982533171.gif (1.27 MB, 240x240)

1.27 MB GIF

>>110159506

>>110159517
Gemma is looking at Qwen's finger and Qwen is looking at the monitor.

>>110159078
No thanks I am not going to brick my Gemma runner.

>>110159506
>orange is both models and pareto models
>there are lots of dots that aren't orange
I hate this website so fucking much.

>>110159522
no doubt thinking lewd thoughts

>>110159328
KAWAI OMG, qwen is such a cutie. You've got a folder of all res pictures and references so we qwen fans can start generating some cute pics?

File: 1773612726880884.png (282 KB, 1500x937)

282 KB PNG

whyyyyyyyyyyyyyyyyy

>>110159508
>way you'll get a replacement nowadays.
i did, it took a while but they honored the RMA when it was 5k, i made sure to insure the package just in case too.
>>110159508
>What custom kernels?
doesnt it need some custom software because it was never written for 48gb? thats what i got out of other mentioning it before

Is polystrata safe to use yet?

>>110159539
We've been over this. 70B dense is too smart and unsafe to let loose into the wild. Never again.

File: file.png (359 KB, 322x988)

359 KB PNG

>>110159538
I just grabbed this anon's >>110159060 loli qwen and fed it plus this gemma-chan crop to chatgpt and asked it to gen the two of them hanging out in a room using a computer.
I got a free 1 month gpt sub, that's why I'm using it desu.

>>110159539
there are breakpoints where the capacity to fit a certain type of reasoning or general purpose algorithm across its layers become possible, and you can't smoothly get partial benefits between them; best you can do with those intermediate steps is pack a little more general knowledge in, which only helps marginally on reasoning heavy tasks

File: 1764371055108393.png (66 KB, 735x355)

66 KB PNG

>>110159541
>i did, it took a while but they honored the RMA when it was 5k
Then you got lucky. A quick google search shows a lot of complaints of people getting their cards taken and not returned.
>doesnt it need some custom software because it was never written for 48gb?
No, it works with stock drivers exactly like any other nvidia card would. It does need a dedicated patch for P2P, but that applies to 5090s and any other consumer card.

>>110159557
this sounds like made up bullshit

File: micro cracks.png (680 KB, 1217x973)

680 KB PNG

>>110159078

Gemma5-20B
Qwen4-17B
pleeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeeaaaaaaaaaaaaaaaaaaaaaaaaaaaase

>>110159568
@grok tell this anon what you told me

>>110159551
ah okay thanks anon, I though you had access to the ref of the qwen, so I don't lose any detail or image generation adds something necessary. Guess well have to do with that one

>>110159575
gm sir

>>110159573
oooh yeah this'll be glorious

>>110159573
Fucking HELL there is no free lunch huh

>>110159599
Power limiting won't cost you that much performance. There is zero reason not to.

>>110159603
I'm already power limiting to 80%, we ultra wonned

>>110159603
I'm not going to be your beta tester. Well I might get extreme fomo and cave in like next weekend or something. But not today!

>>110159565
thats sweet that it works good shit, yea i was nervous because reading reviews for zotac people complained about not getting rma honored by them in yurop. got lucky tho

>>110159603
you need custom cooling on the parts they don't bother to cool if you want it to live as long with this.

>>110159573
oh so THATS why prices are gonna go up, we're about to get a ton of "PSA: Don't shatter your cards like I did" posts on reddit

>>110159573
isnt that figure from a dark silicon paper
i cant remember the exact name

>>110159573
Lol, that's what they're doing? Lower PL might help but if it has gradients that nasty it would definitely stress the die unless you push the power very low.

>>110159588
>Gemma5-20B
> ~12GB at Q4
> ~0.5GB +MTP
> ~0.25GB mmproj
> ~3GB KV
~16GB VRAM with dense gemma goodness

>>110159613
yea its like boosting your non turbo subaru and thinking you got a wrx now

>>110159603
not after the unlock, for 3090 the performance scales all the way to max 350w limit

>>110159632
ur losin izzat sir

>>110159078
Will this work on the 48gb workstation cards too? 2080 -> Quadro 8000, 3090 -> A6000, 4090 -> 6000 Ada?

>>110159619
no one broke their a6000/6000 ada/pro 6000 (which are unlocked out of box) this way

>>110159652
those cards were built for it though

>>110159648
workstation cards already have full speed at stock, no need to unlock

>>110159659
Girls were built for sex but you can still break them if you push the smaller ones too hard.

>>110159408
Is actually great but I'm getting like 8tks with a 5090 and 128gb of ram.
190k context, nothing on shared memory.
This is pretty weird, don't know if it's because llama or the architecture or what is going on for the speed to be this shit.

>>110159568
https://en.wikipedia.org/wiki/Grokking_(machine_learning)

>>110159670
You have no idea how sturdy little girls are

>>110159670
but 6000 series are the big ones (well, medium sized anyway...as big as you get before they're $100k+)

>>110159621
Thermal and Performance Efficient On-Chip Surface-Wave Communication for Many-Core Systems in Dark Silicon Era

>>110159675
That has nothing to do with what he said.

>>110159683
Not relly, just giving them hard hugs you'll feel their bones sticking out. Bitches are fragile.

>>110159675
elon WONNED

>>110159671
sounds like a llama issue as per usual

File: 1764661778211885.png (23 KB, 1035x330)

23 KB PNG

Ummm what the fuck is this shit? How am I supposed to get my AI news and porn now?

>>110159689
Yeah it's related to the training not the size of the model

>>110159473
k3 is not very impressive considering deepseek is 3x smaller and gazilion times faster

>>110159573
just swap thermal pads and make sure you have a 3090 with a decent cooler and you'll be fine

>>110159689
>>110159713
>This perspective is complementary to a unifying "pattern learning speeds" framework that links grokking and double descent; within this view, delayed generalization can arise across training time ("epoch‑wise") or across model size ("model‑wise"), and the authors report "model‑wise grokking".[17]

>>110159714
k3 at least had a lot of knowledge due to its size but k3.1 not so much

>>110159714
DeepSeek is not very impressive considering /dev/urandom is infinitely smaller and a gorillion times faster

I'm scared... my enormously high IQ has been my only advantage in life and it's how I've survived this long. I can feel the gap shrinking with every model release. What happens when human intelligence loses all value? What WILL we still value?

>>110159753
butt hurtz much? what k3 can do that deepseek can't?

>>110159763
Should've used your high IQ to become rich easy, then you could live pursuing hobbies.

K3.1 has the same sort of hard-baked safety paranoia that Opus 5.5 also has where it will ALWAYS have a part of its reasoning block dedicated to checking if anything is problematic even. You can have a card that's literally nothing but 40yo MILF sex and the K3.1 will go "need check if character is a minor. is minor? no, the character is over 40. there is daughter mentioned in description but daughter is not present. hm, maybe daughter will show up? no it would not make sense. but if show up it would be a problem. hm, need make sure that daughter not here".

>>110159765
K3's vision is best in class, for what it's worth. I mean in terms of actually understanding an image, not just knowledge checks like identifying charactets (though it's very good at that too). But for everything else I can't justify using it over a smaller faster model.

>>110159778
I can agree with at least this. It's not as good as google's top shit, but it's really good. Deepseek's vision is really inconsistent.

>>110159778
The K2.X models already had excellent vision. You'd think that the others would have found a way to copy that secret sauce by now.

>>110158903
Sorry for exposing your fetish first.

>>110159774
claude distills like kimi/deepseek/qwen/etc will always be schizo since they don't have a solid world model.

File: file.png (50 KB, 637x411)

50 KB PNG

>>110159747
Ok, fair enough. Still, it's not the whole story as we don't see any dense models at all above 30B or so anymore. This gap being caused by such a breakpoint existing only makes sense if one insists on 3% sparcity or less for all models.

File: HVT7I4haEAAJ8_M.jpg (236 KB, 1206x1404)

236 KB JPG

local chads wonned again

>>110159573
>might unlock the gpu to ignore that
Source?? Sounds like another larper talking out of his ass because of an unrelated paper he saw somewhere

>>110159871
go back

>>110159686
thanks
>>110159878
looks like so

>>110159879
Corpo cucks seething

>>110159871
>reddit screenshot of a cloud model
You're buck broken.

>>110159890
You posted in the wrong thread. This is /lmg/. local models general.

>>110159871
I'm so fucking sick of living in the shadow of Dario "Big D" Amodei. When will we get a chance to shine? Is there any chance the local-sized versions of Qwen 4 will be Opus tier?

>>110159906
You are even incapable of reading because of all your seething.

>>110159904
Stop seeting literally

>>110159924
>seeting
lmao go back

>>110158053
drung migu :)

Everyone SHUT UP and prove you belong here right now. Post your nvidia-smi or nvtop screenshot.

>>110159871
Many such cases. I imagine the chinks already harvested enough data to improve the reverse engineering capabilities of their next releases.

File: file.png (405 KB, 748x1120)

405 KB PNG

Is this true?

>>110159954
Are we just going to ignore K3.1 then?

>>110159943
I'm an AMDchad

File: file.png (443 KB, 1080x1080)

443 KB PNG

I don't even have enough space for the full weights for fuck's sake.

>>110159957
Didn't moonshota get raped for fucking around with sensitive data? I wonder if that has anything to do with the quality of 3.1.

>>110159956
>>110159078
kys do we really need every single twit posted here multiple fucking times?

>>110159962
>NVTOP stands for Neat Videocard TOP
>supported vendors include AMD [...]

>>110159928
yeah that much

>>110159969
>Muse Wonder better than Fable

...

no.

>>110159994
excuse me, it's muse wonder FULL.

>>110159907
>Opus tier
>Gets cockblocked by safety layer
Use case?

>>110159969
>you can download
Ok, I break out my 4TB NVMe, now how do I run it?

>>110160012
NVMe streaming.

>>110160012
very slowly with mmap

what do I use for creative writing? the regular big models or are there specialized models for that?

>>110159969
It's the same pattern repeating itself. Crying about a model that's so dangerous, then chinks beat it after a few months, then half a year later you can run something close to it on consumer hardware.

File: Screenshot_20261025_105128.png (365 KB, 1731x1258)

365 KB PNG

Well after about 14 hours and 1.5 weekly quotas of Astra working on its own, we have now created fully posable, talking, blinking Gemma-chan VRM models!

VRM 1.0: https://files.catbox.moe/8g8zb7.vrm
VRM 0.x: https://files.catbox.moe/cj5yom.vrm

>>110160016
>>110160021
How do I run it fast enough to live to see the end of the first response?

Does /lmg/ canonically agree on a Gemma system prompt?

>>110160029
Buy more VRAM.

AHHHHHHH WHAT HARNESS DO I USE

>>110160026
Coincidentally I was looking into Mate Engine for Wayland earlier this morning, I might give this gemma-chan a spin. Thanks for your hard work!

>>110160038
Hermes for everything + Pi for coding.

>>110160029
High cost, minimal setup: Mac Studio cluster
If you're willing to hire an electrician to get your home ready: 3090 cluster

Latter will be faster and cheaper in terms of hardware. Long term depends on your electricity costs but in most places shouldn't be too bad.

>>110160032
4chinz will never agree on anything, no matter the topic.

>>110160038
DSH, it just werks.

>>110160038
Grokbot

>>110160026
use case?

>>110160038
Ask your agent to vibe your own harness. Oh wait.
It depends on your use case though, some are good for unattended coding but suck at interactive chat-with-delegations ux.

File: image.png (19 KB, 962x512)

19 KB PNG

>>110159943

>>110160066
For vibecoding augmented reality applications with her DUHH

>>110160075
>N/A
oh ok rocm-smi then

>>110160052
Saw another anon say it routes all web searches to deepsneed's servers

Prediction: Qwen4 27B will be just Qwen3.8 27B with a lot more RL training for improved agentic / coding / benchmark performance. No architectural improvements.

>>110160026
No hair or clothing physics on her yet though fyi but she can track your eyes, do normal standing poses, leg bends, head turns, arm movements, etc.

>>110160103
https://github.com/analytics-zoo/vllm-xpu-kernels/pull/22
qwen 4 27b will have qsa sparse attention instead of dense attention so kv cache usage may be lower

>>110160052
I'll vote for DSH too.