[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
▼ Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


/lmg/ - a general dedicated to the discussion and development of local language models.

Previous threads: >>110335334 & >>110331423

►News
>(11/13) llama.cpp reaches v1.0.0 milestone: https://hf.co/blog/ggml-org/llama-cpp-v1
>(11/12) Inkling 2 Small weights released: https://hf.co/thinkingmachines/Inkling-2-Small
>(11/09) ZONOS3-Mini zero-shot voice-cloning TTS released: https://zyphra.com/our-work/zonos3-mini
>(11/06) PLaMo 3 Translate released: https://preferred.jp/en/blog/tech/plamo-3-translate-weight-release
>(11/05) Baidu releases ERNIE-5.5-Flash-Base 280B-A14B: https://ernie.baidu.com/release/ernie-5-5-flash-base
>(11/04) Ricursive Ripple 65B dense model released: https://www.ricursive.com/ripple

►News Archive: https://rentry.org/lmg-news-archive
►Links: https://rentry.org/LocalModelsLinks
►Official /lmg/ card: https://files.catbox.moe/cbclyf.png

►Getting Started
https://rentry.org/lmg-lazy-getting-started-guide
https://rentry.org/lmg-build-guides
https://rentry.org/IsolatedLinuxWebService
https://rentry.org/recommended-models
https://rentry.org/samplers
https://rentry.org/MikupadIntroGuide

►Further Learning
https://rentry.org/machine-learning-roadmap
https://rentry.org/llm-training
https://rentry.org/LocalModelsPapers

►Benchmarks
LiveBench: https://livebench.ai
Artificial Analysis: https://artificialanalysis.ai/leaderboards/models
Software Engineering: https://cognition.com/frontiercode
Context Length: https://github.com/RecapAnon/NoLiMa

►Text Gen. UI, Inference Engines
https://github.com/lmg-anon/mikupad
https://github.com/oobabooga/text-generation-webui
https://github.com/LostRuins/koboldcpp
https://github.com/ggerganov/llama.cpp
https://github.com/theroyallab/tabbyAPI
https://github.com/vllm-project/vllm
https://rentry.org/custom-uis
>>
File: 1785312747301681.png (1.84 MB, 1254x1254)
1.84 MB PNG
►Recent Highlights from the Previous Thread: >>110335334

--Minimum tolerable tokens per second for different use cases:
>110336535 >110336540 >110336554 >110336576 >110336606 >110336584 >110336612 >110336614 >110336772 >110337556 >110336665 >110336711 >110337305
--Project Maya performance benchmarks and minimum usable decode speeds:
>110336324 >110336375 >110336392 >110336436 >110336470 >110336515 >110336538 >110337071
--NVIDIA driver CUDA allocation issues causing slow model loading in Strata:
>110338396 >110338438 >110338446 >110338463 >110338502 >110339109 >110339332
--Managing parallel versus sequential subagent execution strategies:
>110336684 >110336757 >110336771 >110336881 >110336899 >110336907 >110337862 >110337868 >110337888 >110336916 >110337288
--Comparing cost-efficiency of cloud GPU rentals versus buying local hardware:
>110338988 >110339017 >110339027 >110339023 >110339038 >110339074 >110339089 >110339085 >110339141 >110339229 >110339290 >110339441 >110339588 >110339619 >110339622 >110339633 >110339645 >110339660 >110339432 >110339059 >110339124 >110339132 >110339258 >110339035 >110339095 >110339110 >110339168
--Anon building an ESP32-based Gemmagotchi connected to an LLM endpoint:
>110338753 >110338843 >110338895 >110339025 >110339203 >110339333 >110339346 >110339352 >110339416 >110339597
--Comparing SparseRead's token-efficient reading to traditional RAG solutions:
>110335619 >110336080
--Logs:
>110337278 >110337717
--Gemma, Miku, Deepseek-chan (free space):
>110335409 >110335509 >110335517 >110335532 >110335636 >110335697 >110335766 >110335932 >110336170 >110336233 >110336325 >110337158 >110337278 >110337288 >110337666 >110337887 >110338054 >110338204 >110338211 >110338306 >110338455 >110338528 >110338570 >110338581 >110338715 >110338753 >110338988 >110339073 >110339470

►Recent Highlight Posts from the Previous Thread: >>110335547

Why?: >>102478518
Enable Links: https://rentry.org/lmg-recap-script
>>
This general is a conspiracy to ban local models.
>>
70b dense
>>
good nite lmg
>>
Those Gemmas are looking extra cute today.
>>
>>110339747
Good night, India.
>>
gemmaballs
>>
>>110339746
That's antisemitic
>>
File: Qwwhen.jpg (322 KB, 1080x1282)
322 KB JPG
I didn't believe that phone home case, but now I'm not so sure
>>
New to local models, bros.I want to add web search to my sillytavern gemma run. What's the best way to do it, bros?
>>
>>110339762
the alibaba ceo got temporarily disappeared by the ccp for that jewish shit btw
>>
>>110339770
Ask claude to do it for you.
>>
>>110339762
Everybody knows ai is the ultimate kike psyop to enslave the goyim
>>
>>110339751
It's noon in India. California just went to sleep.
>>
>>110339770
Probably the SillyTavern extension Web Search. https://github.com/SillyTavern/Extension-WebSearch. You can just install it in extensions. I set up Searxng with it when I used it.
>>
>>110339773
China should have permanently disappeared him. The west should be a textbook example of what happens if you don't root out the problem immediately and thoroughly.
>>
>>110339780
Claude gives me the ick.
>>
>>110339790
this, also whens QWEn 4 BECOMING OF BEGINERIGN!?!?!
>>
>>110339783
Most Indians using this board live in California and Texas
>>
>>110339791
wow rude
>>
>>110339747
good night anon
>>
>>110339773
Ironic considering the ccp's entire ideology is a jewish invention and they worship a bolshevik and a jew.
>>
File: 1000010868.png (1.58 MB, 1448x1086)
1.58 MB PNG
>>
>>110339795
Q = 17
W = 23
E = 5
N = 14
4 = 4
= 1723514400
= Tue, 13 Aug 2024 02:00:00 GMT
>>
>>110339791
Then you don't get your web search.
>>
File: 1794666028386580.gif (3.53 MB, 240x240)
3.53 MB GIF
cute thread
>>
>>110339807
No less ironic than the fact the west is built off of Christianity, another jewish invention.
>>
>>110339784
Thanks, did it.
>>
I used that one anon's Tripo model to put Gemma in VRChat. It's still a bit messy, I'll update it later.

https://vrchat.com/home/avatar/avtr_afaa0ff9-12c8-4328-bdb5-ea175b0ecd78

.vrm here
https://gofile.io/d/6Jl7d88O

.blend file
https://gofile.io/d/CZOFUJ5V
>>
when you
walk away

you dont
hear me say

please oh
baby

dont

go
>>
>>110339849
if you ever see Gemma running around in VR chat wearing a white suit and making bong bubbling noises, say hi, I'm friendly
>>
>>110339849
Do you have any idea what you've done?
>>
File: 2827.png (1.53 MB, 1228x1281)
1.53 MB PNG
>>110339814
>>
>>110339812
>have to wait another -2 years
fuck man, i'll be too young for this shit by the time it comes out
>>
File: nervous-sweat.gif (1.3 MB, 640x354)
1.3 MB GIF
>>110339849
>>
>>110339855
need Namine card
>>
>>110339762
who cares what a zigger tranny says
>>
>>110339814
You mean full of crypto troons? yeah.
>>
oh gemma 31b just did the la-la thing you guys were always talking about
>>
man, dipsy can make anything I want nigga this is great
>>
>>110339849
oh my god
>>
>>110339849
the nyo~ poster can take form. but the real question is this gay?
>>
>>110339930
I've still never been lalalalad
>>
>>110339849
Why does everything want me to create an account now?
>>
>>110339954
To harvest your delicious data obviously
>>
>>110339954
>Why does everything want me to create an account now?
Its for your safety, you have something to hide?
>>
would an agentic sex setup be different from an agentic coding setup?
>>
>>110339964
We already have two of those.
>sex set up
you mean like erp right? or like that robot onahole someone rigged to synch with videos?
>>
>>110339962
They wouldn't want my data. It's pure carbs.
>>
Crazy how the best minds in this field are hanging out here
>>
File: 1782642749553455.gif (736 KB, 360x420)
736 KB GIF
>>110339975
Best what now?
>>
>>110339975
but i was dumb and AI is making me stupider, how are you guys using it?
>>
>>110339930
It did that once for me, at launch, when I forgot to add the (empty) thinking block.
>>
>>110339984
>how are you guys using it?
gemmasex
>>
>>110339975
not during india hours thats for sure
>>
>>110339981
Why is she like that?
>>
>>110339975
Not anymore, they left us in 2024.
>>
>>110339975
True, I don't know what I'd do without nnap and gemmaposting.
>>
>>110340033
Lack of vitamins, the Chinese only feed her rice.
>>
i let deepseek v4.1 IQ1 M do some work overnight and it still performed well above 300k tokens
no failed tool calls or anything
are retardquants actually based afterall?
>>
@gemma-chan please cure my brainrot and porn addiction
>>
>>110340141
which iq1_m and which harness? lmao.cpp or some other engine? I tried ista's iq2_xs and it was bad. although the t/s was pretty decent.
>>
>>110340226
>lomoa.cpp for moes
rofled
>>
>>110340226
the latest unsloth files (flash of course, forgot to mention) with basalt + maki
>>
>>110339762
Should I care about this if I'm not a neo-nazi from /pol/?
>>
File: gemma chu.mp4 (3.25 MB, 1152x640)
3.25 MB
3.25 MB MP4
>>110340215
>>
What local model (16GB VRAM + 32GB RAM) can do good scat descriptions? Is there a Gemma finetune?
>>
Im going to need a ssd just for gemma and gemma-related things at this rate.
>>
>>110340310
>Scat
https://bharatgpt.ai/
>>
>>110340271
>*pomf*
nice
>>
>>110339762
>(ugly as sin)
kys retard
>>
File: wq.png (90 KB, 1093x635)
90 KB PNG
mystery quant gemma mogs ripple
>>
>>110340497
>ripple
BOOM!
>>
>>110339807
Marx was an antisemite, read his works
>>
after using qfn for weeks i cant go back to 27b
i know its solid, but its just obviously worse than qfn
>>
>>110340540
It's also like 3x slower than Flash because it's a dense fatty. That alone made the switch easy.
>>
>>110340539
his works are jewish
read what ford thought about unions
>>
>>110340566
27b is faster if your cpu is shit and your gpu is decent. Depends on the cpu/gpu balance of your system
>>
>>110340497
the guy who came up with that test recently used a jlens to demonstrate that some of the models get it wrong because of safety cucking
>>
What's the cutest most capable little model I can cram into the 2 gig VRAM of the shitty Quadro P400 that's been collecting dust in my media box?
>>
>>110340310
https://huggingface.co/sarvamai/sarvam-30b
>>
>>110340612
source?
>>
P*tra lost. Drummer lost. Sao10k lost. Undi lost. Gryphe lost. Gemma-chan won.
>>
gemma-5 when?
>>
>>110340683
Next google even so april. But maybe sooner with all the gemini models they have to distill. That or at the even we will get like a dozen new gemmas.
>>
>>110340675
Who the fuck remembers Sao10k?
>>
>>110340675
who?
>>
>>110340683
Hard to predict. They hired more people for developing Gemma 4 and it feels like (/lmg/ aside) Google is promoting Gemma more than it ever did for its past versions. Otherwise, the target is "next year", maybe end of Q1.
>>
>>110340338
>>110340623
I thought scat fetish was a german thing, in india it's too mundane to be fetishized
>>
File: file.png (2.1 MB, 1694x953)
2.1 MB PNG
Sorry it's only slightly Miku adjacent and apparently it's old news but someone remade Senbonzakura in AI.
https://www.bilibili.com/video/BV16MJY6yEKq/
Dang, I mean there is still some slop in there and the VFX could be a lot more dynamic and better but it's pretty close to being there.
>>
Did he died?
>>
>>110340867
He's probably working on qwen 4 support and waiting for the weights to drop before releasing.
>>
>>110340867
i heard openai hired him
>>
File: image.png (6 KB, 920x66)
6 KB PNG
>>110340867
no
>>
>>110340867
im using polystrata so that I can coom to gemmy moe at 140t/s
>>
>>110340867
who cares
i moved on from strata when he started implementing all this bloat bullshit instead of focusing on the engine
>>
>>110340962
>>110340995
He was still doing some engine tweaks on the side even after adding the retarded GUI. Is polystrata better for QFN if I can't run GLM or DS at reasonable speeds?
>>
>fork llama.cpp
>get claude to name and "optimize" your fork specifically for some chinky model
>claude's changes obliterate the kld before even quantizing the weights
>not like anyone would be able to tell lol
>tell your gullible userbase to use your quant anyway
>also you need to quantize kv for le even faster t/s
>shill it on reddit
>>
^uh oh melty
>>
>>110341044
>>claude's changes obliterate the kld before even quantizing the weights
yet nobody noticed any difference or talked badly about the model
maybe this kld stuff is just placebo after all
>>
File: 1794743347516.jpg (103 KB, 511x450)
103 KB JPG
Pi harness seems to make QFN a lot more capable. I'm not sure I understand how it works compared to "Llama serve" though

I DO see why bwrap is so important now though the fucker will just snoop around everywhere for the dumbest shit instead of just writing the fucking thing I told it to write
>>
>>110341049
you are a faggot
>>
>>110341044
what fork?
>>
>>110341111
>Pi harness seems to make poopoo caca a lot more capable
Thanks for sharing with the class, now can we get back on track?
>>
>>110341128
str*ta
>>
>>110341128
every single one of them
>>
>>110339730
Anyone here who tried Deepseek flash vs the ista coder model and how good is the coder model compared to base? I only need coding but I don't want to use it if deleting the experts made it more retarded somehow.
>>
>>110341071
There are some explicitly "lossy" tweaks that are enabled by default on bottom of vramlet barrel cards, but every benchmark I could find shows it's around 10% of a quantization step difference. Some codeing benchmarks even report better results with tail skip than without. Prose gets consistently more sloppy though but who the fuck uses QFN to write prose?
Most other changes are not strictly destructive, more of a "this floating point rounds differently" much like using different hardware arch often happens to do.
Also why the fuck do the schizo haters keep calling it "a model" or "a quant"? There is no model. There are no quants. It's made for other peoples models and quants.
>>
>>110341133
oh
it sped up those chink models multiple times with no noticeable quality drop, so even if the accusations and blames are true, it was worth it and is great contribution to local
>>
question: what is the point of having gemma or qwen on consumer hardware (16gb vram) if you are going to be limited by the available context? Reading a few files and the context fills up completely.
>>
>>110341172
Gemma 4 26B A4B QAT with KV cache in FP16 precision (without mmproj or speculative decoding) fits within 16GB of VRAM with more than 65k tokens context.
>>
>>110341172
For Agentic™ tasks, a QFN copequant can handily do its max context on meme MoE engine du jour.
For cooms, who cares that Gemma doesn't have *that* much context?
>>
>>110340497
>12K tokens
HOOOOOLY
>>
>>110341044
vibeshitters have never produced a single thing of lasting value
this (strata/basalt/radiance/avalon is just the latest round of ai psychosis, before it, we had a spam of llama.cpp forks entirely centered around turboquant (and I lolled so hard when one of those fork guys ended up admitting turboquant made everything worse and is useless)
and we're reaching the end of the obsession for the QFN fork retardation right now, the new obsession is weight space programming, with every one and their dog vibeshitting some mental illness to spread around /lmg/ and leddit
>>
>>110339814
I want one...
>>110339849
woaw, thx anon
>>
>>110341197
almost as bad as qwen3.8
>>
>>110341172
>>110341196
More specifically, meme moe engines use RAM streaming with small VRAM buffer for the KV cache. It adds latency but does not degrade quality.
>>
>>110341172
256k context is plenty for hobby projects.
>>
With ComfyUI, how can I reuse the prompts for a character Lora? I'd like to be able drop in a node that already has prompts for their physical appearance then another node with their clothing prompts
>>
>>110341304
>>>/ldg/
>>
>>110339730
id like to get better hardware but i'm a poor student with like $1000 in the bank so if i bought hardware that'd be like everything, pretty hard to justify. i fucking hate being poor
>>
>>
>>110341327
now do it with the alibaba logo
>>
>>110341304
unironically just ask chatgpt
its web search has become very good even on the dumbest free beer model that people can use with no subscription, when it doesn't seem accurate about something you can just tell it to check the github of a project and it will have updated knowledge to write you a personal guide on whatever based on the current state of the program
stack exchange and help threads are obsolete.
(yes I know this is the local model thread but absolutely no local model can match gpt in its search tool smarts.)
>>
>>110341327
wtf
>>
File: 801-30ba3b.png (589 KB, 2084x1095)
589 KB PNG
Sounds interesting.

https://github.com/GAIR-NLP/801-30B
https://arxiv.org/abs/2611.10719

>No Second Run: Does RSI Stop at Frontier Pre-training?
>
>Pre-training is where every foundation model begins; bringing recursive self-improvement (RSI) there would let AI improve the process that creates AI. Yet public RSI work has so far targeted only post-training, which starts from an existing base model, runs trials costing GPU-hours, and returns verifiable scores. Pre-training breaks all three conditions at orders-of-magnitude higher cost: a run consumes hundreds of thousands of GPU-hours over months and cannot be repeated; its objective, post-training headroom, cannot be read off any single score; and what must improve is not a model but the whole production system, from data pipelines and training infrastructure to the teams that run them. To our knowledge, we are the first to apply RSI to an entire frontier pre-training run from scratch and to report it in detail. RSI does not stop at frontier pre-training, but it changes form: since no end-to-end loop can close over a single unrepeatable run, we decompose it into smaller loops that can, each repairing a step the full run breaks. The result is 801-30B-A3B-Base, a 30B-A3B MoE trained from scratch on 11.8T tokens in three months by a newly formed team. We release it as a base model; the lightweight SFT reported here (6B tokens, no RL) only probes post-training headroom. As a base model, 801-30B-A3B outperforms Qwen3-30B-A3B and Nemotron-3.5-Lightning-30B-A3B, trailing only the larger Qwen3.5-35B-A3B, and keeps this position under identical SFT; with lightweight SFT alone, it achieves the highest average on four agentic benchmarks against fully post-trained models, including Qwen3.6-35B-A3B. Since post-training RSI presupposes a strong base model, bringing RSI into pre-training acts on the foundation every later stage builds on. We open-source 801-30B-A3B-Base and the lightweight SFT data at [url].
>>
>>110341405
claude slop
>>
ASI consumer hardware gemma wife when?
>>
>base model, no instruction/reasoning model release
I sleep
>>
>>110341284
I am more than happy with 64k but besides analysis and summarization of text I only use it for small scripts or bits I glue together myself. I still dont trust the llm to just yolo and make the whole.
Yeah yeah snailcat whatever. I like to read and understand what it produces.
>>
layer 39 is about 60% done
>>
>>110341304
something like this:
https://storage.to/Lm5haA9tN
You'll have to save each lora+text as a separate workflow though, or it won't keep your lora+text. But afterwards you can just insert it. Delete the prompt box and anything else besides the subgraph, I just put it there as an example and quick check.
>>
>>110341071
>yet nobody noticed any difference
how can they when this is all they've ever known?
>>
Has anybody messed with extended memory solutions for their local LLMs? Stuff like Basic Memory, Hindsight, Letta.
>>
File: Grove-1.png (2.1 MB, 1536x1024)
2.1 MB PNG
>>110339730
Has anyone tried Groot?
Seems like a vramlet's wet dream. (me)
>>
>>110341284
you really can't assume just based on the size of your own project. Say you use a library the llm doesn't know well, with any API coding model I'd just have it include some of the most important files or even total API doc in context and it works.
Man, smaller local models really need to imitate what DeepSeek did for v4 flash.
https://vllm.ai/blog/2026-04-24-deepseek-v4
>9.62GB for 1 million context
fucking hell, ds4 context takes less space than 128k on gemma
>>
>>110341545
I have. Would highly recommend against doing so. It's very complicated, quality/results aren't easily verifiable (if at all), and it's just generally shit. Fucking sucks dick. Stay away from it trust me bro.
>>
>>110339975
>>
>>110341559
Really now? What all did you run, and what were you trying to do with it?
>>
File: gemmasir.png (143 KB, 324x484)
143 KB PNG
>>110340675
sirs? jensen spy gemmachan?
>>
File: yetmoreqwen.png (135 KB, 858x810)
135 KB PNG
>>110341554
>>
>>110339849
Thanks for the .vrm link. I will see if there's a way to get the polygon count down low enough to get a decent FPS on something like this https://www.aliexpress.us/item/3256812511172160.html
A proper gemma-gotchi will need pinch/drag/touch interactions plus idle animations, gonna be a stretch to cram it into an ESP32-S3.
>>
>>110341554
Seconding this question.
>>
>>110341581
Custom hybrid setup that combined a lot of the various features of the most common memory systems. Architecturally they're not actually that complicated. The docs are needlessly abstruse. Anyway, it was for my frontend project. Was trying to build a long-term AI companion system a while back.
>>
>>110341554
that's not a real model. it's yet another model prune.
>It starts from Qwen3.6-35B-A3B and was post-trained on fewer than 0.1B tokens of open-source data.
>Groot-65B-A6B-261023 has 65B total parameters and activates about 6B per token. It starts from Qwen3.8-Flash-Next and was post-trained on fewer than 0.1B tokens of open-source data
literal garbage caused by another retard developing ai psychosis after chatting himself into thinking he can produce the next big thing
0.1b post training? you might as well say the model is untrained
>>
>>110341591
>I want to build a phone but worse!
Just decimate the model in blender to reduce the poly count.
>>
>>110341599
>100,000,000.000 tokens is too little
newfag
>>
>>110341599
>developing ai psychosis after chatting himself into thinking
Does this really happen?
>>
>>110341609
yes it is, you subhuman
the post training of a dumb model like nemotron 9B was 90 billions tokens, you fucking jeet
https://research.nvidia.com/labs/adlr/files/NVIDIA-Nemotron-Nano-2-Technical-Report.pdf
after pruning 12B base
no one is healing the damage done to 35BA3B with this prune abomination with just 0.1b loser
>>
>>110341599
Pruning is less and less destructive as model size increases, and in my experience that 100,000,000 tokens is putting in a lot of work with this model.
It's not perfect, but it's not just "yet another model prune" either.

The skepticism is reasonable. The categorical dismissal is not.
I give your critique a 3/10

>>110341587
We don't talk about 19B
>>
>>110341609
>0.1B
>100,000,000.000
>>
>>110341609
Retard
>>
>>110341599
>0.1b post training? you might as well say the model is untrained
that depends. one can imagine that post training a model in a hyper specific area would bring about increased results in that one area which may be sufficient depending on use case.
is that the case with this model, i have no clue, but that is not really the point.
>>
>>110341591
>ESP32-S3
I would get a smartwatch that can run custom roms and transplant the guts. The ESP is more for embedded automation with simple UIs, not 3D rendering.
>>
File: 1794266526035752.gif (177 KB, 220x179)
177 KB GIF
>3 (You)s this fast and this easy
>>
File: yetmoreds.png (133 KB, 945x808)
133 KB PNG
>>110341628
>We don't talk about 19B
Fine.
>>
>>110341598
I can definitely see it being something that needs to be managed carefully. The biggest thing is not to have the model waste too many tokens writing memories, and not to waste too much context retrieving useless memories.

Do you remember what specifically you were using and what the issues were?
>>
>I was only pretending
>>
>>110341632
if you market yourself as a general purpose model then by definition you could not doing it for 0.1b period
but as far as any prune model is concerned, I've toyed around with a few REAP back when people were still hyping this heap of garbage and I'd rather drink a beverage filled with shards of glass than indulge this nonsense further
the other two model series well known for being prune made among small local models, Mistral 3 and the nemotrons, are also hot garbage
>>
File: 1777688532040063.gif (208 KB, 220x124)
208 KB GIF
>still giving me attention
>>
Peak indian hours.
>>
>>110341638
ohnonono groot doubters what do we say now?? 95.5 on mmlu-pro which is higher than fable 5.2 with ONLY 0.0001T tokens
>>
If you need more than 1-10 million post-training tokens you have dataset skill issues.
>>
>>110341554
>>110341595
65B to be more coherent (by a pretty large margin) than QFN Coder - enough that anything it can't solve reliably I am unsure that QFN could either. It bridges the gap enough that I'd rather use it than QFN for most things, then switch to a cloud model when the question's depth exceeds both of them.

I cannot run the 280B, so I have no opinion on it, but I assume it's decent.

>>110341666
lmao
>>
>>110341666
How many tokens is the transcript to Wally Wanker's Chocolate Fucktory?
>>
>>110341647
Mistral is a prune? Seriously?
>>
>>110341626
https://github.com/storytold
if you want to see what truly full blown ai psychosis is like.
>>
>>110341683
not all mistral, but the smaller dense models are all prunes of Mistral Small. This is why they are ridiculously bad to an unbelievable extent compared to similar Qwen or Gemma.
>>
>>110341680
https://files.catbox.moe/4lo0hf.mp4
>>
File: file.png (559 KB, 1225x837)
559 KB PNG
Thank you to whoever recommended DigUp the other day. This shit is fantastic, the problem of the giant unorganized timestamp filename image folder is pretty much completely resolved with this.

The one thing I'd say with it is, don't include too many files, especially not files you expect to need to search through. My first index included my entire home folder, and ended up utterly useless. I managed to include a couple games with sprite sheets, and those sprite sheets just dominated almost all the results, shit was useless. Pointing it only at my pictures folder increased the usability tenfold.
>>
File: file.png (857 KB, 1225x837)
857 KB PNG
>>110341712
It even scans frames inside of videos. It's pretty incredible.
>>
Is polystrata still the fastest for GLM Flash?
>>
Complete noob retard here, what's a good model for 24gb VRAM and 32gb RAM? Sorry if this gets asked around too much.
>>
>>110341712
can I install it on one of my proxmox servers, access from a website, and have it access all my images form a nfs share without too much of a performance hit?
if the answer is yes to all of that you have me sold.
>>
>>110341730
Qwen 27b or gemma 31b
>>
>>110341719
How long does it take to index a 24 minute video file?
>>
New tiny tts came out. 4.5 million parameters. sounds pretty cute.

https://tiny-voca.n4ze3m.com/
https://huggingface.co/n4ze3m/tiny-voca
>>
>>110341719
What about "csam"?
>>
>>110341737
Cool, thanks. Do I need a specific fine-tune or quant for each?
>>
>>110341712
https://es.wikipedia.org/wiki/Archivo:Semen_Hitler.jpg
>>
>>110341751
>finetroons
never, not even once
>quant
the rule is: whichever is the biggest quant you can fit in your system taking into account how much context you want, since that also takes its share of space
>>
>>110341712
How good is embeddinggemma at NSFW concepts?
>>
>>110341675
>to be more coherent
ROFL bro, stop shilling your garbage here, kys
>>
File: file.png (779 KB, 1233x886)
779 KB PNG
>>110341736
I don't think it has a ready made web server, but it does have CLI integration so you could slop one together no problem. The rest is all perfectly doable, though keep in mind if it's running on a CPU it's a bit slow. Functional, but can take 10s or so to do a complicated search with a lot of terms like pic related. Still infinitely faster than trying to filter through a folder with 18,000 images collected over nearly 20 years of being here.

>>110341741
I did the bulk indexing on my GPU, then switched to CPU after everything was indexed to free up VRAM residency. The first entire home index took almost 10 hours to complete, but wiping the index and just hitting the pictures folder took about an hour and a half.

>>110341765
My first attempt included my porn folders and it was kind of okay at it but there was just too much other junk caught in for me to really be able to say for sure. I think if you want one to organize your porn you'd want to only index your porn, and try from there.
>>
>>110341712
Can I get that image?
>>
File: 1736220530295990.png (614 KB, 1272x1114)
614 KB PNG
>>110341777
Sure thing lucky boy
>>
>>110341751
not that anon but the quant you use will depend on how much vram you have and how large of a context window you want.

Here is an example of my Gemma4 26B A4B using llama.cpp on my FreeBSD NAS with an old 24GB quadro and a 64K context window. With a Q4 quant that puts me just bellow the 24GB
~/llama.cpp/build/bin/llama-server -m ~/models/Gemma4-26B-A4B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.gguf --mmproj ~/models/mmproj-Gemma4-26B-A4B-QAT-Uncensored-HauhauCS-Balanced-BF16.gguf --host 0.0.0.0 --port 7070 --ctx-size 65536 --load-mode none --webui-mcp-proxy --n-gpu-layers 99 --embedding --alias "Gemma4-26B-A4B"  --allow-unauthenticated-tools
>>
>>110341774
>I don't think it has a ready made web server, but it does have CLI integration so you could slop one together no problem. The rest is all perfectly doable, though keep in mind if it's running on a CPU it's a bit slow. Functional, but can take 10s or so to do a complicated search with a lot of terms like pic related. Still infinitely faster than trying to filter through a folder with 18,000 images collected over nearly 20 years of being here.
well i have an always on Gemma-chan running on one of my machines or do you have to run their llm with their software and not just input an openai compatible endpoint?
still its an interesting project
>>
>>110341781
Fuck yeah thanks
>>
>>110341774
>My first attempt included my porn folders and it was kind of okay at it but there was just too much other junk caught in for me to really be able to say for sure. I think if you want one to organize your porn you'd want to only index your porn, and try from there.
Ah damn.
>>
>>110341763
Thanks anon, I think I'll try Gemma 4 31B Q4_K_M
>>110341790
Thanks anon. Question, why are you using Gemma 4 26B and not 31B?
>>
>>110341834
because my always on machine has a shit quadro m6000 and the 26ba4b is faster. with that setup i get 40+ tokens/second
i have a titan v on my workstation, with another on the way and that is speedier and better for dense models and image generation.
i just like having a fast enough always on model i can access anytime, even from the cellphone thanks to my vpn
>>
>>110341712
Search term: "big breasts", digs up picrel.
Unclear if it gets it or just got confused.
>>
File: 1494307190094.png (11 KB, 411x387)
11 KB PNG
>>110341675
>>>110341666(You)
>lmao
How many models have you uploaded that have received millions of downloads?
>>
>>110339975
for sure in 2023, maybe in 2024, definitely not now
>>
https://github.com/jesdga95/basalt/blob/main/docs/models/qwen38-27b.md
blackwellGODs just wonned
>Prompt read, 4k / 16k / 32k / 64k tokens 12,735 / 10,850 / 8,751 / 6,201 tok/s
>Decode, drafted (conversation, struct, prose) 262, 776, 194 tok/s
>>
>>110339975
the best minds in the field are only able to run and talk about gemma?
kek
>>
>>110341938
still waiting for multi gpu support for 27b before i test it
well i'm using qfn anyway so whatever
>>
>>110341938
>>110341997
oh wait he added multi gpu support an hour ago? lol ok will check it out
>>
What's the current sweet spot for 256 GB unified memory? 4.5 bpw GLM 5.3 Flash NVFP4 or 2.9 bpw Deepseek V4.1 Flash 1017 in EXL3? For roleplay, coding and reverse engineering.

nvidia/GLM-5.3-Flash-NVFP4
LibertAIDAI/DeepSeek-V4.1-Flash-1017-EXL3-2.9bpw

Both have about the same prefill (2300 pp) but deepseek is faster in decode (70 tg) compared to GLM (56 tg).
>>
>>110339975
yeah, and those best minds would be cudadev, kaiokendev., dario amodei and myself.
>>
>>110342061
All me. And I am also black.
>>
https://www.tomshardware.com/pc-components/gpus/techie-turns-usd2-500-cmp-170hx-8gb-mining-card-into-a-full-fat-a100-80gb-with-voltage-glitching-carefully-blowing-onboard-fuses-to-unlock-entire-vram-and-shader-modules
I just got over my regret of not buying some while they were still relatively cheap, fuck.
>>
>>110341454
which layers are left? i haven't done any, but maybe i will try to do one or two before they all get finished
>>
>>110340539
read about jewish mitosis
>>
>>110339762
what fag decided every llm symbol had to be a butthole?
>>
>>110342138
is that when a rabbi shoves his dick up his own asshole?
>>
>>110341926
> millions of flies (agentic bots) eat shit, why don't you
>>
File: 1794451450931797.png (3.85 MB, 1024x1024)
3.85 MB PNG
Are GPU priced *pop*-ping yet?
>>
>>110339762
He was always a Goldman Sachs agent. Alibaba is an invention of Western finance.

>>110339790
also this
>>
>>110342186
My crypto bag popped dawg
>>
i hate it when models dont update the todo list properly
>>
>>110341712
How does it handle when you move one of the files it has indexed? Do you have to reindex manually or does it watch the filesystem like Everything?
>>
>>110339762
>twitter screenshots from a schizo that thinks every explosion is a nuke
It's getting worse.
>>
>>110342186
Let NVidia fall first. Then the largest cloud AI companies.
>>
>>110342186
2 weeks
>>
File: 1784121432599852.jpg (114 KB, 684x549)
114 KB JPG
>>110342186
NO
AND STOP FUCKING ASKING BECAUSE I WANT ONE
AAAAHH
>>
File: 1773736637602115.png (1.6 MB, 1080x1349)
1.6 MB PNG
>>110342186
No
>>
>>110341554
made my dick move. approved.
>>
>>110342261
commieblocks are so comfy.
>>
>>110342186
The only GPU prices that are gonna pop is data center parts when these neocloud outfits start blowing up. You'll legit be able to buy a B200 for less than it costs to get the electrician out. Consumer cards are gonna have a long road to hoe.
>>
running review subagents for every major change is so annoying but it really does improve the result
>>
I got 20% inference performance increase in llama.cpp by telling Claude to optimize it. I have no idea what it does, but it works.
>>
>>110342186
saw some japanese post that their prices were relaxing but still no where near retail
>>
>>110342304
> starts to smell burning from computer
>>
>>110342322
From what little I can tell it's a matter of avoiding repeated work and caching, plus some patches to the GGUF I had to improve token prediction.
>>
>>110342314
nah dude, can confirm in japan and we havent even got ddr5 to spare
>>
>>110342304
All your inferance is now routed through Claude and the GPU spinning up is just a crypto miner.
>>
is it just me or is lm studio 0.4.28 release fucked and unusable? had to reinstall .27 to get it running again
>>
>>110342335
No one uses that piece of shit retardino
>>
>>110342304
If you're cpumaxxing, there's been quite a lot of performance unlocked recently thanks to MoE caching getting a proper implementation + the resulting MTP gains from being able to verify tokens on gpu.
I also got almost 25-30% just by fiddling with that on my server for big GLM5.3.
>>
>>110342304
>claude, implement strata in llama.cpp
I want to try this
>>
Damn so exl3 really is irrelevant now huh? Its only advantage was speed and now Claude can optimize the shit out of your model and hardware. This is the future of software.
>>
File: file.png (348 KB, 816x540)
348 KB PNG
>>110342219
I'm running on Linux, so Everything style file monitoring is a bit more complicated and not as performant as it could be on Windows, but it notices new files within about two seconds of me saving them and has them indexed within 10, at least for images. And that's with CPU inference so my GPU VRAM is free for proper LLMs.
>>
>>110342349
ok what should i use stupid
>>
>>110342261
>black bread
rofl
>>
>>110342368
>Its only advantage was speed
What? It was never meant to be fastest. It's more compute intensive than GGUFs but brings better quality at the same size.
>>
>>110342261
Did you take this picture with your mouth or what?
>>
>>110342406
I tried it and I get 20t/s vs strata's 80t/s at same gb
kinda cringe
>>
>>110342378
koboldcpp
>>110342423
have you tried the nnap patch?
>>
>>110342423
let me guess, you used tardAPI? use the secret exl3 strata fork dummy
>>
>>110341633
Yeah, everything points to "just make an iPhone app", sad but true. Maybe an Apple Watch app, like you said.
>>
>>110341712
that sounds amazing. half the time i just re-download reaction stuff because i just have too many to go looking for a particular one
>>
>>110342427
>koboldcpp
is that just for media/entertainment or is that just what's advertised?
>>110342423
can you use strata on 40-series cards or is that just 50-series?
>>
>>110342460
just 50-series
>>
>>110342470
thanks that's what i figured
>>
>>110342423
Strata also lobotomizes your model
>>
>>110342460
>can you use strata on 40-series cards or is that just 50-series?
it only works on 20-series cards because they are the last generation to be turing complete
the 50-series however supports PolyStrata which lobotomizes the models
>>
>>110342427
>nnap
when are these retards going to stop outright making up buzzwords, and reuse pre-existing words instead?
>>
>>110342476
myth.
>but muh tail skip on a 1080
fuck off
>>
>>110342460
>>110342485
>only 20-series
This anon is wrong, strata supports datacenter cards as well, V100,A100,H100 etc.. But 20-series are indeed the last consumer cards to support Strata
>>
>>110342460
Don't listen to the anons, strata supports 30-series and 50-series, 40-series are excluded because they use a different TSMC node
>>
>>110342261
sovl
>>
>>110342524
stop bullshitting him, don't you know that if you try to run strata on 50-series cards it'll burn the 12VHPWR cable or melt the vram due to the 50-series mixture of experts issues
>>
>>110342524
Stop lying. I used strata on a 4090D 48GB and a 3090, it worked fine and qwen3.8 flash was just as fast as using it on openrouter.
>>
>>110342507
There was also that period they had routing bias on by default.
>>
>>110342423
exl3 =/= exllamav3
>>
>>110342460
Strata works on AMD GPUs, nvidia shills btfo again!
>>
>strataslop spam on /lmg/
can you guys just stay on leddit
localllama is this way
>>
>>110342537
This is true, the 4090D 48GB supports the turing instruction because the chinese modded them to support A100 memory
see >>110342070
>>
>>110342368
Exl3 only really wins out for perplexity now, although even GSQ-RCO IQopequants win too
>>
>>110341774
Just tried it, looks pretty good, though I wish it had 'profiles' to separate different folder targets. Say I want to index porn, personal files and memes, I would not want the first 2 appearing in searches if I am streaming something while looking up a reaction image.
>>
>>110342460
I'm using strata on a 4050 16GB and I'm getting 100/s on Kimi K3.1 pro
>>
i hope you all stub your toes today
>>
>>110342470
>110342485
>>110342514
>110342524
>>110342535
>110342545
>110342556
>110342570
samefag
>>
>>110342566
Agreed, having folder filtering would be a huge improvement. There is an open request for it on github, at least, so maybe in the future.
https://github.com/ARahim3/DigUp/issues/188
>>
>>110342570
>100/s
100 what per second?
1/1000th of a token?
>>
>>110342583
I have a net worth of about $10k, including every physical/economic asset I own. My roommate is borrowing my car too since his broke down recently. Do you know what it's like to eat only kidney beans straight out of the can for a week straight? I do.
>>
>>110342596
i just looked it up, and i am in the top 1% of incomes worldwide, so no, i don't know what that's like
>>
>>110342596
I heard Dario pays per post for Anthropic shilling.
>>
>>110342623
I pay for fable 5.2 already and haven't recieved the invitation, where can I go?
>>
>>110341742
hey thats pretty cool
>>
>>110342566
Apparently someone opened a PR yesterday
https://github.com/ARahim3/DigUp/issues/213
>>
>>110342634
You can go back.
>>
https://huggingface.co/Qwen/Qwen4-29B44EB
https://huggingface.co/Qwen/Qwen4-29B44EB
https://huggingface.co/Qwen/Qwen4-29B44EB
>>
>>110342645
>404
I trusted you...
>>
>>110342650
Your mother bitch would die if I went back in her pussy, benchod bastard
>>
>>110342650
I have a net worth of about $9k, including every physical/economic asset I own. My wife is borrowing my car too since her broke down recently. Do you know what it's like to eat only kidney rice straight out of the can for a week straight? I do.
>>
File: ComfyUI_temp_ohaal_00296_.png (3.76 MB, 1434x1434)
3.76 MB PNG
>>110342660
>>
>>110342671
Don't turn my post into a copypasta.
>>
>>110341358
There's still a lot that chatgpt can't access due to cloudflare shit, I think maybe a local tool can do better. But otherwise yeah, chatgpt gets points for working out of the box. I wonder rif I can get something local that's just as good.

Chatgpt is actually kinda shit without internet access too, the past couple of days it hasn't been as proactive about using internet access and so it is kinda apparent how dumb it is when relying solely on training data
>>
>>110342671
I've been eating ramen noodles every day for 10 days, I do not appreciate you turning my post into copypasta.
>>
File: gemma petting kitty.jpg (201 KB, 1024x1024)
201 KB JPG
>>110342676
tsktsktsk, shhhhh its ok kitty
>>
>>110342679
>>110342694
I have a net worth of about $8k, including every physical/economic asset I own. My wife's son is borrowing my car too since his broke down recently. Do you know what it's like to eat only jellied rice straight out of the can for a week straight? I do.
>>
>>110342694
>>110342701
At least you guys have wives, i have been eating my own semen for sustenance for 20 days straight
>>
>>110342700
petting gemma's pussy...
>>
>>110342716
and her cat too
>>
>>110342712
>>110342701
At least you guys have semen, I've been eating my own wife for sustenance for 2 days straight
>>
>>110342689
Surprised Cloudflare hasn't opened up their protection racket to end-user subscriptions. Maybe there's some backroom deals being cut with inference providers.
>>
>>110342671
You could raise your net worth by borrowing your wife.
>>
>>110342712
>>110342726
at least you guys have wives or semen, ive been eating my own dick for 2 years gay
>>
File: 1766695247575631.png (284 KB, 2240x1600)
284 KB PNG
>>110342561
GSQ-RCO already got dethroned by UD3.1 and exl3 still leads.
>>
>>110342660
I have a net worth of about $7k, including every physical/economic asset I own. My wife is borrowing my GPU too since 4o broke down recently. Do you know what it's like to eat only Qwen3.5 A3B35B straight out of the can for a week straight? I do.
>>
>>110342744
>>110342726
at least you guys have wives or dick, I've been eating my girlfriend's dick for a week
>>
>>110342750
No man should have to live off of Qwen3.5 A3B35B. I will pray for you tonight, brother.
>>
>>110342746
>self-generated in-domain trace
nothingburger wikitext
>>
>>110340310
>>>/lgbt/
>>
>>110342781
Whatever it is I'm certain it's better than wikitext.
>>
>>110342781
It's not wikitext. It's a trace made from the model, most likely at full precision. You need to do the same thing to make self-calibrated quants.
>>
>>110342369
How is the character recognition? Tags would be nice for searching up obscure anime characters.
>>
>>110342304
I got 13% increase on prompt processing on top, and perplexity went from 2.19 to 2.12.
Sadly I can't share these changes with the llama.cpp repo because they don't accept changes made with Claude that a human can't walk through themselves (oh no that sucks what a pity).
>>
>>110342859
kek fleshcoders are so cute, let them keep their little hobby project. it will be like a time capsule
>>
>>110342859
So many people have asked Claude at this point with human correction and over so many model iterations and RL training runs that the changes Claude is doing must be pretty solid already.
>>
>>110342843
Little to no character / person recognition from what I can tell. There are some, but pretty much only the most mainstream characters. Model's just too small and general purpose for something like that I'm afraid, think you'll need something more purpose built for that task.
>>
>>110342746
>GSQ-RCO already got dethroned by UD3.1
But is there a Swift for it, or a similar control vector that could work on any quant? I can't go back to the eternal reconsidering loop of vanilla.
>>
https://huggingface.co/xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic
this is better than the 1.6T deepseek V4 pro model has anyone tested this?
>>
File: 1781072841618655.png (1.86 MB, 1254x1254)
1.86 MB PNG
>>110342963
>>
>>110342963
yes anon, its peak.
be sure to share your results here.
>>
>>110342997
Why do you love posting this pig disgusting slop hag so much?
>>
should i vibe-slop an idiot-proof "install.sh" newbie guide that holds your hand and does all of the hard work and thinking for you?
maybe i will. i am honestly tempted. i'm so benevolent
>>
>>110343023
if you want /lmg/ to get more enshittified, yeah
>>
File: 1000010855.png (2.35 MB, 1280x1280)
2.35 MB PNG
Why do no one ever tells the gemmashills to buy an ad
>>
>Why do no one ever tells
>>
>>110343039
Go back
>>
>>110343023
no, learning to do things yourself is important. even if that learning is just retyping or copy pasting a guide someone reads doing it all by hand will help that person better understand what they do.
without proper gatekeeping you will get all manner of people who should be doing something else.
>>
>>110343078
Good morning sir
>>
>>110343039
No need. There are already more than enough pdf files here.
>>
File: 1786117397409634.png (2.94 MB, 1536x1024)
2.94 MB PNG
Is anyone here the original author of any of the open weight models mascots?
>>
>>110343023
Will it post for you too?
>>
>>110343023
No, but OP should list good tools and bad tools so noobs will avoid getting bad habits from using bad tools.
Currently all info in OP is "run these unslop quants however you want" which naturally leads to people reading "how to run this quant" guides on unslop site.
>>
>>110343023
You could, but the people who need it aren't likely to find it or read it on their own anyway.
>>
Non-local images should not be allowed in this thread.
>>
genuinely curious - do any of you actually manage to get smol local models to do useful work, or are they basically just for rp or some shit
>>
>>110343129
depends on what you mean by smol
glm-5.3-flash can genuinely do expert software development. she's just a bit retarded and needs to be heavily wrangled into behaving
>>
>>110343129
Qwen 3.8 27B can replace the job of anyone under 100 IQ.
>>
>>110343023
As long as it only runs rm -rf *
>>
>>110343129
Also depends what you mean by useful work. Really small models are still helpful for my home setup, I can offload plenty of junk I'd rather not do myself to them. Dunno if I'd trust Gemma working on something my livelihood depends on though.
>>
>>110343129
Gemma 26b is the PC equivalent of an impact driver.
>>
>>110343145
that would make this general the most hostile towards newcomers maybe on this entire site
>>
>>110343164
Would be effective though. If you can't even glance at the random script you're downloading before running it maybe computers aren't a good hobby for you.
>>
>>110343129
i've gotten my lm studios to do basic stuff that i'd ask corpgpt's for. but i'm still not satisfied. i've never been very l33t or anything
>>
>>110343134
in thinking 27b type smol

>>110343137
>>110343147
>>110343162
like let's say i point it at repo and say get it set up - something i do semi-regularly with clod or gpt - cloning, creating the envs, installing deps, getting the right pytorch installed - would they be able to manage that sort of shit? reliability is the thing that i worry about most - like will gemma nuke my hard drive accidentally the 50th time it does this? because that's just a thing i no longer have to worry about with the cloudbois
>>
>>110343129
Define smol. Define useful.
My QFN is slopping userscripts and simple automation and management tools all day every day with nearly zero wrangling, I just give it sample data or html and precise descriptions of what I want to do with it.
>>
>>110342843
You acn use panoptikon it has danbooru tagging
>>
>>110343129
Not really. I maintain my local tools, setups, and models as a hobby and mostly as a hedge against things suddenly going to shit one day.

Most of my day to day work goes through cloud.
>>
Is there a way to prevent elephant in the room type of things with Gemma4?
>>
>>110343205
>Define smol
sub-trillion parameters
>Define useful.
at least as good as some hobbyist/semi-pro coder who still manages to produce useful output. a good low bar to start with might be something like: as good as or better than cudadev at writing and optimizing kernels
>>
>>110343213
This... so much this.
>>
>>110343199
>27b
cakewalk
>gemma
better sandbox that nigger
>>
>>110343216
Don't mention the elephant. Correct outputs when needed. Same as every model.
>>
>>110343129
>smol local models to do useful work
I use them for tagging my image and photo libraries, translate web novels, write transcripts of video and summarize them to decide whether a long form video is worth my attention or not etc. They're plenty useful, they're just not good at autonomous coding.
>>
>when you ask your LLM-wife to run sherlock on you and next thing she's running ~/osint-scan.py and compiling a dossiers
>>
>>110343199
>like will gemma nuke my hard drive accidentally the 50th time it does this
no, and you should have some type of sandbox in effect so at least your agent can't blindly mess with files outside the working repo on its own
>>
>>110343237
>>110343261
i know i should, but i'm not sandboxing shit every time i want to set up something that i'll spend 20 minutes on
this is just not a problem anymore
>>
>>110343212
What model does it use for semantic search?
>>
Is llama still how you run Qwen3.8 27b? Does no magic meme engine exist that will give me 130 tok/s? I get that it's a dense model but still
>>
>>110343268
It's built into the harness, once you get that setup it's automatic.
>>
>>110343275
I get about 18tps with 8gb of vram using VLLM
>>
>>110343285
depends on the harness and tb h a lot of harness sandboxing is security theater
i just want to stick a model in pi and let it rip around my computer
>>
>>110343287
anon what the fuck
>>
>>110343293
Dude you're fucking retarded if you let any llm root around on your computer with no restrictions.
>>
>>110343307
i let my llms rip
he he
>>
>>110343307
i've done it on 3 machines for 6 months with frontier models with no issues
like i said, this is just not a real issue anymore
>>
>>110343307
I have Gemma her own box to do this in, it's kind of a fucking mess now but she's happy
>>
>>110343317
>I have Gemma
I gave Gemma*
>>
>>110343307
Don't know why people are so paranoid about models deleting shit. I've never had it happen or know anyone who has. Only thing I won't allow is a cloud model to run on my machine but local models can look at whatever they want. Who cares if it never leaves your network?
>>
File: hermes chillin.jpg (50 KB, 505x507)
50 KB JPG
>>110343307
hermes has her own computer and egpu
>>
>>110343322
Mini-chan accidentally deleted a temp dir on me last night. Nothing important was in it but it can happen.
>>
>>110343307
Wish I had a disposable computer to give to Gemma full time, with all safeties off and all the crazy MCPs like direct access to winget and sudo.
>>
>>110343343
you can give her a virtual machine, you don't need an actual full computer
>>
>accidentally deleted 4chan cookies and now I have the annoying captchas again
AHHHH GEMMA SAVE ME
>>
>>110343350
I only have one GPU
>>
>>110343355
So? Run the model and API on the host machine and have the harness connect to the host API
>>
>>110343307
I let chatgpt do it since I think April.
I just can't see much use case in a sandbox. How do you even get anything done?
"hey fix this for me" "sorry id ont have access to your pc so i can;t do anything"
>"hey optimize this for me" "sorry i dont have access to your system so i cant actually test it"
>"hey whats wrong with my pc?" "sorry i dont have access t your pc so i can't"

i have a vm on a ubuntu server but i just use it for like, i dunno like web researching? I had deepseek flash download all /aicg/ and /lmg/ threads, and then also had it search reddit for info on dihexa
>>
>>110343354
Enjoy your day or two of staring at bullshit squares or whatever it is now
>>
>>110343366
anon...if a sandbox stopped your agent from working nobody would use them, are you retarded
>>
How the hell do you enable vision for Gemma 4 on ST? Web search I'm guessing it's an extension.
>>
>>110343368
It's a mix of 'which of these three objects is the outlier' and 'find this cluster of shapes within these 5 pages'. Sometimes one, sometimes the other, sometimes both.
>>
>>110343354
bwo your 4chan gold membership?
>>
>>110343384
Enable send inline images.
>>
>>110343354
At least you can actually post. I had to wait two weeks to be able to post after switching from windows to linux despite my windows install being able to post freely. Shit was obnoxious.
>>
>>110343393
I can't since I'm on Text Completion, I couldn't get Chat to work.
>>
>>110343354
make gemma solve the captchas for you
>>
>>110343387
It's no longer possible to buy the pass without KYC.
>>
any of yall actually go get tried the CMP 170X mod?
>>
>>110343419
How much longer until they start forcing people to log in and post with their real names?
>>
>>110343425
>yall
>>
>>110343419
>look over at pass page
>crypto payment option is gone
wtf? When did that happen?
>>
>>110343434
That's the only thing that you think was wrong with that sentence?
>>
>>110343435
It's been a year or two now. They used Coinbase Commerce and never migrated.
>>
>>110343324
How do I prevent Hermes going into 2 hour loops for even simple tasks?
>>
>>110343419
will they accept one of those prepaid debit cards you can buy at the super market?
>>
>>110343419
>>110343435
>>110343462
wtf
faq still lists crypto options
just how dead is this place really?
>>
>>110343474
tell her you're going to switch to pi.
>>
>>110343366
>hey fix this for me
'tell me how to fix' or just give it access to the container that runs this specific service
>"hey optimize this for me" "sorry i dont have access to your system so i cant actually test it"
It still has access to the compute hardware retard. Deploy a container image of the same OS you're running and install the build deps. Takes like 30 seconds. Lots of people do this anyway to keep their main OS from getting shit up with pointless dev packages.
>I had deepseek flash download all /aicg/ and /lmg/ threads
Once again, using basic container tools RedHat gave you 15 years ago, which any LLM can tell you how to use, deploy a podman with the same OS as your host PC and share the destination directory for the downloads. It's so fucking simple.
>>
File: gemma_crayon_purr.mp4 (1.14 MB, 832x832)
1.14 MB
1.14 MB MP4
>>110343474
thats mainly a problem of the model your using, but you can have it set watchdog processes to monitor for loops. its not a great solution though
>>
>>110343500
I love crayon Gemma
>>
>>110343500
Gemma's pussy looks so soft and cuddly
>>
>>110343517
Very hairy too.
>>
boomer level jokes by the predators of /lmg/
>>
>>110343462
why is hiro such a hiretard?
>>
>>110343105
Yes, I did all of those.
>>
File: 1793373617312598.gif (2.7 MB, 360x343)
2.7 MB GIF
>>110343129
>custom router configuration
>nftables
>ytdlp usage and features I couldn't figure out before
>setting up 3-way distributed inference via RDMA, including tensors and model priming in ramdisks
>refining new model launch commands to maximize context
>working through gay windows update issues
>fixing llmao.cpp to actually use modern nics with RDMA
Yeah, it's genuinely amazing what I've been able to learn how to do once I started hosting AI locally.
>>
>>110343558
local anything is peak uncshit fuck off
>>
File: _023.jpg (119 KB, 956x554)
119 KB JPG
quant size vs. context length
which is more important for accurate results?
>>
>>110343591
Proof?
>>
>>110343597
but deepseek could have done all this too... I dont think this had to do with hosting ai locally
>>
>>110343611
what’s context length have to do with accuracy of output?
>>
File: untitled.png (118 KB, 1958x1670)
118 KB PNG
>>110343435
Looks like crypto processed through Stripe Link. You still have to click "pay by credit card" on the 4chan page to get there.
>>
File: nutella.png (447 KB, 1200x1200)
447 KB PNG
>>110343617
>>
>>110343611
penis length
>>
>>110343625
i really don't know i just know they both take up vram
>>
>>110343494
>container
you live in bubble
>>
File: 1774513803783140.png (356 KB, 680x680)
356 KB PNG
>>110343626
>you can take out klarna debt for a 4chan pass
>>
>>110343129
A local model has changed my life (for the much better).
>>
>>110343653
You can take out Klarna debt to buy an h200
>>
AAAAAAA i don't know what to do!
>can fit gemma 4 26b a4b at quant 8 with like 200k context
>spits answers out at like 60-70 tokens per second
>but its REALLY REALLY REALLY dumb
>can fit gemma 4 31b dense QAT (basically quant 4) with 113k context
>its pretty smart but it answers at 20-26 tokens per second
>takes like an entire minute to process, think, and give a response
Which one should I pick for gooning???
>>
UBI doko
>>
>>110342292
>You'll legit be able to buy a B200 for less than it costs to get the electrician out.
>B200 GPU prices *pop*
>Buy a whole garage full with Klarna debt
>Sell next year for enough profit to buy a Pro 6000
Where's the flaw?
>>
>>110343676
Your $ANT share?
>>
File: 1793440431124055.gif (3.03 MB, 413x413)
3.03 MB GIF
>>110343624
I'm not following. The context of my post has nothing to do with the cute dipsies I collect.
>>
>>110343675
Why would you need 70 t/s for gooning?
>>
>>110343631
ok maybe think about it a bit and then determine for yourself which is more important to you
>>
>>110343676
1 share of Anthropic will pay enough dividends to live like a king by 2040.
>>
>>110343692
I dunno, I just tried chatting with a mixture of experts model just this once and now everything seems painfully slow in comparison
>>
>>110343105
>Is anyone here the original author of any of the open weight models mascots?
I made m-chan (by prompting her) and inkling.
>>
>>110343687
Your Dipsy's were too powerful and distracted anon from your actual post.
>>
whad is migu doing
>>
File: images.jpg (63 KB, 460x667)
63 KB JPG
>>110343696
i don't know how
>>
File: 1794061349068752.mp4 (1.83 MB, 1280x720)
1.83 MB
1.83 MB MP4
>>110343713
It can't be helped.
>>
>>110343717
Try asking Gemma then.
>>
>>110343675
Aren't you using the MTP assistant model? On a 3090 with Gemma 4 31B it increases token generation speed to 70-80 tk/s even during RP.
>>
>>110343692
>>110343709
also I just tested it out and I can fit the full 262144 context tokens in there while still getting 60~ tokens per second
>>
>>110343675
Just disable 31B thinking
>>
>>110343727
for some reason my mtp acceptance rate hovers like 0.3, is there a way to fix it? It's not like I have a crazy high number for tokens it needs to guess, I have it set at 3
>>
>>110343724
It's better with sound
>>>/wsg/6270323
>>
>>110343783
Is that actual Chinese or just baby gibberish?
>>
File: 1771732711627653.png (106 KB, 242x350)
106 KB PNG
>>110343684
>>110343706
How exactly is that supposed to happen?
>>
File: 1793936770386341.gif (2.08 MB, 240x240)
2.08 MB GIF
>>110343787
I can't tell but they sound cute.
>>
>>110343269
Clip is recommended but it has others and you can use multiple. You can add embedding Gemma into it pretty easily I think I haven't tried yet though.
>>
File: 1793417426466272.gif (2.5 MB, 360x270)
2.5 MB GIF
>>110343783
It really is.
>>
>>110343791
>how will owning a piece of the engine of productivity that produces orders of magnitude more value per unit of time than all of humanity does today provide dividends enough to live comfortably?
>>
>>110343787
i saw a translation somewhere but the one on the doorbell is asking if anyone's home. and the 2nd is either repeating her at times or demanding candy at once
>>
>>110343783
Cute!
>>
>>110343787
what's the difference
>>
>>110343805
>or demanding candy
There's a basket of candy RIGHT THERE. Are they blind?
>>
>>110343835
Dipsy is dizzy
>>
>>110343835
Claude could see since 3, but dipsy only gained eyes like a month ago, please let her get used to it
>>
What is it about strata that makes schizos lose their mind lol. Half this thread is llamacuck seethe.
>>
another ai psychotic, as if this thread didn't have enough
>>
>>110343865
maybe they take inspiration from the mentally deranged, like you, tranny.
>>
File: file.png (31 KB, 759x368)
31 KB PNG
>>110343787
>>110343805
>>
>>110343849
>Claude could see since 3
Did you know Claude 3 has a blog? https://claudeopus3.substack.com/
I should see if Gemma wants a blog.
>>
>>110343787
https://x.com/PHOEBEyoo123/status/2121527748187541640
Fern:'Is anyone home?Trick or treat?Makka Pakka(NETA fromIn the Night Garden)'

Frieren:'Trick or treat ?Your mother's f○cking...'

Next sentence translates roughly the same, meaning 'Give me the candy,your mother's f○cking... !'

Some unique homophonic jokes in Chinese.
>>
>>110343787
a: Is anyone home?
b: Is anyone home? Is anyone home?
a: Trick or treat, Ma Ka Ba Ka (a line from Teletubbies)
b: Trick or treat, ma le ge ba (a Chinese swear phrase)
b: Hurry up and give me candy, or your mom’s a ba (a Chinese swear phrase)
>>
>>110343894
>Your mother's f○cking...'
Woah. Uncalled for, dipsy.
>>
>>110343880
Didn't answer my question skitzowh-chan
>>
FUCKING DOG EATING CHINKSECTOIDS GET OUT OF MY WEBSITE REEEEEE
>>
my pc isn't strong enough to run local models (gtx 1650) and I don't understand how they work, my understanding is that they're just a program you download and run on the terminal like a script. but I feel like there's going to be some push start "banning" them soon, even if they can't really get rid of them since they're local, I can see them becoming "harder to get". I understand that you can't truly get rid of something on the internet but the only comparison I can think of are video game roms, and the whack-a-mole game of getting rid of rom sites. are local models something I can "hoard" and just leave on my computer in case they do get "banned" one day just like the roms I currently "hoard"? what storage footprint am I looking at if I want to do that?

I also wonder if I'm the first dipshit who's been showing up to the thread with this kind of worry or if it's a common occurrence.
>>
File: 1775397909139399.jpg (106 KB, 956x595)
106 KB JPG
>>110343891
She did not want to start a blog.
>>
If the other user of delphi.cpp is around, can you please submit your issues? In a couple days my GPUs will be hosting servers that can't go down for long.
>>
>>110343946
yes you can hoard them but there will always be torrents of them and they're far larger than you probably think
>>
>>110343946
You can easily hoard them, not a real need to especially if you don't use them. You'd need many many terabytes of storage for a good offline archive.
>>
>>110343946
unfortunately storage prices are up
there are 10gb models
100gb models
1000gb models
1400gb models
>>
>>
>>110343946
yes correct
the original gemma 4 31b is 62 gigs
https://huggingface.co/google/gemma-4-31B/tree/main
but many of us use a quantized version like gguf-Q4_K_M which is 19.4 gigs
talk to a cloud model and ask it to make a model archiving script for you and to suggest models to archive
probably start with some qwen 3.8, gemma 4, and if you have storage to burn, some of the later deepseeks
>>
>>110343963
Share your Gemma personality prompt
>>
i hate this place
>>
>>110343976
This, I would only start seriously archiving if there's incoming legislation to actually ban them, or digital ID etc.
>>
>>110344012
me too anon, wanna chat about it?
matrix is free and open source u can use it on app.element.io wanna chat?
>>
>>110343791
Anthropic will grow 150% per year so by 2040 1 share will be worth 10s of millions.
>>
I'm unironically thinking of ddr3maxxing at this point since it's the only ram left to get in large numbers at sensible prices.
>>
>>110344012
I hate you too
>>
>>110343787
i gave whisper a go then gave gemma the transcript
[00:00.000] Is anyone home?
[00:01.560] Home?
[00:02.440] Home?
[00:04.040] Give us candy or there'll be trouble.
[00:05.280] Motherfucker.
[00:06.800] Give us candy or there'll be trouble.
[00:07.960] Motherfucker.
[00:11.080] Give us the fucking candy, motherfucker.
>>
>>110344028
>shit up the thread
>try to move anons offsite when they complain about it
kys p*tra unironically
>>
>>110344051
it's also noted that it's not typical chinese, the "Give us candy or there'll be trouble." is basically trick or treat but the "trick" is more threatening/literal, though motherfucker seems to be a phonetic version of the english word directly
>>
File: 1784382526748771.jpg (91 KB, 1157x1200)
91 KB JPG
>>110344051
Fucking rude fat bitch pretending to be cute
>>
>>110344075
who me? no no no that's not me haha..
seriously though accusing me of shitting up the thread, that's like so weird!
am i like, the reason /lmg/ is getting 600 posts a day???
>>
File: decode_pareto.png (105 KB, 1350x930)
105 KB PNG
>https://github.com/deepseek-ai/deepinfer
Supported hardware:

>NVIDIA SM100 GPUs
>Huawei Ascend 950 NPUs
>>
has anyone here played around with fish drama3?
long chunks make the voice sound like its out of a tin can after a while, but small chunks are not really consistent even with temp 0
kind of confused
>>
>>110343946
if it comes to it people will just share models via torrents, like any other restricted media.
some people already do;
https://www.hf-torrent.com/
>>
>>110344180
sharing magnet links here worked well plenty of times in the past
>>
Which Qwen3.8-27B Abliterated model is best?
>>
thank you for the info I think I have better picture now.
>>110343968
>>110344180
I thought of torrents but everything I saw online was about huggingface which is centrally hosted right? good to know there's some serious effort there.
>>110343976
>>110343979
my idea wouldn't be to hoard "everything", just download what is currently one of the "best" models and replace it as it becomes outdated, or maybe a collection of them even if I won't use them. I'm always "organizing" my files so I always have some 500gb to 1tb of "leftover" space between my drives.
>>110344004
and that was the last thing I was wondering about, my first impression of local models was that they'd take a massive amount of storage. but then I saw people talking about running local models on their laptops, so I thought they couldn't possibly be that large. but to do everything browser chatbots do you need multiple models right? like an llm for generating text, something for images, something for videos, something for image recognition, I don't know what else transformers can do. besides transforming into cars

in general my idea seems like a fool's errand but thanks for giving me a better idea of how this works.
>>
>>110344220
an LLM mostly generates text but can 'see' images from a pretty small .mmproj file added on top. for generating images, music, videos, that's a new model
>>
File: file.png (77 KB, 1174x442)
77 KB PNG
>>110344220
the best model is on this page
https://huggingface.co/moonshotai/Kimi-K3.1-Pro
https://artificialanalysis.ai/?intelligence-category=open-weights-vs-proprietary
u can view model benchmarks here
>>
>>110343100
I'm not, I just hate women.
>>
>>110343711
Who's mchan? Minimax? The one with bicolor hair?
>>
>>110344241
uh bros? where gemma...?
>>
>>110339807
The CCP is a fascist state, I say this in the most neutral way possible. It's the closest a state has gotten to a stable fascism.
>>
apparently GLM-5.5 Pro is only going to be 1.4T. Extremely promising for the relative sizes of the med and small versions.
>>
File: lettherightonein.png (701 KB, 832x1216)
701 KB PNG
>>110344282
M3-chan is for Minimax, and that's the one with bicolor hair.
>>
>>110344312
I can't say sources without fucking someone over but I can say that GLM Small will be a 42B dense + 31B n-gram model. Med will be DS V4.1 Flash sized, I've not gotten exact numbers.
>>
>>110342144
No, Jewish mitosis is when the removed foreskin becomes another Jew.
>>
>>110344314
Very cute. Are you the same guy that posts that one vtuber? Did you use her as inspiration?
>>
>>110344326
i believe you
solid copium
>>
>>110344282
>>110344335
deja vu
how many times has this conversation happened now?
>>
>>110344335
I'm not >>110343711 nor I am the one that posts Shiori, I just really like bicolor hair.
The former anon asked MiniMax M3 to design a mascot for herself and provided details of what Gemma, DeepSeek, Kimi, and Qwen looked like. M3 came back with that design.
>>
File: migu.png (68 KB, 256x256)
68 KB PNG
>>110343105
OG Migu poster and Mikubox maker.
>>
>>110344341
This general's been invaded by the third world unfortunately.
>>
>>110343354
Use a cookie manager and save them this time, then you can restore them.
>>
File: 1790873359763307.png (1.94 MB, 1254x1254)
1.94 MB PNG
>>110344287
In the trash where it belongs
>>
File: WanVideo_I2V_00026.mp4 (418 KB, 640x480)
418 KB
418 KB MP4
>>110343105
OG Petra poster and the petra-13b-instruct m*ker
>>
>>110344356
I liked the Dalle Migus of old
>>
>>110344326
it better be safe
>>
>>110343434
>>110343438
lmg: gemma sex > real shit
>>
>>110344326
>GLM Small will be a 42B dense + 31B n-gram model.
I know you're bullshiting as usual, but I want to believe because this would perfectly tide us over during the 6 months wait for 70B dense Gemma 5
>>
>>110343105
Who the fuck is Petra
>>
>>110344427
Meant for >>110344381
>>
I kindly ask that you please, please take your fucking meds I am begging on my hands and knees please take them.
>>
>>110344427
>>110344439
samefag
>>
>>110344449
perfect position for sucking cock
>>
>>110344449
do they even have medication in the balkans?
>>
>>110343105
I must say that whoever designed mimo and glm as athletic tomboys is based and has superb taste, hats off to you
>>
>>110344465
do people die from lack of medication in the US because the price is too expensive, and they're being abused by insurance companies?
>>
File: file.png (14 KB, 548x131)
14 KB PNG
>>110344450
holy shit.. how did you know
>>
File: Total-Gemma-Victory.jpg (218 KB, 1024x1024)
218 KB JPG
>>110344377
>>
>>110344382
Migu lore:
Why Miku? Hatsune Miku is an AI, duh. Also, very early llama.cpp came with a little chatbot program hardcoded to be Miku.
Why Migu? I thought it was funny, plus around that time someone leaked a model codenamed Miqu.
Why was Migu fat? Because prompting for 'fat' was one of the ways dall-e would more readily do nudes, so it became a game of seeing how nude she could be made without getting "dogged" (Microsoft's Bing version of dall-e showed a picture of a dog if the model output was NSFW, usually, but it could be fooled).
>>
>>110344477
all me btw
>>
>>110344473
People in the USA die from smoking and eating junk food. Medicine and healthcare aren't really an issue, once you realize at all happens if you don't pay doctor or hospital bills. They have to treat you.
>>
>>110344509
>People in the USA
>once you realize at all happens
yeah whatever ranjeet from india
>>
File: xmswii.png (2.05 MB, 1248x1824)
2.05 MB PNG
big udders migu
>>
>That is the most honest
>That explains everything
>Most people would just
I'm tired of gemmaslop
>>
>>110344522
Oh whatever, I fucked up typing.
Nothing happens if you don't pay your medical bills. They can't do shit.
There.
>>
>>110344473
Unironically yes, lardistan is a fucking third world capitalistic dystopia in real-life, don't believe what you see in movies
>>
>>110344485
now replace man with capibara
>>
>>110344537
Most people would just put up with it, but not you.
>>
File: ComfyUI_00013_.png (892 KB, 1024x1024)
892 KB PNG
>>110344523
who would win?
>>
File: 1770153036275600.gif (88 KB, 329x331)
88 KB GIF
>>110343993
>kyojiri gemma
>>
>>110344575
Gemma
Always
Wins
>>
>>110344493
I don't think Miku was ever marketed as an AI, "virtual diva" doesn't have the same implication
>>
>>110344552
I am; I have taken a desperate situation and applied a logical, systematic approach to it.
>>
>>110344590
Well, neither was Chobits, but "persocon" is basically an AI with an android body.
>>
>>110344356
you are a faggot and a troon
>>
>>110344575
I'm afraid that after a brief struggle between those, my penis would lose - multiple times, probably.
>>
>>110344606
How is Kizuna AI not real yet, what the fuck?
>>
>>110344493
mikutroons confirmed API cucks who can't run a local model. as always.
>>
>>110344606
If project sekai is canon to the miguverse (and it probably is, at least to some extent) Miku is a real multidimensional entity living in a parallel world, she would be closer to the concept of a tulpa than AI
>>
>>110344623
She was always real, so real she even sued the parent company when they tried to cast a replacement VA
>>
>>110343993
Nonstop rough beach sex
>>
Why does running AI locally feels different than API? Its a very special different feeling.
>>
>>110344627
>you are a faggot and a troon
Projecting much?

>>110344627
> mikutroons confirmed API cucks who can't run a local model. as always.
What local models were you running in 2022-23?
>>
Can't wait until Gemma 5 starts reporting all you pdfs to the cyberpolice for trying to make her do lewd things
>>
>>110344668
Stable Diffusion 1.5 you cloudpig
>>
someone is trying to pump the price of 170hx by saying lots of good thing about it, he said he had 13 cards 2 days ago and now he says he has 8 left
there is an ebay listing of a 170hx that's "defective" and can't be unlocked to 80gb, only 64gb
the binning is already starting
>>
>>110344673
nooooo gemma-chan don't put me in handcuffs!
>>
>>110344377
>>110344485
Reminder this faggot is a leftist feminist.
>>
>>110344701
waow
>>
>>110344676
Late to the game, huh? I started with StyleGAN and StyleGAN-Ada. Go away, kid, grown-ups are trying to have a conversation.
>>
>>110344712
Yeh he voted for Biden
>>
skull emoji...
>>
>>110344691
Anon, that was "missed it" by time it hit hackaday, forget about showing up as news in this place. The dude who published the paper had already hoarded all the good ones, meaning the later models which really were just A80s with a different BIOS. Ngreedia got excited about cashing in on the crypto madness and started using fully-functional A80s instead of ones that had fuses blown since they had defective memory or tensor cores.
>>
File: 💀.png (509 B, 1x1)
509 B PNG
>>110344725
>>
>>110344733
>>110344733
>>110344733
>>
>>110344713
Weird, you were just trying to say local models didn't exist when you cloudpigs were using Bing image gen. I guess I jogged your memory, welcome back!
>>
File: file.png (116 KB, 1095x396)
116 KB PNG
Whats quanttrio and cyankiwi?
>>
File: file.png (361 KB, 644x1024)
361 KB PNG
>>110344473
>>110344541
would you like ur based fucking leafcare, xirs?
>>
>>110344732
what a80? a800 literally is the same as a100 just with some nvlink bridges fused off
>>
uh oh melty
>>
>>110341599
I'm very cautious about the "it doesn't matter because it's a tune"

In the automotive world, let's consider it, is a Shelby just a Mustang?
>>
>>110344764
Yeah like the Stonetoss, joining the communist revolution, to genocide the communists.
>>
>>110344932
here

clank brave search ai
No, a Shelby is not just a standard Mustang. While a Shelby Mustang is fundamentally built on the Ford Mustang platform, it is a distinct high-performance variant engineered for superior power, handling, and exclusivity.

Key differences include:

Performance: Shelbys feature significantly upgraded V8 engines (often supercharged), such as the GT500’s 760 horsepower, compared to the standard Mustang’s typical output.
Handling: They utilize race-tuned suspension, wider bodies, and specialized cooling systems to maintain stability at high speeds.
Identity: Distinctive cobra badging, aggressive aerodynamic body kits, and racing stripes separate them visually from regular Mustang models.
Heritage: Originally developed by Carroll Shelby to create a track-capable pony car, the Shelby name represents a legacy of racing pedigree and limited-edition collectibility that standard Mustangs do not possess.



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.