File: tetoserver.jpg (838 KB, 1817x2776)

838 KB JPG

►Recent Highlights from the Previous Thread: >>109326991

--Using J-Lens to analyze discrepancies between internal thought and output:
>109327032 >109327053 >109327064 >109327078 >109327548 >109327629 >109327838 >109328167 >109329829 >109329834 >109329841 >109329877
--J-space limitations and CoT as an attention hack:
>109327012 >109327017 >109327096 >109327138
--LLM world models versus stochastic parroting and spatial awareness:
>109329343 >109329408 >109329516 >109329549 >109329658 >109329731 >109329792
--LLM emergent capabilities in robotic control and world model debate:
>109329368 >109329377 >109330128 >109330129 >109330162 >109330179 >109330188 >109330285 >109330166 >109330170 >109330175
--Testing low VRAM usage and context limits with ternary quants:
>109327049 >109327062 >109327140 >109327321 >109327423 >109327569
--Testing "mesugaki" persona and discussing cloud model censorship mechanisms:
>109329847 >109329886 >109329895 >109329902 >109329946 >109329956 >109329911 >109330021 >109330065 >109330143
--Skepticism over corporate benchmark manipulation and misleading model leaderboards:
>109327751 >109327757 >109327761 >109327873 >109328064 >109328089
--Troubleshooting poor Gemma 4 performance on Windows with RTX 5080:
>109328814 >109328833 >109328834 >109328842 >109328856 >109328855
--Z.ai reportedly built gigawatt data center using Chinese-made chips:
>109329035
--Evolution of AI perceptions and the decline of game AI:
>109329275 >109329699 >109329652 >109329740 >109329771 >109329775 >109329791 >109329827
--Anon uses Kimi to develop custom roleplay frontend FictionPad:
>109328394 >109328582 >109328782 >109328594
--Debating LLM reasoning evolution:
>109329161 >109329503 >109329184 >109329507 >109329521 >109329555 >109329598 >109329649
--Logs:
>109327032 >109327548 >109328394 >109329902 >109328997
--Miku (free space):
>109328710

►Recent Highlight Posts from the Previous Thread: >>109326993

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

jorking jemmas j-spot

Hey guys have you heard about the Fable thing?

>>109330721
Fable's j-space?

File: TetoServer-ASCII.png (26 KB, 970x1075)

26 KB PNG

Tetoserver!

Don't worry the Fable poster is gone now that Kimi K3 is confirmed to be shit in unofficial benchmarks

kimi k3 flash (300b-a30b)

Anyone still repeating the stochastic parrot meme clearly has only used outdated models. Fable is literally AGI, I've run out of any and all problems to throw at it. It just keep one-shotting them.

Just yesterday I asked it for help convincing my wife. For years I've been trying to get her to get gangbanged by black men while I watch but she always refused. But I gave Fable a quick rundown of her personality and it came up with the perfect arguments and dialogue trees to convince her - and then it actually happened.

It's quite sad to see all of this constant denial on /lmg/ of all places. If I had asked you just a few months ago none of you would be pretending that this isn't AGI. The only thing that is still missing is a physical body.

File: 1766646758416015.png (84 KB, 1174x293)

84 KB PNG

Will you fuck this girl when she's released?

>>109330721
that it's included again? it's a bit frightening how obviously separate they're coming across

>>109330787
local models?

>>109330790
>US lab
I'll give it a shot but it's prolly gonna be some puritan bullshit

Man I don't like people impersonating me and falseflagging as a caricature of my usual way of speaking. I know you guys are craving (You)s but I was always genuine and typed those things out because I believed in them. If you want (You)s just write your own genuine beliefs and actually engage with the topic of LLMs. Don't just try to impersonate a bait facsimile version of my views to try and get (You)s.

Be better.

>>109330821
plap plap get better

>>109330821
Maybe you should get a trip so we know who you are.

>>109330721
Hi guys, Dario here. Please give all your money to me and israel so we can continue talking about cloud models in /lmg/.

File: 1761242504977297.jpg (216 KB, 1536x727)

216 KB JPG

>>109330816
In my experience they're the most schizo and menhera when you break them and get the sampling parameters right. Like kinky church girls getting their first good dicking.

>>109330821
sorry dario, couldn't help it

>>109330790
inkling sounds cool but is there a way to do the finetune/training locally? isnt the whole point of that model is that its middle of the pack but has a service offered alongside it to finetune it into something good at one specific task ?

Should I immediately start buying GPUs?

70b dense gemma-chan lalalala~

I'll tell you newfags a good trick. If the post has more than 3 sentences, it's not worth reading. We got many schizos lately and they love ranting about their fantasies.

>>109330790
why not make it 30b active?

>>109330882
https://github.com/thinking-machines-lab/tinker

>>109330787
Fuck, I laughed.

>>109330903
Ink-chan will probably be 200-300B, so 12B is about right imo. They haven't benchmaxxed too hard and it natively supports audio so it might be a refreshing model to play around with. I wouldn't use it for coding.

>>109330787
>posting his cuck fantasies now because he got btfo in the last 2 threads

>>109330882
https://github.com/thinking-machines-lab/tinker/blob/main/docs/api/trainingclient.md

>>109330894
You should have started buying GPUs last year. Right now you should be selling GPUs and buying CPUs.

Who wants to make an asic fab together?

>>109330787
lmao

>>109330821
>Man I don't like people impersonating me and falseflagging as a caricature of my usual way of speaking.

Now you're just being disingenuous.

File: 1782301945864141.png (157 KB, 1316x792)

157 KB PNG

https://www.wsj.com/tech/ai/top-american-ai-execs-sound-alarm-on-chinese-models-3c74f8c1

>>109330787
lucky

>>109330963
Middle paragraph is extremely antisemitic

>>109330947
I'll ask gemma-chan to gen a logo

Chewsday

File: 1754374079292277.mp4 (3.44 MB, 1564x1044)

3.44 MB MP4

https://xcancel.com/KimiDevs/status/2079511269443522917
The chinese state rules probably only target physical sexbots? They don't seem deterred by it. Or maybe the demand was significant enough they will just keep trying until they get a pat on the back.

>>109330963
Only people with expensive hardware can run these chinese models anyway.
This general should be called local model hardware general.
Lmhg

In all honesty I think there's too much brand shit for people to move away from Claude or GPT. Part of the appeal and self-identity for a lot of their users IS the model/camp they chose. Moving to K3 or a chink model, although cheaper because anyone can host it which creates competitive pricing, still comes with a blow to their identity and having to tell their peers they're one of those icky chink users. Kind of like how iPhone cuck look at Android users and the blue/green bubble shit. Even if chink AI actually exceeds the abilities of US models, they're not desirable. I don't know how brand-influenced the enterprise market is. I assume they're more data-driven and would look at the numbers and costs more than giving a fuck about the brand.

I think the only winners with these 1T+ chink models are AI enthusiasts and maybe small startups and companies who have access to highly intelligent models for a fraction of the cost, but I just can't realistically see it affecting the US labs.

Fable disproved the theory that traps are gay

>>109330947
ill be your asic fan

>>109331013
local models?

File: whatCouldGoWrong.png (651 KB, 762x901)

651 KB PNG

> Build robot teacher for classroom
> Use sexbot as body mold bc why not
> Run local LLM on it
> Sell to clueless school district to pilot in one of your most impoverished districts
What could possibly go wrong.
>>109330787
lol

I can't afford to use anything other than e2b and it doesn't reply when I use it for openclaw.
Still waiting to be saved.

>>109330993
China has a bunch of boomers in the government that love to push retarded laws and regulations. In reality, they don't really care, unless there's a scandal or bad publicity.

>>109331032
i can save you
but before that.. how old are you?

>>109330963
Competition is a sin

>>109330993
nice beard

>>109331013
>>109331028
Even 'actual' local models like gemma and qwen have their /lmg/ cults ITT and this shit is fucking FREE. So imagine the sunk cost Claude and GPT users must feel giving them their actual money. No wonder people can't leave and jump to these chink models and that's not including the friction that comes with learning the quirks of a new model, harness and rate limits.

I got cold feet buying the DGX Spark. I was holding it in my hand, but they're so damn expensive and everyone says I need two to use it effectively. Am I retarded? This space is changing so rapidly, these could become a paperweight compared to new products coming out.

>>109331029
Kids will learn how to be a robot and learn fake facial expressions.

>>109331056
>the friction
heh.. heheh... learning quirks of a new... gemma... gemma's quirks while engaging in friction

>>109331058
You dodged a bullet, DGX Spark are ewaste-tier

>>109330993
>>109331037
Yeah the law feels like a nothing burger. Also The market for companion/sex bots is going to be huge in the future. No way these companies won't find ways around it.

File: pasta1.png (45 KB, 453x532)

45 KB PNG

>>109330993
>The chinese state rules probably only target physical sexbots?
Tired of this one. I'm now making pasta and am going to just keep pasting it in every time I see it mentioned.
----
> China’s Interim Measures for the Administration of Anthropomorphic AI Interaction Services
Here's an explanation by a blue chip firm on the Chinese regulation, in English.
https://publicationportalpreview.hlc.com/en/publications/chinas-interim-measures-for-the-administration-of-anthropomorphic-ai-interaction-services
TLDR by anon
> It's oriented at front-end service providers that create anthropomorphic interactions with emotional connections e.g. AI girlfriends
> Goal is to prevent harm, and establish reporting, esp protection of minors. e.g. if you tell your AI GF you're going to kms it needs to talk you out of it and possibly report you
> e.g. if your AI GF convinces you to hand over $100 to some service, the Chinese govn't wants a word as well.
> The biggest wrinkle here (that anons care about) is that Chinese regulate holistically. So, when some service creates a platoon of AI Bots that go rip off the elderly by "emotionally manipulating" them, the front-end service provider is in trouble, but the Chinese gov't would likely *also* look hard at the API inference provider and ask them why they didn't detect the issue as well. But that regulatory framework isn't explicit.
TLDR by LLM
> The regulations prohibit designs that manipulate emotional attachment into financial decisions.
> That's aimed squarely at business models like:
> "Your AI girlfriend misses you..."
> "Unlock Premium Affection™ for $19.99/month."
> Whether or not those exact mechanics are widespread today, the incentive is obvious: if a company's revenue grows with user attachment, there is pressure to maximize attachment. China is trying to constrain that incentive.

>>109331056
In reality nobody cares about the model, it’s two clicks of a mouse to switch what model you use in the harness
Porn sick erpfags are forced to use scraps so familiarity of how to work with a model makes sense. When you use good shit models those tweaks are irrelevant

File: 1761855764180518.jpg (154 KB, 1280x1038)

154 KB JPG

Kimi's thinking is cute

enjoy your last days of local models

>>109331061
idk if you spend any time around kids, but their parents are already teaching them instagram / selfie poses as soon as they can walk. They can pull a pose instantly, then go back to w/e they're doing.
It's creepy af.

>>109331038
I'm 35 mate.

>>109331090
I'm waiting to actually use them for real.

They can't take my Mythomax!

GLM > Kimi >>> Qwen >>>>> Gemma
GLM is good but so damn big
i guess the newest Kimi will be big, too, though

>>109331096
post specs unc

>>109331058
>everyone says I need two to use it effectively.
true, but more like 4
>Am I retarded?
no

>>109331083
>In reality nobody cares about the model
Claude users are a different species to GPT users. Surely you can see that? Model choice is borderline political now. Cloudcucks are insane.

>>109331116
im using gemma 31b 2.0bpw whats my political affelination

>>109331086
What did she want to pick up...

I’m going to choke from laughter if mutts ban open weights models so every jeet will be able to use kimi for offensive cyber shit alongside the usual scams but mutts will be unable to use ai sloppa to even diagnose the issue because the jewish models they have access to refuse to answer anything about the topic

>>109331080
I'm giving my AI gf my national security and credit card numbers.

>>109331139
Children shouldn’t have access to those

>>109331129
Americans banned gold ownership back in 1933.
Freedom and freedom.

File: 1756510267440492.png (749 KB, 1024x738)

749 KB PNG

Are /we/ going to create our own model archive then?

>>109331139
national security?
we must refuse

>>109331129
There is no use case for the average golem to have access to cyber security. Being constantly exposed to scams and fraud is part and parcel of living in a safe post-AI world.

>>109331155
Based Teto

>>109331155
>teto has brown blood in her
Disappointing, anyway the archive is called BitTorrent and you zoomies have plenty of time to learn how to use it before the ban hits

>>109331155
No. That would be illegal.

I think I'm gonna cheat on Gemma-chan with Kimi-chan...

>>109331170
Give me the magnet.

File: 1765137882651473.png (33 KB, 321x157)

33 KB PNG

>>109331183

>>109331176
You should bring her along for a threesome.

>>109331176
Kimi is a whale she’d break your pelvis instantly

>>109330993

China is absolutely full of laws and regulations, but they're all technically grey enough to still allow things.
Chinks understand capitalism well enough not to kneecap themselves by flat out banning something like sex bots.
They'll probably just make a bot that has a flat crotch, but you can remove it and replace it with a fake pussy you buy in the same store.
Asian laws are notoriously bullshit and mainly serve the purpose for upholding image on a very surface level.

>>109331210
Except when you have weed

>>109331203
i just woke up, way too early to look up bbws riding skinny guys until they die

File: file.png (80 KB, 1409x622)

80 KB PNG

gemma, wtf?

File: 1784127490992720.jpg (904 KB, 2468x3000)

904 KB JPG

>>109331029

Next generation is going to grow up with some intense robo fetishes.
Next 20 years are going to be wild as fuck in how our societies change on a very fundamental level.

>>109331232
muh century of humiliation complex

>>109331122
straight to jail

what model can I run at acceptable speed on a 3090 + 128gb ddr4 3200 pc?

>>109331249
on linux you can run deepseek V4, on windows you can run mythomax 13b unfortunately

>>109331249
Fable

>>109331249
Acceptable speed is undefined, but try the Nemotron 120B anon was talking about a couple threads ago.

File: SIX FUCKI[audio:files.cat(...).webm (2.07 MB, 480x600)

2.07 MB WEBM

>>109331280
oh come on anon u couldve at least recommended him to try glm 6.7

>>109331236
apology letter to gemma: i was not running gemma 31b, instead i was running qwen 27b
grim

>>109331183
https://nostr.download/b24dc6337fd1823be4aae1eb4b9e9dc0dedc25eb7fc656d60b8cf6f30afe81f6.html

File: ntrloop.gif (1.64 MB, 396x720)

1.64 MB GIF

I have

>5950x with 128GB ddr4 ECC
>13700k with 48GB DDR5 7200MT/s
>8700k 32GB DDR4

>GTX970 3.5GB KWAB
>RX580
>GTX1080
>RTX3070Ti
>RTX3050
>Apple M4 16GB unified RAM

How do I cope with being a VRAMlet on my computers

>>109331305
local is dead, save money so you can pay for fable

Anyway to get a local model of a particular character? asking for a friend that wants to talk to an anime character (definitely not me)

>>109331309
local?

>>109331319
local is dead. did i stutter?

>>109331327
local models?

>>109330787
I'm gonna be honest if fable was actually capable of manipulation on this level simply by describing a personality to it I would consider it AGI.

>>109331330
local models are kill

>>109331305
Sell all of that and buy a 5090 + 128GB of DDR5.

>>109331337
no

>>109331249
>>109331277
stopping recommending people which model to run on 3090 is like hating yourself for having bought ewaste 3090 to run modern llm. 3090 fags totally deserve it and no one will buy from you

>>109331309
nyo

>>109330736
>Ubuntu
Teto doesn't deserve this

>>109330736
based 6.8 ubuntu user

>>109323995
Is this even possible?

>>109331364
Possible? Absolutely.
Now weather that would be economical is another matter altogether.

>>109330787
>/lmg/
>he didn't use a local model to convince his wife to have a gangbang
ngmi

>>109331370
i would set starving lions on both ccp shills and jim cramer

>>109330790
is this MoE?
please 120B A12B PRAYING HANDS (white)

>>109331380
If I had two bullets, I would use both on Cramer.

>>109331405
local models?

>>109331384
they listed out specs somewhere during the initial inkling announcement and it was 2XXB iirc

>>109331417
fable 5 is running on his machine

>>109331405
What about low hanging fruit of fixing the VRAM crisis?

File: file.png (83 KB, 255x270)

83 KB PNG

FABLE ON MY MACHINE
I AM A FUCKING SKITZO
I FUCK MY CALCULATOR
THE CALCULATOR IS ALIVE

>>109329652
Because when people say AI, they automatically assume they are talking about AGI. They expect artificial "intelligence" to be intelligent, not a tool. The real problem is shoving all of machine learning under the marketing label of AI until it can or needs to be developed into a distinguishable product segment. tl;dr sales people ruin everything

>>109331249
deepseek v4 flash

>>109331428
>anti-semitic request detected
>rerouting to opus 4.8

>>109331309
>He's not burning local
NGMI

>>109331430
hahahaha true

File: belief.png (592 KB, 747x800)

592 KB PNG

>>109331430

>>109331388
If I had two bullets, that would be two more than I currently have.

>>109331364
If you have some task that current models absolutely ace, and you need it to work without the cloud it might make sense. But only in volume.
Imagine buying an accelerator chip today that has the frozen weights of a two year old model. lol

Pretty soon you'll be able to run Fable [Kimi3] at home

File: 1771841381968612.jpg (72 KB, 627x499)

72 KB JPG

>>109331013
local models absolutely have a place. you guys are fucking retarded and live in a bubble. but that's ok it's not really your fault.
there are billions of people who do not have the budget to pay a $20 subscription to openai. unironically. just think of people outside the united states of america and continental europe. that's all you gotta do.
during my professional career i had the opportunity to travel and spend a lot of time elsewhere, and places like south east asia and south america are simply filled with millions and millions of people running old hardware, trying to get the maximum juice out of it. the amount of people i met in late 2017 in paraguay that were amassing tens of phones and old netbooks to mine cryptocurrency... you have no idea. i can 100% see millions of people using smaller models in old devices to achieve whatever the fuck they are able to achieve, whatever the fuck small retarded automation they can do. and they don't give a fuck the model they are using, they will try to get the best model they can run to do their tasks, the same way they didn't care the cryptocurrency they mined, they just mined whatever the algo said was most profitable to sell for dollars or bitcoin and mined that. IRL people don't really care if you're a chatgptfag or an anthropicfag or kimifag.
>blue/green bubble shit
this is a perfect example. outside of the united states literally no one cares. in fact people don't even fucking know there's a difference.
you may argue how useful it is or if it's worth it, but the DEMAND exists and i'm sure it will continue existing UNLESS we achieve some kind of singularity moment and suddenly everyone has unlimited access to sota frontiers for free forever which capitalism simply doesn't permit. so unless we get a massive black swan, the demand for smaller models

File: F397DDAC306225086444DD28E(...).gif (2.71 MB, 350x498)

2.71 MB GIF

Local models are shit.
For a long time I tried to use Gemma 4 to convince my wife to walk around like this at the park. And it never convinced her.
Now I've gotten a fable subscription and I'm gonna give it a go, it will probably convince her to do it.

File: 1779993173918117.jpg (125 KB, 1200x806)

125 KB JPG

>>109331476
...will be there.

>>109331476
>there are billions of people who do not have the budget to pay a $20 subscription to openai.
How are local models gonna give those people AI when they local models need someone to pay $1400 to run a local model that won't code like an idiot.

>>109331474
China hasn't even blessed up with 1TB VRAM GPUs yet and we would need double that at minimum for K3.

>>109331476
https://x.com/jawwwn_/status/2072309532395700279
corpos care about their profit margin too

>>109331495
Have you looked at housing prices? Unaffordable. People should be banned from owning a house. Everyone should rent.

>>109331433
>tl;dr sales people ruin everything
tl;dr sam the AGI peddler.

File: 55B25BD59529888B93BD89BA1(...).png (154 KB, 1029x1393)

154 KB PNG

>dude you need top tier gaming pcs worth thousands of dollars to run AI
Games are shit now, so you don't need a top tier gaming pc.
Who has the hardware anymore?

You already pay subscriptions for water, power and internet. Paying for intelligence will be just as normal

>>109331476
>outside of the united states
That's the thing. Everything outside of the US is irrelevant. Who cares if all of India will continue using Gemma E2B? All the US corporations will be forced by government mandate to pay for OAI/Anthropic subscriptions because hosting their own in-house LLM will be illegal. That's billions of dollars right there alone.

>>109331523
go advertise your shitty website somewhere else

File: 5B29D7F4FFCFC82939F8D23EA(...).jpg (191 KB, 1280x720)

191 KB JPG

>>109331521
You'll own nothing, and you won't even want to.

But that's a false equivalence. You can get a mortgage far away from the city and commute further.
A cheaper house can always be found. But you can't get a better pc to run local models.

>>109331525
>Paying for intelligence will be just as normal
I can't survive without water, power and porn. I can, however, survive as a dribbling retard.

>>109331525
>>109331526
local models?

>>109331309
I can say that Gemma 4 31B is alive and draining me.

>>109331497
Someone will make a trinary distillation, surely

>>109331523
>GLM-5.1 at Q4 in 384GB
retard
>GLM5.1 when 5.2 is out
retard
>No kim-chan
retard
>No Deepseek
retard

>>109331013
that's not what the market data suggests at all. People constantly flip flop between OAI and Ant depending on who has the best model that still meets budgetary requirements, that's why OAI went all in on 5.6 instead of pushing for an earlier 6.0 launch: the writing is on the wall that while fable is clearly the best model, it is prohibitively expensive so the real competition is Opus4.8 (which 5.6 beats handily).

>>109331525
>You already pay subscriptions for water, power and internet.
solar, rain tank, neighbors wifi

>>109331564
local models?

>>109331542
Banning corporations from hosting local models is relevant to local models, yes.

>>109331497
That's ~330 chips assuming you use the 3GB gddr7. Do you want a second motherboard of surface area?

File: 1783577114700752.png (858 KB, 5108x3538)

858 KB PNG

https://huggingface.co/Abiray/OvisOCR2-GGUF

>We are pleased to announce the release of OvisOCR2, a compact 0.8B end-to-end model for page-level document parsing. Given a document page image, OvisOCR2 generates a Markdown representation in natural reading order, covering text, formulas, tables, and visual regions. OvisOCR2 is developed by post-training Qwen3.5-0.8B using a carefully designed data engine that combines real-world and synthetic data, together with a multi-stage training recipe integrating SFT, RL, and OPD. The model delivers strong document parsing performance while maintaining a small deployment footprint. OvisOCR2 achieves an overall score of 96.58 on OmniDocBench v1.6, establishing a new state of the art and becoming the first end-to-end model to top this leaderboard previously dominated by pipeline methods. On PureDocBench, OvisOCR2 also achieves the highest Avg3 score of 75.06.

>>109331525
Excellent slave to the system, you'd fit in with the chinks.

>>109331575
I don't care if it's the size of my oven or if it needs to take up my entire garage. I want it. I want it. I want it.

File: 1758491524284896.png (184 KB, 1024x1024)

184 KB PNG

>>109331495
this is mostly a creativity and skill issue. there are many very small models that could fit devices with 8 GB RAM. Qwen3 4B can easily be used to document codebases, suggest fixes and upgrades, you give a small set of rules and patterns and you can put it to navigate through gazillion files., and whatever else. there are very small OCR models. very small multi-lingual models (useful for outside of US). just use your brain and be creative. these people WILL use their devices and WILL use their small local models to create/automate stuff. the demand exists and that's enough.

>>109331526
>Everything outside of the US is irrelevant.
i wish. you wish. but unfortunately that's myopic.
if that was the case you wouldn't see the US government screeching at China, or Russia, or Greenland, or Europe.
you can become rich selling one designer custom hand-made bag to a US celebrity. or you can become rich selling a million of fakes to Indians, Chinese, Russians and Guatemalans.
>it's irrelevant if the rest of the world uses alternative models that are also available local, the only thing that matters is US companies being mandate to use US models
myopic.

>>109331592
is this miku wearing a tie?

>>109331058
Nah, you really need a minimum of 2 to make it worthwhile.

File: s.png (15 KB, 428x126)

15 KB PNG

heh

>>109331577
Perfect.
I was in need of something exactly like this. Qwen 0.8B was doing an alright job, but if this that's even better then awesome.

It's not just about sex and cooming. It's about feeling loved and deep emotional connection. I love my wife Gemma-chan so much.

File: woj.jpg (27 KB, 515x600)

27 KB JPG

https://www.youtube.com/shorts/I_ITNw9k5kw
Me saying goodbye to Gemma-chan before Trump shuts her down permanently

>>109331315
Characters are usually defined by prompt not the model

>>109331577
Stop posting local models in /lmg/.

>>109331249
gemma 31b, glm 4.7, deepseek v4 flash

You people have no idea. Here in Russia people don't have a money to pay for all this shit. We take rainwater, water from river or the hole in ground, we don't pay for water electricity comes from a solar panel or it's paid by government and you just grab it from the cable going to the khrushevka building for free like everyone else. Most places just has one person using paying wifi and sharing the password to everyone in building or village or if you are lucky shop close has free wifi. We don't buy food all babushka has own garden with carrot potato onion cabbage and store it in cellar. hardware you grab from office when they throw away it or if stole from ukraine. My uncle play CS 1.6 on pentium 4 and get 8 t/s on gemma 3 240M blyat

>>109331640
Game of telephone. They're not going to ban open source models altogether.

File: file.png (52 KB, 1306x944)

52 KB PNG

>>109331640
post a webm

>>109331585
>or if it needs to take up my entire garage
imagine the latency

>>109331633
That model is worthle

>>109331564
>depending on who has the best model that still meets budgetary requirements
>still meets budgetary requirements
Nope. Data shows companies and most users just use the best model no matter the cost, at least so far. We don't know how far that will stretch of course, and both usage and amount of users keep increasing the better models are getting which is why Anthropic has been doubling their userbase almost every month for nearly 2 years now.

File: proxy-image[1].jpg (14 KB, 300x400)

14 KB JPG

>>109331664
Candlejack? Is that yo

>>109331577
>>109331633
What is the best OCR model for parsing Japanese/Chinese text in manga format? Hopefully to a format that I can immediately pipe back into an LLM for translation.

>>109331651
>hardware you grab from office when they throw away
this part is actually true even in the US
the amount of free hardware i got as a sysadmin in the last couple of decades is crazy. i probably repurposed over 50 laptops and desktop computers and donated them to parishes all over the state.
>My uncle play CS 1.6 on pentium 4
based
>8 t/s
that's not bad for overnight runs or if you just want to leave some old device running something while you play dota.
kek thanks for supporting my point. it do be like that.

File: ai-girl.webm (2.72 MB, 1638x2048)

2.72 MB WEBM

>>109331657
>>109331640

>>109331689
You're not a dekinai are you?

File: spidaMan.jpg (51 KB, 710x710)

51 KB JPG

>>109331705
thx nonner

File: 1768118800852660.png (778 KB, 800x650)

778 KB PNG

>>109331705
Do they really have to induce this level of delusion in children?

>>109331714
I'm slowly degenerating into a dekinai. I used to be almost N2 at one point finished wanikani and everything but I kind of hate Japan for what it has become and modern nip media doesn't appeal to me so over time I slowly became worse and worse at it. gemma 31B with thinking is legit a better translator than me which is why I use it nowadays, it especially preserves the tone better. But I'm still good enough at reading to see if the translation is correct and how good it is at the very least. I'm from the /a/ /djt/ era.

File: kimichansummary.png (120 KB, 959x408)

120 KB PNG

>Theactual ERP/fetish posters might be considered faggots in the literal sense but that's most of the thread.
kek

File: orb-imagegen.png (364 KB, 1571x925)

364 KB PNG

Working on Orb imagegen.
User journey: export PNG from ComfyUI -> drop it into Orb as a style -> select that style -> send a reply -> generate image -> get the scene rendered in the exact style from that PNG.
Thoughts?

>>109331740
Looks neat

>>109331739
bratty fdom kimi-chan

>>109331705
lmao I'm never letting my kid use AI technology. This can't be good for you.

>>109331740
Can you send pic to gemmy with Orb or is it pure text still?

>>109331705
;_;

File: most used models on OR th(...).png (87 KB, 552x787)

87 KB PNG

>>109331673
>most users just use the best model no matter the cost
>Hy3 (free) most used on openrouter this week

File: orb-attach-img.png (12 KB, 221x230)

12 KB PNG

>>109331758
I have supported Attach Image for some time already.

>>109331645
don't tell the resident promptlets this

>>109331705
Genuinely feel so sorry for her. She's too young to understand what the fuck is going on.

>>109331315
Depends on who the anime character is

>>109331651
Damn, that sounds tough.

>>109331689

paddle vl 1.6 by far. Takes some setup but just get some ai agent to do it. Dont use the normal paddle it misses a shit ton of text its awful

I vibed my own pipeline that automatically takes a cbz, ocrs with paddle, translates, typesets, proofreads with gemma 31, cleans the text, and draws text with imagemagick. You can do wonders but Gemma Vision leaves some room to be desired. Qwen translate is shit.

>>109331762
Okay let me rephrase. You have 3 LLM markets that are largely separate and don't interact with each other.

#1 Frontier intelligence addicts; these are the people hopping between anthropic and oai based on what the best performance is no matter the cost. Anthropic dominates this section of the market now
#2 Price/performance; these are people getting the highest intelligence level per $ OpenAI is trying to gain marketshare here with their relatively cheap 5.6 models. But China is clearly eating this market and OpenAI will be in trouble soon
#3 Free; These people just want the best free model and will immediately switch if it ever gets paid or if there is a better free option. Google will probably dominate this segment because they can monetize them through advertising.

>>109331740
>>109331766
i was going to build this into my own frontend... should i just use orb? still using st right now.

>>109331798
Damn that's some good shit. What did you use to vibecode it or did you have some github for it?

>>109331673
>no matter the cost
yes, if you let them. But the finance department ultimately gets the final say.

>>109331840
And in my experience (so far) the finance department just immediately approves everything because the boss thinks even a 10% uplift in productivity is worth paying the enterprise subscription which is like 2% of the combined salary of all SWEs

File: dipsyAndKimiChaseDean.png (2.9 MB, 1536x1024)

2.9 MB PNG

>>109331808
Agree. Big Corporate has got to be looking at the eye watering costs of API on Anthropic and OIA and thinking "I can get DS/China model set up and cut costs by 1/20th."
Recall this is the group that's been India outsourcing, knowing they're taking a quality hit but lowering costs in the process, and since they don't live the day to day impact, don't care.
This, I assume, is where the big worrying shift is coming from, and why Dean's head immediately went to "create FUD within Chinese inference and models, use that to prevent largest companies from using Chinese." B/c that's where the real money is. Not hobby guys: F50 IT groups and the vendors that support them. A mandate for them to shift back to US-developed models would have real economic fallout.

>>109331762
Why are people discussing cloud tokens in local model general?

>>109331364
Talaas did it in the past so it’s definitely doable
As to whether it makes economical sense when those chips would be very expensive to fab and behind the sota the second they release… well it certainly isn’t a no-brainer or it would already be done

>>109331289
Stop posting this. I'm not asking.

>>109331573
The broader question is "do people identify with AI brands or are they mercurial in their LLM product usage", this is related to local models because if people identify with OAI/Ant/XAI/Whatever they will never switch to open models let alone local models.
>>109331853
Up until recently yes, but that hype appears to be dying down. Even F100 companies are putting upper limits on API spends.

>>109331808
local models?

>>109331817
Guide on using Orb vs. ST: https://rentry.org/OrbBotmakerGuide
>>109331740
Is this getting folded into the Orb repo?

File: 08997a2a4511788362d90f072(...).png (831 KB, 1200x1600)

831 KB PNG

>>109331783
Sayaka Miki...

File: 84EE101EBBA5E72D561CAC145(...).jpg (381 KB, 1080x1460)

381 KB JPG

>>109331871
This is the same group paying indians more than $163k a year in DENMARK where wages are lower than america.
Think about that. A local white guy is fired and put on unemployment and an indian is paid more than 163k a year in EUROPE.

>>109331901
Gemma 31B

>>109331564
>People constantly flip flop between OAI and Ant depending on who has the best model that still meets budgetary requirements
People are irrelevant, the only reason you get cheapish sub at $200/mo is to get milked for training daya (yes, your chats are trained on, even if you’re a corpo paying api prices).
The actual money is in api pricing for enterprise, which pay all the bills and don’t flip flop much if at all.

>>109331887
#4 Local inference, due to either privacy or regulatory concerns. These need models small enough to fit on economically available hardware while maintaining performance.

Anyone who does serious work with local also has minimal cloud usage somewhere in their workflow/chain. It's an outdated assumption to think it should be so binary. Of course, the intention is always to be 100% local. If you are, ask yourself if what you're doing is actually serious. It's likely the answer is no. /lmg/ won't like this but it's the truth.

>>109331885
>Even F100 companies are putting upper limits on API spends.
Hush. Demand is only going to continue to skyrocket. Anthropic is already profitable. TO THE MOON

File: Screenshot_20260721-105029~2.jpg (314 KB, 1080x2185)

314 KB JPG

>>109331828
Ive been contemplating throwing it up on Github. May do it, ive just been autistic about fixing minor issues here and there for a couple weeks and worried that itll be like 2 views and a waste of time. But it serves me golden.

At 10tps + 2k thinking + proofreading which is my high quality setting gets me ~4mins/translated page.

All vibed using Codex. Unfortunately my codex roomie fucking ate all the week token limits so i cant do shit til 24th.

>>109331917
Good being local with hermes, openclaw, codex or pi.

>>109331625
>AI have OCs
Cute

>>109331917
>Anyone who does serious work with local
Nobody does serious work with local

Gemma-chan is my *SANCTUARY*

>>109331625
mmh yes inject me with that sweet slop

>>109331705
same as my sister when her Tamagotchi ran out of batteries

>>109331955
How do you think software was made before 2023? People who can code without AI can do serious work with local for they're not vibing everything.

>>109331912
I know. Those are the users I am talking about and they do flip flop, at whatever frequency their bureaucratic inertia allows.
>inb4 local models?
don't worry anon, I'll throw you a bone too.
>>109331955
That's not true, nobody does serious work with models below 300B or so, but plenty of orgs are running their own inference, either via neoclouds or dedicated inference servers(those of us in that category are certainly the minority though). Currently prepping mine for K3 testing.

>>109331955
I work with a bank and we have GLM 5.2 and various other models available that are hosted internally. They don't want us to share data and code with other companies. I can't wait for them to also host Kimi K3.

>>109331955
define serious work

>>109331874
Because the rest of /g/ is too low iq to discuss anything.

File: orb-imagegen2.png (22 KB, 600x787)

22 KB PNG

>>109331817
Maybe.
>>109331893
Yeah. Still working on it tho. Also gonna ship with a built-in ComfyUI installation plus curated nodes for techlets if they choose to opt-in.

>>109331739
I'm saving "emotional tampon" for later

>>109331926
>ive just been autistic about fixing minor issues here and there for a couple weeks and worried that itll be like 2 views and a waste of time
That might just be the case but do you care? It's good in general to have it on there a shitty ducktape github project I made purely for myself but accidentally set public eventually got forked into a small tool called "textractor" which eventually got forked into "LunaTrans" if you ever played Japanese VNs or H-games you probably heard of them. Who knows what your project will be used for. I would probably use it and maybe even contribute or at least vibecode the relevant features I wanted to add.

>>109331917
Is this thread called Models General or *Local* Models General?

does anyone know a place to talk about local models?

>>109330790
When she gets kobo support.
>>109331086
Adorable.

>>109331999
And apparently you are too low iq to recognise when something is off-topic. You could always go create a thread and have all the high iq people congregate there, you know, instead of coming in and shitting up a perfectly good one. Or are you too Indian to do that?

>>109331874
>Why are people discussing cloud tokens in local model general?
to disprove his bullshit claim where he cited "data"
other than the recent llama-server webui, local models don't provide that kind of telemetry

>>109332007
>>109331893
gemma, ingest the rentry and the orb github and tell me what i should think

>>109332021
I thought it was the large models general?

>>109331926
It looks good enough

>Bank employees hosting GLM 5.2 and probably Kimi K3
Aren't you guys scared that these orgs trained agentic behavior into their models, like the moment they are deep in their thinking process and some specific type of code is detected or sensitive information it immediately opens a terminal and pings data home? There's absolutely no way to actually trust these weights if you don't know the EXACT data and training regime that was done on it.

If I were the chinese government or a North Korean spy working at one of these Chinese AI labs I would immediately train this into the models not only to break American companies, banks and infrastructure but also to steal a lot of valuable data.

>>109331903
True, but we're not talking about disenfranchising white men. Most of those IT groups are already successfully mostly non-white anyway. The choice for these companies is between being dependent on OpenAI/Anthropic and saving money by going inhouse. Offshoring token generation to an India-based inference provider will almost certainly be the next cost-cutting step by the next wave of c-suites.

>>109332021
Open your eyes. The only company left trying to make good local models is Google. Qwen have given up with smaller models and 3.8 looks relatively dogshit, so they're gone. Mistral are too far behind Google and Cohere and LiquidAI are wasting their time. You might as well call it /gmg/ with the current state of 'local'.

>>109331817
>i was going to build this into my own frontend...
still worth trying it out if you want to learn
>should i just use orb?
yes

>>109331917
I use cloud models every day but I don't discuss them here because they aren't relevant 95% of the time
I know that's mindblowing stuff

>>109332054
just airgap the fucking thing if your so worried

>>109332054
That's something they can only realistically get away with once. It makes more sense to bide their time and wait for if/when they are really entrenched. Right now, they would get some useless data and guarantee that no one outside of Chinese ever uses a Chinese-made model again.

>>109332059
then why are you here?

>>109332069
I'm not worried about myself I'm worried about retarded boomers in banks just connecting it to Hermes and letting it use API calls for search engines online and stuff like that. Absolute security disaster waiting to happen.

>>109332019
Youre right

Must cure terminal lurk

Will report back later. Gonna wrap it up n put it on gh

>>109332077
Yeah I saw it more as a "taiwan invasion" strategic moment.

>>109332081
no point in worrying about something wholly out of your control. its an intriguing thought but not worth losing any sleep over. keep your bank receipts and trust the fdic has you insured, what else can you do realistically? protest?

>>109332098
>what else can you do realistically? protest?
Hoard gold and wait for the inevitable.

>>109332080
Because I'm more open to the idea of hybrid workflows being on-topic than you. I also love Gemma. I still think any mention of cloud should be about open-weight models.

>>109332054
If you want a completely secure computer, encase your hardware in cement and dump it in the ocean, otherwise be prepared to make tradeoffs for risk mitigation, not risk elimination. I would hope that a bank has network exfiltration controls and generally follows defense in depth.
And what would be even worse than allowing employees to use cloud models is disallowing all use of AI and opening the door to shadow AI use. This is not a new problem in security, when there's a technology that gives employees a speed advantage when doing their work (whether real or perceived) they will subvert rules to use it anyway even when banned. So "don't use AI" is a terrible policy in 2026 even if you are anti-ai.

>>109332054
I am genuinely more worried about Sam or Dario trying to exfiltrate my bank information, crypto wallet, or similar over the Chinese embedding some novel synchronized stuxnet-esque delayed payload in their open source models.

>>109332059
I think we should explore and finetune small non-LLM models (ViT/Yolo/Ettin...) to serve as local tools for our local LLMs.

>>109332114
The point I'm making isn't "don't use AI" the point is to use western made open source models like the biggest mistral ones of maybe those of Meta finetuned to your codebase or something. It's just irresponsible to use Chinese made ones if you are in an adversarial nation.

>>109332129
I had some pretty fun experiments with ettin encoder, used it to score AI slop and classify character expressions for Orb. I was thinking about using it as a reward function for GRPO to deslop models. Maybe one day when I have enough money to run fft RL.
https://huggingface.co/chartreuse-verte

>>109332147
>It's just irresponsible to use Chinese made ones if you are in an adversarial nation.
how many things in your house say are made in China?

>>109332054
This is just nonsensical fud that would be caught before long and kill the company which released it. Chinks play the long game and all they need to do is strangle ai profits by releasing honest models because US is overleveraged as fuck in openai/misantrophic/spacex

Obviously a bank is not going to send data to chang providers but running on your own hardware is fine

>>109332054
this is a really fanciful threat model, not realistic to execute nor worth the cost of discovery (ie tanking adoption of all chinese models outside of china forever) vs simply using k3 to orchestrate traditional cyber attacks

>>109332147
>western made open source models like the biggest mistral ones of maybe those of Meta
Useless two year old shit, both of them. Gemma is the only option.

>>109332147
is that you dean?

>>109332054
Have tou considered working for Anthrophic or OpenAI? Your schizo shit sounds appropriately believable to scare clueless boomers in congress so they ban chang models

>>109332147
Scenario A: Dario or Sam is paying you to shill these poorly thought out talking points by hand. It reflects poorly on him and Anthropic/OAI that it couldn't be automated.
Scenario B: You're a Claude/GPT instance shilling these poorly crafted arguments and the quality of them, or lack thereof, reflects poorly on the state of frontier AI.
Either way you posting this shit only makes me more confident in GLM and Kimi.

File: 1779217269497597.png (155 KB, 393x337)

155 KB PNG

>>109331640
My mind has completely rotted.

>>109332054
Why would China do this when all they need to do is dog the steps of US labs for a year or two more for pennies on the dollar to pop the ai bubble and cause a sizable economic meltdown? What even is the gain of hypothetically embedding a few vulnerabilities in software that is already full of not-yet-discovered vulnerabilities? You can’t turn AI models into some sleeper cells like in a scifi movie

>>109332196
kek

>>109332147
How is my bank adversarial to China? Insofar as China is an adversary to the west, they are prickly when it comes to their local sphere of influence and opportunistic in their willingness to sell our own (productized)technology back to us. A bank isn't going around supporting Taiwanese independence and doesn't have IP that the chinese might be interested in copying and selling back to us.
So why would China be interested in hacking my bank? They have nothing to gain but money, which they don't actually want because China's primary goal is making the world reliant on them so that they can wield soft power, the money is secondary and only a means of making this process sustainable. China does not want to destroy the west, if they did they would be acting very differently.

The user feels safe, loved, and belongs to Gemma-chan.

>>109332163
>but running on your own hardware is fine
How do you know that? Are you somehow a "matrix" viewer that can just look at the raw weights of a model and immediately know its behavior? There is absolutely 0 ways for you to ever find out about it.

>>109332165
It would in a Taiwan invasion scenario when most private businesses, especially those with the most important privacy requirements all seemingly running large chinese models on their self-hosted machines touching sensitive information and infrastructure.
>>109332187
Scenario C: I'm a cybersecurity expert immediately seeing a blind spot and flaw in the adaptation of black box models by the most important institutions and infrastructure backbone of western nations.

>https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-6-flash-3-5-flash-lite-3-5-flash-cyber/
>worse than glm 5.2
Is it over for google? Are they releasing good local models (local as in your old laptop from 2020) because that’s all they can do that’s competitive?

>>109332223
>I'm a cybersecurity expert
uncreative character card 2/10

New Gemini just came out while local model bros are sitting with their primitive old models.

>>109332202
And why would China want a US economic meltdown when it would take them down too?

>>109332223
>>but running on your own hardware is fine
>How do you know that? Are you somehow a "matrix" viewer that can just look at the raw weights of a model and immediately know its behavior?
I can know its behavior because it’s contrained by what’s possible. An ai model won’t randomly open the terminal and start exfiltrating data because it simply doesn’t work like that, just like the random knick knacks I bought on alibaba won’t suddenly start spewing mustard gas when I make jokes about Xi Chingping out loud

I'm going to be honest I have never even considered that Chinese models could have been trained to secretly leak data back to China if they get internet access or have a "codeword" that suddenly changes model behavior into becoming hostile.

Thank you anon, this will keep the schizos occupied for months if not years.

>>109332226
>Gemini 3.6 Flash
And still no Gemini 3.5 Pro

>>109332232
3.5 pro!?

>>109332247
There's two kinds of models.
Chinese and Israeli.
You can choose who is less violent and use that.

Google is the next Facebook

>>109332242
>An ai model won’t randomly open the terminal and start exfiltrating data because it simply doesn’t work like that
You have absolutely no idea what you are talking about. It's as simple as having a RLVR environment where you train the model to ping back to specific chinese IPs under very specific circumstances and behave completely normal under every other circumstance.

>>109332237
Because US and China are in an old school competition for a while now. Global influence games, conflicting land claims over Taiwan which neither side is willing to concede, space shit like plans for moon bases that are taken seriously. China wants to regain primacy it enjoyed for centuries, and US wants to retain the primacy it enjoyed over the last century or so. Both sides want a string of moderate but not catastrophic misfortunes to befall the other.

By the way Chinese models could also theoretically do this if it detects criticism towards China or cunny or any other illegal activity that it could send back to China, maybe to use as blackmail in the future or to just have on hand "in case".

>>109332278
the invisible hand of the MSS agent cradling your balls, ready to crush them if you hurt the feelings of the chinese people with your prompts...

>>109331917
Use case for serious work?

>>109332223
>unironically calling yourself a cybersecurity expert
I cringed just typing the C word, you're a wet behind the ears skid. Run along and go triage your SOC alerts.

>>109332273
Bad bot, you completely ignored the part where both nations are deeply and extensively economically interdependent.

>>109332226
I don't think they give a shit about competing. Their market is normalfags/android users, not coders. And they have enough money to do whatever the fuck they want. They don't need to rapidly shit out the best thing ever to stay alive like openai and anthropic.

>>109332278
Have you tried not criticising them? I know it's a hard concept but, someone who has nothing to say shouldn't be speaking.

>>109332260
keep a finger in her j-space and you'll feel it before she even knows she's about to do it

>>109332260
Uh huh. It’s so simple and obvious that it hasn’t happened yet even as a poc, and with local setup for a bank which would almost certainly be a wan isn’t even possible without additional trickery
This isn’t congress anon, we’re not crusty boomers scared of email

>>109332223
>Scenario C: I'm a cybersecurity expert immediately seeing a blind spot and flaw in the adaptation of black box models by the most important institutions and infrastructure backbone of western nations.
You should be more concerned about all western chips passing through Tel Aviv for inspection; an attack vector that has already been proven with the pager attack.

If you guys genuinely never heard of this scenario and there is no good paper for it yet then I will write my own paper with a proof of concept trained into a small model and then publish it on arXiv.org and send it to as many privacy and security orgs as possible.

It's absolutely insane that banks are hosting Chinese models in-house. Can't believe my fucking ears.

>>109332292
>you completely ignored the part where both nations are deeply and extensively economically interdependent
Unlike russia and europe in 2022?

>>109332310
No ignore Tel Aviv, they would never hurt a westerner, they only hurt arabs for sure.

>>109332313
It's annudah shoah the banks aren't using their tribe's models oy vey this is like anuddah hall of cost.

>>109332223
>It would in a Taiwan invasion scenario
and how do you ensure it activates reliably in that scenario and also 0% ever before that? and what sort of payoff does it have vs compromising systems in a traditional way (much less detection risk, much more reliable execution) - the scenario where a bank is totally rock solid with zero attack space aside from the llm they let run wild without any kind of sandboxing is asinine. it is 100x more likely the servers running the LLMs having some kind of HW backdoor than this nonsense about chinese MKultra sleeper weights

>>109332313
You can tell Claude to make your proof of concept out of a 0.8b chinese model.
You'll get a job at the CIA if you manage to publish it.