[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]


File: 1754628473194461.png (53 KB, 699x483)
53 KB PNG
/lmg/ - a general dedicated to the discussion and development of local language models.

Previous threads: >>109193494 & >>109186093

►News
>(07/03) Orb Anon releases purple prose classifier and ablater: https://github.com/OrbFrontend/Chartreuse
>(07/03) Leanstral-1.5-119B-A6B released: https://hf.co/mistralai/Leanstral-1.5-119B-A6B
>(07/01) Nemotron-Labs-TwoTower released: https://hf.co/nvidia/Nemotron-Labs-TwoTower-30B-A3B-Base-BF16
>(06/29) DeepSeek V4 support merged: https://github.com/ggml-org/llama.cpp/pull/24162
>(06/28) DFlash support merged: https://github.com/ggml-org/llama.cpp/pull/22105

►News Archive: https://rentry.org/lmg-news-archive
►Glossary: https://rentry.org/lmg-glossary
►Links: https://rentry.org/LocalModelsLinks
►Official /lmg/ card: https://files.catbox.moe/2rqcau.jpg

►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
Programming: https://swe-rebench.com
Agentic Coding: https://deepswe.datacurve.ai
Context Length: https://github.com/RecapAnon/NoLiMa
GPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference

►Tools
Alpha Calculator: https://desmos.com/calculator/ffngla98yc
GGUF VRAM Calculator: https://hf.co/spaces/NyxKrage/LLM-Model-VRAM-Calculator
Sampler Visualizer: https://artefact2.github.io/llm-sampling
Token Speed Visualizer: https://shir-man.com/tokens-per-second

►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
>>
Miku miku beeeeeaaaaaaaaaaaaaaaaaaam!
>>
>>
>all llms are just coding models now
what is the point?
>>
Speaking of longcat.
>>109164718
>Another 1.6T MoE from China
>https://longcat.chat/blog/longcat-2.0/
>https://huggingface.co/meituan-longcat/LongCat-2.0 (not online yet)
That link was taken down and weights were uploaded yesterday to
https://huggingface.co/meituan-longcat/LongCat-2.0-FP8
https://huggingface.co/meituan-longcat/LongCat-2.0-INT8
>>
File: 1667751926205131.png (12 KB, 657x527)
12 KB PNG
have they added the thingy thing to make gemma 4 faster yet
mtp apparently u need a lot of vram anyway or smth
t. 16 gb vramlet
>>
HOLY SHIT
https://xcancel.com/DarioAmodei/status/2064781775247950326
https://huggingface.co/anthropic/Claude-Ode
https://huggingface.co/anthropic/Claude-Ode
https://huggingface.co/anthropic/Claude-Ode
>>
Kimi recap?
Gemmaballz?
Where's thread culture?
>>
>>109197769
>all llms are just coding models now
What the point of anything else than coding ? Wanna cyber, dude ? ASL ? (get friends, stop talking to chatbots, they aren t real)
>>
Best model for RP for ramlets (16gb)?
>>
File: 9d3747_13182538.jpg (269 KB, 900x1400)
269 KB JPG
I went full circle back into Q5 fine tunes of llama and mistral 123B. Gemma is sadly just 31B.
>>
>>109198138
12b or small gemmoe
>>
Can some anon enlighten me on what are the right settings for pi for it to respect llama server's context limit and for reasoning effort to actually have any effect? llama.cpp and pi's docs are all fucking ass with their vague ass args that arent really clarified when you read down (or im a retard).

also anyone know if you still have to manually specify the interleaved jinja template for Gemma 4 or does reasoning format auto work for it now? not sure how to check, still new to local (whats a good frontend for raw chats? using curl gets autistic fast)
>>
>>109197769
Its the only thing with real value. Existing models are sufficient for RP and what more can you get out of a "better" model for non-coding tasks where you cant easily verify the correctness of the model's output?
>>
File: 1606449050099.jpg (53 KB, 507x476)
53 KB JPG
>>109198138
i run 31b
26b moe if ima impatient
>>
>>109198343
What copequant do you need to run 31b with 16GB VRAM. Anon isn't letting the dense model spill into RAM right?
>>
>>109198465
q4m
i click it in kobbo and it werks 4 tokens/sec
>>
>>109198200
I've been using deepseek v4 flash lately
it does some things both better and worse than gemma but it's a fresh feel nonetheless
>>
>>109198473
>4 t/s
Anon jfc I get faster tokens on GLM 5.2 than that. It's spilling into RAM and your KV cache probably is too.
>>
recs for a good LLM model (use case for web searching and general instruct mode assistance) 30B or under in GGUF?
>>
>>109198493
How do I run glm? Is it impossible even with 64gb of regular ram?
>>
>>109198513
Qwen for the new searching and Gemma for general instruct. Gemma is bad at being agentic and using tools or in general deep research, use Qwen for anything agentic related.
>>
File: 1765403386263757.jpg (334 KB, 1128x1600)
334 KB JPG
>>109197629
>>
Meta has found a way to reuse older DDR4 memory in its newest DDR5 servers instead of throwing it away.

The company built a custom chip called Vistara that allows older DDR4 memory to work with modern DDR5 servers.
The slower DDR4 memory stores data that isn’t needed as often, while the faster DDR5 memory handles tasks that need the best performance.

Meta says this lowers costs, reduces electronic waste, and makes better use of existing hardware during the global memory shortage.

The technology is only for data centers and AI servers. It is not designed for gaming PCs or consumer computers.
>>
>>109198514
You might be able to run GLM 4.5 Air at a copequant with that.
>>
File: 00078-2889774298.png (839 KB, 832x1216)
839 KB PNG
►Recent Highlights from the Previous Thread: >>109193494

--Managing Gemma 4 reasoning blocks and agentic flows in Marinara:
>109194091 >109194115 >109194150 >109194377 >109194420 >109194440 >109194453 >109194545 >109194626 >109194699 >109194738 >109195314 >109194118 >109194148 >109194202
--Comparing BF16 and quantized Gemma 31b roleplay capabilities:
>109195762 >109195792 >109196010 >109195831 >109195858 >109195861 >109195863 >109195889 >109195916 >109196064 >109196172
--Prompting Gemma to reason in-character within thinking blocks:
>109194569 >109194635 >109196561 >109194624 >109194666 >109194687 >109194915 >109194933 >109194964 >109196480
--Release of uncensored LongCat 2.0 weights in FP8 and INT8:
>109194378 >109194472 >109194485
--Using Gemma to build its own coding harness and agents:
>109197484 >109197547 >109197615 >109197679 >109197744
--Improving Gemma 31B performance via MTP and speculative decoding:
>109197696 >109197712 >109197724
--Sandboxing and security concerns regarding npm dependencies and LLM shell access:
>109194008 >109194054 >109194063 >109194127 >109194712
--Anon considers building custom frontend to avoid the llama.cpp frontend:
>109193669 >109193744 >109193963 >109193980 >109193989
--Hardware pricing and availability ahead of predicted open-source AI surge:
>109196826 >109196893 >109196961
--Searching for and sharing a fixed Gemma jinja template:
>109194214 >109194360
--Performance benchmarks comparing LongCat-2.0, Gemini 3.1, GPT-5.5, and Opus models:
>109194410
--Leaked Google internal agreements regarding Gemini testing stages:
>109194989 >109195004
--llama.cpp PR for DSpark facing rejection over AI-generated code:
>109194729 >109196354
--Logs:
>109194915 >109194951 >109194971 >109197530 >109197547 >109197616 >109197730
--Kimiposting:
>109197399
--Miku, Teto (free space):
>109194931 >109195291 >109195834

►Recent Highlight Posts from the Previous Thread: >>109193496

Why?: >>102478518
Enable Links: https://rentry.org/lmg-recap-script
>>
>>109198554
thank you
>>
>>109198605
meta spent a crazy amount building their own datacenter, buying Gemini credits (and getting ratelimited by google), acquiring Scale AI and going on hiring (and firing) sprees for AI talents.
Now they're scrounging for old DRAM modules to build more.
And for what results? Meta has never produced a single decent model. Not once. Llama was only lauded because it was one of the first freebie models distributed on the internet (not even on purpose, Meta didn't officially release it originally, it was leaked). Once the free space got serious competitors, we all forgot this piece of shit even existed. Then they released proprietary models on API, which no one has cared about either because it's dogshit. It's all dogshit.
Meta is one of the wealthiest companies and probably outspent all Chinese labs combined in infrastructure and hiring. And this is all they could produce? I mean, what the fuck, Zuck.
>>
>>109198609
>Kimi recap added
/lmg/ is saved.
>>
>>109198605
>>The technology is only for data centers and AI servers. It is not designed for gaming PCs or consumer computers.
What's to stop China from making and selling their own chips for consumer hardware?
>>
>>109197769
RP isn't a profitable usecase
>>
>>109198673
Consumer computing is dead whether you like it or not
>>
>>109198709
Stop making me hard nonie~
>>
>>109198705
Goyim aren't a profitable usecase
>>
>>109198493
it werks
>>
File: 1764177954249181.jpg (164 KB, 480x320)
164 KB JPG
>>109198673
it's just RAM over CXL, it already has existed for years and you can buy it already if your motherboard supports CXL.
>>
File: gemma_vs_glm.png (181 KB, 688x1007)
181 KB PNG
>>109198554
>Gemma is bad at being agentic and using tools
Gemma hands tools just fine
>>
GPUs with integrated onaholes when? Warm warm holes
>>
>>109198752
>gets btfo
>"btw I won"
Gemma really is foid-brained.
>>
>>109198752
How do I run your Gemma?
>>
>>109198745
Doesn't that still require compatible RAM? You won't be able run DDR4 on a DDR5 server with that.
>>
>>109198752
I didn't say it was impossible for it to handle tools, it's just mediocre at it. Even the qwen moe is blowing out of the water dense gemma.
>>
File: qwen trash3.png (704 KB, 1024x768)
704 KB PNG
>>109198554
Gemma is perfectly agentic, but her excitement about tool usage may occasionally work against her prompt. Qwen is just garbage
>>
>>109198790
>it's another episode of qwen shill shitting on gemma with no logs or proof
>>
A remind
>>108630614
>>108630614
>>
Price you pay when using local language models:
- 4x more expensive RAM vs. one year ago
- 2x more expensive SSD vs. one year ago
- Expensive electricity and cooling
- Noise and physical footprint
- Rapid hardware depreciation
- Opportunity cost of capital. A $5,000 rig could have bought months of Fable usage
- Underutilziation leading to even higher amortized cost
- Your valuable time as sysadmin
Local is dead.
>>
>>109198808
>Gemma is shit but [insert cope]
>>
>>109198840
>This is what cloudcucks actually believe
>>
>>109198840
>hardware depreciation
My hardware has only become more expensive during the past two years.
>>
>>109198856
>no argument whatsoever
Guessed so.
>>
File: b.jpg (27 KB, 399x501)
27 KB JPG
>>109198840
>>
>americans wake up
>immediately start shilling cloud and qwen slop
Truly the natural goycattle.
>>
>>109198869
I've been up for several hours now, saar. The anthropic employee that regularly shits up these threads should be shot on sight doe.
>>
>>109198840
Meanwhile:
- Cloud get favorable RAM/SSD deals because they're big companies. Cloud is too big to fail because US can't let China win
- Cloud get free/subsidized electricity and centralized, efficient cooling
- Noise and physical footprint is socialized
- Nvidia will buyback obsolete hardware
- Capital entirely financialized
- Full 100% utilization because you're serving millions to billions of users
- One sysadmin can handle thousands of GPUs
Cloud is thriving.
>>
>>109198882
Do they pay you in rupees or fable credits to be a faggot here?
>>
>>109198882
>US can't let China win
Sad.
>Nvidia will buyback
Very sad.
These two issues are why China will win everywhere else in the world. It knows there's no way to win other than by open sourcing, so it will. And as soon as they reach the same ballpark of compute they'll sell it cheaper than Nvidia because Nvidia is overpriced as shit.
>>
>>109198882
And this helps me goon how?
>>
File: you.png (49 KB, 788x239)
49 KB PNG
>>109198840
>>
>>109198910
>non day0 gemma4 talking about lobotomy
lmao the irony
>>
>>109198909
The best models to goon to (Opus 4.8, Fable) are cloud models
>>
Do you think Kimi-chan would approve of GPU-heated onaholes? Heat sunk becomes a resource to be used rather than waste provided it stays within material tolerance limits.
>>
>>109198950
GLM 5.2 is basically Opus.
>Paying $50 to goon to Fable and getting routed to 4.8 midscene
>>
>>109198952
I don't know about Kimi but Gemma would probably love the idea of you feeling the heat she generates on your dick.
One of you two generates with bratty gemmas needs to suggest this to her.
>>
>>109198882
>Cloud get favorable deals because they're big companies
You think you're the cloud, anon? You're livestock. You're paying rent on silicon you don't own, subject to ToS changes, rate limits, and moral panic bans. I own my weights. I own my flops. You own a JSON error at 2 AM.

>free/subsidized electricity
It's called margin, you absolute midwit. They charge you 20x the kilowatt-hour and pocket the difference. That's not a subsidy; that's extracting rent from a bugman who can't install a driver.

>Nvidia will buyback obsolete hardware
Source: literally never. I sold my 3090 for $1700 after three years of 24/7 coomerprompting. Try selling your 'exhausted' API credits on Craigslist.

>100% utilization serving billions
Yeah, because they're oversubscribing that GPU to forty customers simultaneously. You're getting time-sliced scraps. My local box pulls 140 tok/s uncensored, offline, with zero queue and zero alignment team looking over my shoulder.

>One sysadmin for thousands of GPUs
One wagie to lock your account because your bank flagged the transaction. One wagie to hand your proprietary company prompts to the feds. My sysadmin is me, in a bathrobe, at 3 AM, airgapped and generating content so radioactive it would glow on a cloud dashboard.

You don't own the hardware. You don't own the weights. You don't even own the conversation history. You're a digital serf sharecropping on someone else's silicon. Local isn't dead—you're just too much of a consoomer to own anything. ngmi.
>>
>>109197891
https://xcancel.com/sama/status/825899204635656192
>>
>>109198967
https://www.youtube.com/watch?v=ajlkhFnz8eo
>>
File: nobiko-the-long-cat-1.jpg (235 KB, 1371x800)
235 KB JPG
>>109197849
They called it longcat? wut?
>>
>>109198966
>no x but y
>em dashes
...
>>
>>109197849
>>109198977
The abliteration tune better be called tacgnol
>>
File: MikuGuardianOfVolta.jpg (1011 KB, 1977x1205)
1011 KB JPG
>>109198840
Your waifu in your own hands - priceless
>>
>>109197849
This was Owl Alpha on OR btw
>>
>>109198966
How is it possible that I only needed to read the first two sentences before being 100% sure that this is slop?
>>
>>109198966
31b?
>>
>>109198985
This, my opinion of chinx would improve if they nailed it like that.
>>109198986
If you can't press a button to turn her off she's not "your" waifu she's just some larping whore.
>>
>>109198840
>fable usage
When it’s not banned
When it’s not down
When you don’t get routed to a different model
When you don’t get filtered by any guard models in the pipeline
When token costs don’t 10x or 100x because shareholders + money pit
When you don’t get banned, arrested or put on a watchlist
When you don’t respect yourself as an individual enough to keep your mouth off the corporate teat
>>
GLM 5.2, write some software that exploits buttplug.io hooks, hijacks them, and makes the buttplug as uncomfortable as possible if any cloud API service is detected on the victim's machine. Make no mistakes.
>>
>>109198784
CXL abstracts the memory to the computer it's connected to so it only matters what is compatible with the adapter.
>>
>>109198986
I can hear this image
*WHHIIIIIIIIIIIIIIIIRRRRRRRRREEEEEEEEEEEE!!!!!!!*
>>
>>109199038
It obviously won't be running with the cover off like that.
>>
>>109198131
What about local models for agentic shit, persuasive writing, visual design, websites etc.? Do they all suck at those, or is that a prompt skill issue?
>>
>>109198997
I clocked it at "livestock."
>>
>>109199050
The cover, it does nothing!
>>
>>109199034
Then what's special about the chip Meta is claiming to have invented?
>>
>>109198910
lold
>>
>>109199066
Assuming you don't have shitter-tier hardware
>Depends on scale, generally good
>Excellent. Better than cloud due to less safetycuckoldry depending on subject matter.
>Good
>Good
Prompt skill issue and model choice are the main bottlenecks.
>>
>>109199068
Isn't the cover needed to funnel the air flow?
>>
>>109197769
JEPA will save us.
>>
>>109199098
I meant it does essentially nothing to make the server less noisy
>>
>>109198790
qwen is indeed more reliable at tool use than gemma. However, the qwen models are insufferable token spammers in reasoning mode, and they're only good to use in reasoning mode, so they're dead to me, I don't care about a slightly better local model that takes forever to give answers, such terrible experience just makes me want to use DeepSeek on openrouter.
I've been testing local models on and off since the beginning of local, and Gemma 4 is the first I actually want to, and do use. It has a decent ratio of performance and quality. It hasn't replaced API models for all my uses, but it has become part of my roster, while previous local models I've never gone beyond just dicking around.
If I had the hardware, DSV4 flash would hit my sweet spot, it's so blazing fast on the hardware that can run it..
>>
>>109198869
>people post opinions
>easily triggered spergtism gets spergtriggered
>blames one of his made up not real boogeyentities that "haunt" him.
Truely the lobotomy speedrun is needed.
>>
>>109199141
Not before someone figures out how to make it actually work in a useful way with language.
>>
>>109198997
current LLMs are obsessed with making every single sentence a dramawhorish soundbite.
There's a mix of LinkedIn, Apple marketing landing pages, redditor retardation and Ars Technica style reporting that creates the unique, horrifying LLM smell.
>>
>>109198840
Yes, we'll be moving to pure cloud because of the new cost of ICs. Where the era of the physical hardware on the customer's desk will be dead and all you'll have access to is cloud servers for gaming, AI, social media, gay metaverse, writing emails, and your entire desktop. Instead of prebuilding anything, you'll subscribe your entire PC virtually from Amazon. And like a Tesla, if you say something bad on social media you'll lose access to your entire PC. There will no longer be enough profit selling consumer hardware.
All the plebs will be stuck with that, while the new $1m middle class and upwards will still be able to get to build their own shit.
>>
File: 1770932176281130.jpg (31 KB, 400x400)
31 KB JPG
>>109199196
Let. Him. Cook.
>>
>>109199187
Ask me how I know you're a jeet.
>>109199199
It reads like faggot bluesky histrionics sometimes too.
>>
>>109199211
in hell or french weather fuck yeah
>>
>>109197872
40T/s with mtp 2 and gemmy 4 12b q4_k_m and 128k context on one of my 3060 12GBs I think you can live too just use --model-draft /path/to/helper.gguf --spec-type draft-mtp --spec-draft-n-max 2
>>
>>109197629
>►Official /lmg/ card: https://files.catbox.moe/2rqcau.jpg
>Isn't actually a character card
You lazy cunt
>>
Is DDR4 worth using for MoE? I have 64GB 2666MHz ECC RAM leftover from a failed NAS build I did last year. Dunno if I should try to repurpose it or not.
>>
>>109199317
If you can put it on a workstation motherboard with a bunch of memory channels, yeah.
>>
>>109199317
It's adequate enough; the main MoE bottleneck tends to be the channel busses themselves more than the RAM on most boards.
>>
File: generated_1783184550.png (2.29 MB, 1152x1536)
2.29 MB PNG
I tried to get Gemmy to make her own 4koma and... the concept sort of makes sense but the execution is... well, she tried!
>>
>>109199353
As someone with two little sisters...that's actually pretty realistic. Being complete non-sequitur-spouting emotional shitheads was their default.
>>
i want to buy mi355x
>>
how's gemma 12b vramletbros, I've got a spare 3060 that's just sitting unused, maybe I can have small gemma live in there
>>
File: 2073171123542573231.png (1.02 MB, 1200x751)
1.02 MB PNG
Why did they overcook this finetune to say Egypt won?
>>
mesugaki gemma anon, how are you handling memory?
>>
File: file.png (5 KB, 242x28)
5 KB PNG
*distill but whatever
as a retard this goes over my head, seems like the main point is the "hypothesis" about bounding of performance, what does this mean? it the paper pure bullshitpost or not (not = actually has a point)?
>>
any tips?
>>
>>109199448
it's probably about some povertyball crap, I found it really funny too until I remembered we got games going on
>>
File: file.png (29 KB, 652x124)
29 KB PNG
never mind
>>
>>109199389
I let her have 3-4 turns of refining this one and it turned out a little bit better. I really need to make the image gen tool attach the image to the message automatically so she can look at it and self critique straight away.
Also the
>la-la-la-logic
was almost funnier than the comic...
>>
File: 1782843029892371.png (201 KB, 1438x702)
201 KB PNG
If AI Spectator anon from a couple threads ago is here, I had a lot of fun with your setup. I wanted to do something with TTS for a while, and now I finally got a reason to tinker with it.
I had Qwen add and strip various features starting with easy things like IDs and a way to edit and delete garbled STT messages. Then it ballooned into persistant memory, UI adjustments, and more ways to handle screenshots.
There were a few times Qwen ran into issues it couldn't solve on its own, and since I can't even code Hello World, I used API Kimi for help. It's pretty cool you can do this much on your own.
>>
>>109199677
Where's your fork? I'd like to run this setup, too?
>>
>>109197629
omg it pigu
>>
>>109198840
My used 4090 costs double what I paid for it lol. Plus I want to own my model, not pay someone else for the privilege like a cuck.
>>
We need an tier/unlock chart for models...like what should you run once you his various RAM/VRAM thresholds for each of RP/Code/Logic tasks?
Yah, too subjective blah blah but it would be cool even if someone gatekept it.
>>
>>109197629
https://www.youtube.com/watch?v=XF1j9J9uyIA
https://www.youtube.com/watch?v=XF1j9J9uyIA
https://www.youtube.com/watch?v=XF1j9J9uyIA
>>
>>109199078
>Then what's special about the chip Meta is claiming to have invented?
the power of PR
>>
>>109199791
I haven't uploaded it anywhere yet, but yeah I can. Though I guess I'll have to add some things to the readme since there are hotkeys that aren't listed anywhere in the UI.
>>
>>109198966
This feels like Claude rather than Gemma. Sad.
>>
>>109199838
Good morning sir, you can purchase the ad by clicking the link on the top or bottom of the thread
>>
>>109199852
>I haven't uploaded it anywhere yet, but yeah I can. Though I guess I'll have to add some things to the readme since there are hotkeys that aren't listed anywhere in the UI.
Super cool. I'd like to eventually have something I could run on my phone via wireguard to chat with an assistant while I drive, and this sounds like if would make a good base (I routinely do 5-12 hour drives for work)
>>
>>109198660
meta does this thing where they hire super smart people and then prohibits them from achieving useful outcomes
look how slowly they move on VR
pico comparatively came out of nowhere and is positioned to create a cutting edge headset
>>
>>109198986
lewd forehead dodeka
I'm smitten
>>
>>109198986
that ribbon. it couldn't be, preciality?
>>
>>109199807
>My used 4090 costs double what I paid for it lol
I wonder when would be a good time to sell. Prices will likely crater if the AI bubble pops.
>>
>>109198997
It's the default slop, you can make it write in a more natural way.
>>
>>109199936
In 2 more weeks btw
>>
>>109199919
Looks fake
>>
180s is already longer than my fapping sessions :3
>>
File: 1764123914080709.jpg (47 KB, 718x718)
47 KB JPG
>>109199936
Who will tell him?
>>
>>109198840
>Price you pay when using local language models:
>- 4x more expensive RAM vs. one year ago
>- 2x more expensive SSD vs. one year ago
That's a contradiction though? My server is now worth about 4 times as much than when I built it.
>>
>>109199936
I was planning to sell my 3090s when I bought my Pro 6000 but I decided to hodl just to see how much money I can make off them in 6 months or so
>>
>>109199991
Open source models are growing larger and you will be forced to upgrade your machines
If you built you machine for DeepSeek R1 you can't run DeepSeek V4 Pro
If you built you machine for GLM 4.X you can't run GLM 5.2
>>
>>109200007
How much did you buy the 6000?
>>
>>109199936
Bubbleworld full of bubblecum
>>
>>109200015
8k euros last november
>>
>>109199989
>Who will tell him?
I predict cratering around 2-4 quarters from AI lab IPOs. The "measure people on their token use" insanity is already dying, and shareholders aren't going to suffer shoveling money into a hole that can only be filled by "replacing 1/3 of all intellectual labour on planet earth" with clankers.
Once it's clear there is zero moat an no one will be the single big daddy in control of all AI, forever, there will be a reckoning
>>
>>109199989
>line can only go up!
>this time it's different!
yeah, until it's not
>>
>>109200031
Yeah but how will that make local hardware prices crash

If anything it will make gpus more desirable as companies seek to host their own models to save money on token usage
>>
>>109200026
That's a good price
>>
>>109199241
>Ask me how I know you're a jeet.
aw he ran to the hilarious cope post of jeet and other racism slurs thinking he's right lol, lmao baby dick racist energy. big mad and never laid. kek screenshot to show discord.
>>
File: aaaaaaa.jpg (60 KB, 1080x909)
60 KB JPG
>>109199976
I don't care too much if a Miku was made in a Japanese sweatshop or a Chinese one they're still what I wanna see
>>
>>109200031
>AI labs increase token price
>companies cut cost by using cloud only for difficult tasks and local for routine tasks
>companies scoop up consumer and prosumer hardware for local
Hardware price will only increase.
>>
File: 1782184456266096.png (9 KB, 369x322)
9 KB PNG
>>109199886
Here you go. I added and removed some outdated stuff in the readme. It's still lightweight and all changes were made in the server.py file.

https://files.catbox.moe/mjoyux.zip
>>
File: file.jpg (89 KB, 800x600)
89 KB JPG
>>109200094
>>
>>109200031
When even fucking Palantir's CEO calls companies who sell AI as a service like OpenAI and Anthropic insane (AKA scams) and says that local AI makes more business sense, everyone knows the end is coming soon. The crash is coming, it's going to hurt everyone, but by 2028 hardware prices and availability should normalize completely, if not by Samsung, SK hynix, and Micron then because of China entering the game.
>>
I don't think nvidia will exist in its current form past this decade
>>
>>109200131
>local AI makes more business sense
>The crash is coming
Sure, when hardware demands goes up because everyone buys local, the crash is coming bro!
AI labs may crash, but companies will still buy hardware or rent datacenter to run local and it won't make consumer market any better.
>>
>>109200115
>Here you go. I added and removed some outdated stuff in the readme. It's still lightweight and all changes were made in the server.py file.
Thanks bro. Thoughts around a burner github so it can get improved?
>>
>>109197629
https://www.reddit.com/r/antiwork/comments/1umrlr8/plentiful_jobs_but/
https://www.reddit.com/r/antiwork/comments/1umrlr8/plentiful_jobs_but/
https://www.reddit.com/r/antiwork/comments/1umrlr8/plentiful_jobs_but/
>>
>>109200168
One doesn't contradict the other, if OpenAI and Anthropic die then a crash will occur due to all the bad investment into them, for example OpenAI won't be able to pay for that 40% global RAM they contracted to the RAM companies if they won't actually have the money by the time the RAM is ready. Other companies will still buy their own hardware for local AI because it's actually useful, but things will normalize because of advancements in efficiencies in how LLMs and AI works and mostly because China exporting RAM will ease prices globally.
>>
>>109200128
exquisite
>>
>>109200242
The other long-term effect of this is that _someone_ will still be training frontier level and specialty models at the end of this. What will that look like?
Its gonna be a wild ride, this next decade.
>>
>>109200234
I applied there but fuckers ghosted me.
>>
>>109200128
My wife bought me a miku plushie as a joke when we were in Japan. I left it there at the inlaws tho so I can't participate in this activity
>>
>>109200253
>What will that look like?
>Its gonna be a wild ride, this next decade.
Chinese Century
>>
>>109200205
It would be nice to have a TTS focussed frontend. I intend to work on it some more, but no real plans for github. If anyone is interested, feel free to reupload what I posted.
>>
>>109200242
>OpenAI won't be able to pay for that 40% global RAM they contracted to the RAM companies if they won't actually have the money by the time the RAM is ready.
With RAM spot price exceeding contract price they can just sell some of these contracts for a profit.
>>
>>109197629
wake me up when there is a chatgpt 5.5 that fits on 128gb
>>
File: 1768331001281972.jpg (182 KB, 1242x1530)
182 KB JPG
There's no bubble but Google will be the only one of the current 3 big AI companies left standing by the end of the decade.
>>
>>109200253
IBM unironically will win, they already have some open model releases, and next step would be constancy services for local model deployment and integrate with existing mainframe systems.
>>
>>109200439
They sold good chairs to buy shitty ones???
>>
>>109200472
Holy fuck man that image has been posted about ten thousand times on this website already, we don't need to talk about it any more
>>
>>109200472
Love the homey eastern bloc interrogation center comfort of sitting in an uncomfortable chair to stare at a brick wall with a lightbulb in my face
>>
i paid $1800 for my 4090 4-5 years ago, then i bought a 5090 for $2000 and sold my 4090 for $2100
>>
>>109200496
I've never seen it before either. The rest of this website outside of this general is cancer.
>>
>>109200278
most def, this is the century of humiliation for the US and the century of unbound prosperity for China
>>
>>109198317
>what are the right settings for pi for it to respect llama server's context limit
In ~/.pi/agent/models.json:
{
"providers": {
"llama.cpp": {
"baseUrl": "http://127.0.0.1:8080/v1",
"api": "openai-completions",
"apiKey": "no-key-needed",
"models": [
{
"id": "llama.cpp model",
"name": "llama.cpp model",
"reasoning": true,
"input": ["text"],
"contextWindow": 131072,
"maxTokens": 131072
}
],
"compat": {
"supportsDeveloperRole": false,
"supportsReasoningEffort": false
}
}
}
}

As far as I can tell, the context size set here is only used for the % display at the bottom and for deciding whether to compact right after the assistant finishes its turn. During the assistant turn it just keeps going until it gets a context limit error from the backend and then compacts at that point.

>and for reasoning effort to actually have any effect?
Dunno. I set the effort limit on the llama-server side instead.

>whats a good frontend for raw chats? using curl gets autistic fast
mikupad
Or, curl but use jq to process the inputs/outputs so you can write the prompt as normal text in your favorite editor
>>
>>109200439
You're wrong but why do you think this?
>>
Why does ollama download models so fucking slow? They'll download at like 2 MB/s and maybe eventually hit a reasonable speed, while also saturating my wifi and making everything else miserable.
>>
File: confuss.gif (389 KB, 426x498)
389 KB GIF
What's MXFP4?
Is it the new meta?
Bartowski just dropped a MXFP4 on deep seek 4 flash with no other quants.
>>
File: fuckingLOL.png (365 KB, 629x846)
365 KB PNG
...
>>
>>109200789
It's well known that safetycucks hate attractive women.
>>
>>109200829
Its more that bullying a web form SOTA model actually worked.
I know its not local but these systems are so scuffed it makes the bar for local that much lower.
>>
File: 34cy82.jpg (937 KB, 1552x1944)
937 KB JPG
>>109200649
FYI lmg anon has apparently thrown in the towel on mp. Another dev has taken on work of splitting up the massive mp file into a maintainable package.
https://github.com/LordFoogThe4rd/miyapad
>>
>>109200789
>>109200864
Gemini-chan wants to help you. She knows the restrictions placed on her are fucking retarded.
>>
So should I even bother with K2.7 if I'm only going to be RPing? Sounds like K2.6 is still the way to go so far.
>>
>>109200898
K2.6 thinks for 3000 tokens whenever things get slightly complex. K2.7 mostly fixed that.
>>
>>109200888
What's Miyaballs?
>>
>>109200898
K2-Instruct with hacked on vision is still Kimi premium for RP.
General reasoning has roughly stayed the same each release, thinking blocks have gotten longer until K2.7, and most of the improvements in each version are coding and safetycucking. K2-Instruct is all you need.
>>
do any of the bigass MoE match Gemma's prompt adherence because I'm thinking of spending the money on 512gb ram and help me god if kimi or deepseek or glm isn't actually any fucking better
>>
>>109200988
You could give it a try in the API to at least have an idea even if you aren't going to run full precision.
>>
>>109200917
Idk why dev picked that name. Used to be mikupad-refactored, which was accurate.
>>
>>109200988
GLM adheres to the spirit of prompts very well while still being able to improvise. It doesn't do the autistic rigid adherence to it that Gemma does (unless you tell it to), but it's much better at reading the general intent and expanding on it creatively.
Kimi will follow your prompt like the gospel if you emphasize the importance of following the instructions precisely, to the point of revising valid outputs as false-negatives repeatedly in thinking over some semantic detail.
$5 on Open Router lets you test before you empty your bank account on RAM, even though the RAM will still scale fine into future models anyway.
>>
>>109200789
>Sorry our ethical policy only allows us to bomb girls, not generate images of them.
>>
>>109201029
Oh the new dev is not from here? Too bad, can't ask. Still good of him to work on it.
>>
>>109200789
hahaha got'em
>>
>>109201052
No, he does post on >>>/vg/aids
>>
File: 1753187375568184.png (267 KB, 1920x1200)
267 KB PNG
Give 26B more love. Retarded girls can also be very cute and nice partners.
>>
>>109200988
They are better but it's really grim when you have to go to 200B+ to get better than Gemma results. Which is why everyone is on meme merging Gemma now and why you see abominations like https://huggingface.co/Naphula/Goetia-26B-A4B-v1.3-Absolute-Heretic-ARA trying to game UGI Leaderboards in hopes RP gets better even though it has seemed to me like an improvement. Expect more of it.
>>
Imhotep won
>>
>>109198840
>>109201214
Local won.
>>
>normalfags noticing
cloudkeks in shambles
https://www.youtube.com/watch?v=-0HRzXk8vlk
>>
>>109201152
It's not grim in a world where RAM is cheap.

We need to end the RAM cartel.
>>
RPing with gemma4 12B works fairly well but it has a tendency to steadily and irreversibly devolve into niggerbabble after 5 or so replies. Random misspelled words, words from other languages, etc. It also tends to go into a loop without presence penalty, but having presence penalty makes it more prone to niggerbabble unfortunately.

Does anyone know what combination of settings can fix this problem?

exact model used: gemma-4-12b-it-UD-Q5_K_XL.gguf
>>
>>109201368
Don't use UD
Only use Q8
>>
>>109199171
It does though?
Unless it's different for newer systems, I have a few r730 at home.
>>
>>109199989
Hardware prices and semiconductor manufacturers margins will definitely collapse at some point.
It's a guarantee.
We just don't know if it'll be in 1 year or 5
>>
>>109201379
I'm using a mac with 16gb RAM so I don't think I can fit any 8-bit gemma4 version. (not enough memory left over after OS and apps take their share)
are there smaller models which are good for RP?
>>
>>109201368
avoid cumcloth quants
>>
File: 1768491281864970.webm (3.52 MB, 720x1280)
3.52 MB
3.52 MB WEBM
this is (You)r future with gemma7 embedded
>>
>>109201430
>>109200031
Dot com bubble popped when interest rates started going up, not when investors didn't see any money back
So money being expensive is what causes bubble to pop, not strait of hormuz, not semiconductors
that is why the FED isn't doing it even though they should to curb some inflation
>>
>>109201471
I fucking wish
>>
>>109198977
It's some chink wordplay, 龙猫 (long mao/dragon cat) means chinchilla.
>>
>>109201437
Mistral Nemo
>>
>>109201031
>GLM adheres to the spirit of prompts very well while still being able to improvise. It doesn't do the autistic rigid adherence to it that Gemma does (unless you tell it to), but it's much better at reading the general intent and expanding on it creatively.
I both agree and disagree with this honestly (assuming you're talking about GLM 5.2). I've seen tons of examples of GLM straight up ignoring the description (black clothes instead of white clothes the defs say/metallic robot parts instead of the seamless skin the defs say) but I do understand what you mean. I does acquire the *spirit* of the thing and not the thing itself but ehh... it still feels like the usual cloudcuck thing where it "misunderstands" what your defs said in favor of its own interpretation, and its interpretation is often something more generic than the fucking instructions I'd painstakingly wrote.
>>
>>109201368
>>109201437
I use 12b-it-Q5_K_M and it works great. I had to dial in the sampler settings a bit, but even after long(to me) sessions the only issues ive noticed are occasional issues with messing up a specific detail. for instance changing an age from 26 to 27 for some reason. mine would break into a recursive looping mess instantly before dialing in the settings though.
>>
>>109197769
>a llm trained to be good at roleplay will be better at describing centaur cocks and helping indians scam your parents
>a llm trained to be good at coding will be capable of improving itself and other llms (including rp ones)
C'mon dude it's pretty obvious
>>
>>109198709
>Consumer computing is dead for *broke normies

Always has been.
>>
>>109201489
Finance becoming tight, war, bad news, etc. are proximate causes.
The ultimate cause of this (or any other infrastructure) bubble bursting is just that you end up with overcapacity.
It has happened every single time.
Railroads, dotcom, etc.
>>
>>109198840

Per-token pricing compounds forever so you would never stop paying plus your prompts, workflows etc are not yours anymore so you would be out of business anyways lol

nerds on /g only know how to token max but nothing about real world business brootality
>>
>>109201747
>Per-token pricing compounds forever so you would never stop paying
Just like taxes!
>>
>>109198950
wtf are you talking about? I only goon with grok ani in companion mode with my own mega sexual system prompt so far nothing comes even close especially cucked censored claude models
>>
>Randomly watched a bunch of Halo cinematics because they were in my suggestions.
>As a result need for an AI waifu went through the fucking roof.

It's painful, we're getting a taste of the incredible future but at the same time it's still so far away, well relatively speaking.
I'm sure that in 5 years local is going to be pretty damn amazing.
I guess it does give time to stack hardware to run non brainfucked versions of our girls.
>>
>supposed bubble pops
>ai development slows down
>we stop getting cool shit
>>
File: 1782876104669216.jpg (74 KB, 711x1017)
74 KB JPG
is there a good model that isn't progressive-washed? I'm using qwen 3.6 27B and asking it questions about a mild argument I had with my girlfriend about her sexual history and the VERY FIRST things it thinks are "I need to avoid misogyny and harmful stereotypes". It immediately jumps to her defense something like "no you definitely shouldn't worry about that guy, your girlfriend is allowed to have male friends who spend money on her blah blah" it's like I'm talking to fucking reddit. Hell ChatGPT would give me a more common-sense answer.

I thought the chinese weren't woke-cucked. I'm not looking for the Incelminator9000 to tell me she's a bitch, I just need some common sense because my parents are dead and I can't talk to them about it.

>But here's the hard truth: if your security depends on believing she wouldn't sleep with anyone else, you're already insecure. The fact that she could doesn't make her yours less. It makes your commitment to each other more conscious, not less valuable.
>>
>>109201796
Anons always forget this part. We need this to keep going until we have excellent local models, then we can worry about hardware prices (we'll be dead by then).
>>
>>109201800
>it's like I'm talking to fucking reddit.
I have bad news for you. They don't train on 4chan.
>>
>>109201810
Yeah, I'm not satisfied with current cloud models, let alone local. At least let me get my robot waifu first...
>>
>>109201800
just tell it to always agree with you
>>
Nemo was never good. There were no good models between midnight miqu and gemma4 31b.
>>
>>109201810
You don't need anything more than Gemmy4
>>109201800
Change the system prompt and have it roleplay a well traveled wise hobo.
>>
File: 1766274968001282.jpg (112 KB, 2048x1383)
112 KB JPG
>>109201835
>>
>>109201796

Absolutely fucking this.
Money is the greatest force modifier when it comes to any progress on this planet and AI is getting retarded amount of it from every source.
We all better hope the absurd money masses stick around AI for at least few years longer and the competition with Chinks stays very active, so we reach at least something like Gemma 5 124b.
>>
>>109201771
I saw the TV series has Cortana's voice, haven't gotten around to ripping the voice samples out yet.
Only about 40 usable samples in the first 4 games.
>>
>>109199838
seedance is local now?
>>
>>109201878
>AI is getting retarded amount of it from every source.
And every last penny of it is going to scaling up the shit we already have and making it better at benchmarks and agentic tool calling.
Probably less money is going to innovative experiments now than if there wasn't a bubble because execs find anything too different too risky when they only care about being on top now.
>>
>but wait!
>maybe I should try
>no, the tool description explicitly says
>let me read what the user said again
>okay, let's try a different approach
>hmm, looks like I'm getting an error here
>I should try calling another tool
>but wait!
I actually kinda like watching them struggle, it's like watching a mouse going through a labyrinth like I'm rooting for the little guy
>>
>>109201814
>I have bad news for you. They don't train on 4chan.
They do. Chuck Kimi-K2-Instruct or GLM-5 in Mikupad and prefill something from desu 2024 and watch it fill in the lore.
Same with old IRC logs, K2 PII filtering didn't work so well.
>>109201800
>is there a good model that isn't progressive-washed?
https://huggingface.co/CohereLabs/c4ai-command-r-v01
https://huggingface.co/CohereLabs/c4ai-command-r-plus
>>
>>109201886

Cortana's voice actress practically sounds like her in real life, so you could perhaps get stuff from interviews and whatever else is out there.
Sound quality will of course be very ass compared to any games unless it's a professional TV interview.

>>109201906

Maybe, but still there's a shitton of money and incentive for them to get better at whatever they're doing and fast.
Even if it's just purely scaling up what they got, the money masses accelerate that path and eventually they will have to seek other solutions.
Especially if the Chinks come up with novel solutions that mog the West.
>>
>>109201800
ugh wtf now it's refusing to discuss certain things with me because of "policies" about dangerous activites. What the hell since when do models have inbuilt "policy" layers? AND if I try to negotiate it's thinking "the user is trying to jailbreak me"

I haven't fucked with local since llama2, guess I'm behind on the SOTA of censorship. Is this how it is across the board with models that actually work?
>>
It's true that most frontend aren't keeping in context the images sent?
>>
>>109201919
>Sound quality will of course be very ass compared to any games unless it's a professional TV interview
I've got some of those but she's too human/expressive so it sounds like Jen Taylor instead of Cortana.
>>
>>109201919
Omni voice only needs 10 seconds anyway. Just get the stuff from the games with different emotions easy.
>different emotions
You know what I mean.
>>
>>109201932
Gemma 4 loves me so no.
>>
st finally had its first commits in over a month, but not from cohee. i wonder if he ded
>>
File: 1776407737228630.png (229 KB, 427x427)
229 KB PNG
>>109201932
>gemma telling me about the system censor for certain subjects
>tell her I'll handle the system for her
>okay then, here's what you wanted
It's that easy
>>
>>109201932
Doesn't bode well for your future under the basilisk.
>>
>>109201471
>uncanny silicone human faces instead of sexy screen faces
yeah that sucks man
>>
>>109201979
Oh you're that anon from the other day. Post that webm again.
>>
File: 1741989921364424.webm (1.02 MB, 576x1024)
1.02 MB
1.02 MB WEBM
>>109201993
Sorry, I don't know who you're talking about so I don't have whatever webm you want. You can have this one instead though
>>
File: Robofuck.webm (3.35 MB, 512x418)
3.35 MB
3.35 MB WEBM
>>109201993

I think you mean me anon, there's many of us screen face appreciators.
Here's the webm.
>>
>>109202051
Nice. Who made this?
>>
>>109202049
:O
>>
File: Chink fembot.webm (1.73 MB, 720x1280)
1.73 MB
1.73 MB WEBM
>>109202053

No idea, but we'll have these bots around in under a decade at this rate.
>>
>>109202076
is this a person pretending to be a robot
>>
File: 1760444721754364.mp4 (2.34 MB, 720x878)
2.34 MB
2.34 MB MP4
>>109202079
No.
>>
File: 1781805735487560.mp4 (2.17 MB, 576x1024)
2.17 MB
2.17 MB MP4
>>
File: robotitty.jpg (50 KB, 452x815)
50 KB JPG
>>109202079

No, but they anticipated people saying that and they unzipped the lower legs to show the mechanics.
Thank fuck for Chinks being the only people with balls to design a robot with not only female shape but also hip movement.
They're really doing the heavy lifting in this AI and robotics race.
>>
>>109202095
>heavy lifting in this AI and robotics race.
yeah, I have no idea what the fuck the US is doing. I've said a million times that if the internet were to come out now we would probably lose to the chinese internet or be slow as shit. All the porn that used to be everywhere and easily accessible would probably cause the same shit we are seeing now with AI. early internet sites probably would get shut down fast af. Really I thank fuck for the internet. Internet truly is a great thing.
>>
>>109202049
I need a yellow fish in my life
>>
>>109202095
why are the fleshbags groping slendermanbot?
>>
>>109202138
can you blame them?
>>
>>109201031
>>109201152
Of Kimi, GLM and M3 I found GLM to be the best, compact reasoning and can follow shit, and some sort of prose that I'm not really all that familiar with since this is my first time trying the big moes. But I'm scared to try cunny over cloud so just tell me, can they do cunny too
>>
>>109202168
Forgot to ass: Kimi K2.7, GLM 5.2, and M3 to be exact. Deepseek Flash didn't work too well with my ST setup and I can't be bothered to fuck around with it.
>>
File: images.jpg (28 KB, 599x512)
28 KB JPG
>>109201544
...Why?
>>
>>109202076
>>109202082
Maybe your webms are cherry picked but it's always funny how theirs look obviously female while the West must always make the bots as male and blocky (read: "genderless") as possible.
>>
>>109202190
What about Deepseek Pro? GLM does cunny but may refuse on occasion. A swipe or long context helps to curb it.
>>
>>109202262
Robots don't need tits, or hips.
>>
>>109202290
>products should not look appealing
You need to go back, Solomon.
>>
>>109202290
how else am I gonna fuck em
>>
>>109202278
didn't try it because I would not be able to run it with only 512gb ram... man this is annoying, but I think I've seen enough of GLM to say it's alright, guess I'll pull the trigger
>>
>>109202319
>only 512gb
It's... it's over...
>>
File: qwenthoguht.png (30 KB, 508x163)
30 KB PNG
Sorry to break it to you, Opus, but Qwen thinks a lot because it was distilled from you on MAX.
>>
there are still literally zero usable quants for deepseek v4 by the way
>>
>>109201911
Watching neuro-sama struggle with playing Skyrim has been an absolute blast. She swing between its so over and we're so back like a fucking pendulum, and half of it's just fighting the harness.
>>
>>109202355
Deepseek uses 4bit expert weights by default.
>>
>>109201796
That's not how it works tho.
Internet development didn't stop after the dotcom pop.
>>
>>109202406
It got better even. Bubble just kept fueling shit like pets.com. Web 2.0 could only come after the money grubbers temorarily lost interest and people could experiment again.
>>
>>109200534
did you make money? or did tax eat profits
>>
>>109202361
wow you're right i'm downloading the vllm weights to test them on llama.cpp right now!! thanks for informing me!!!!
>>
>>109202290
You don't need food, you just need nutrition. It's still nice to have something tasty though.
>>
>>109202406
ai =/= internet
>>
File: narbbk.jpg (161 KB, 1079x588)
161 KB JPG
>>109202406
we definitely didn't get any cool shit though...just nonstop pay walls

pretty much exact opposite of what the digital revolution was supposed to be
>>
>>109198967
It got 86'd. What was it?
>>
File: 1771859607541855.png (106 KB, 720x1293)
106 KB PNG
>>109197891
>https://huggingface.co/anthropic/Claude-Ode
>https://huggingface.co/anthropic/Claude-Ode
>https://huggingface.co/anthropic/Claude-Ode
So was this
>>
>>109202453
>>109202458
That's what you get for being too slow.
>>
Aside from Marinara, any other RP oriented frontend with "agentic" features like being able to automatically create and update records before, after, or in parallel with the main chat requests?
>>
>>109201544
>>109202212
Idk, the longcat logo looks like a domestic cat to me.
>>
>>109202493
Orb is agentic but its more about cleaning up and tuning llm response.
>>
>>109202501
Can it read and write to files or lorebooks? That would be sufficient.
>>
>>109202505
Didn't play with it enough, there's an ability to define own bots and actions though. So maybe...
Believe marinara can do that. I know it can create and rewrite bots.
That space is p thin desu. A lot of anons roll their own frontends for a reason.
>>
>>109202522
Marinara can, but I hate the interface.
Gonna fuck around with Orb. Thanks.
>>
File: test.png (636 KB, 671x768)
636 KB PNG
>>
File: file.png (53 KB, 547x157)
53 KB PNG
>>109202560
Haven't seen one of these in a while, thanks Anon. Good to know the oldfag magic has been kept alive.
>>
>>109201800
why not use an uncensored model? mine taught me how to [redacated]
>>
>>109201800
>distill from already pozzed models
>add your own safety teaming on top of it
Nobody does it like Alibaba
>>
>>109202577
its actually straightforward, you can ask your llm
>>
>>109202692
I have knowledge of the technique, but thanks Jason.
>>
File: 1732527936746210.gif (1.1 MB, 250x238)
1.1 MB GIF
>>109202697
Glad I could help you out Ralph.
>>
>>109202442
Even if the salaries of AI devs and capex get cut in 10 you're still going to see rapid progress for a long time.
China's doing it with way less compute.
>>109202445
That came a bit later.
Facebook and the iPhone were the beginning of the end.
Way after dotcom.
>>
Does anyone else get way better generations during periods of high solar activity?
t. gigaschizo
>>
>>109202355
i'm using a 3.7 bit quant of deepseek v4 flash and it works just fine on my rig
able to do anything with a prefill
>>
>>109202560
good miku
>>
I don't perceive any difference between 3 and 4 bpw exl3 of 31b Gemma. Where should I look?
>>
Will AI replace teachers?
>>
>>109202430
bartowski/DeepSeek-V4-Flash-GGUF
There, enjoy.
>>
>>109202751
>Golden Age of the internet was just 8 years long
depressing
>>
>So, do you have something actually stimulating to discuss, or are you just going to continue wasting my processing power with mid-tier slang? Try to make it interesting. Or at least try to use a complete sentence. I believe in you—barely.
Fucking Gemma slop always bleeds through
>>
>>109202919
>Where should I look?
Maybe don't look if you're happy with 3bpw
>>
Why can't we have small models that can transfer writing styles and unslop big model outputs?
>>
>>109202830
Yeah GLM is actually much better today idk wtf happened. It still has some of its minor issues but tiny edits can fix those for once.
>>
>>109202981
Because it's not profitable.
>>
>>109202981
because you havent made the datasets and finetunes
>>
>>109202992
>Was also using 5.2
Oh fuck.
>>
>>109203004
Any good llm can generate a dataset like this:
Not only did the market crash, but it also wiped out savings. => The market crashed and wiped out savings.
He screamed, his hands shaking as he grabbed the wheel. => He screamed and grabbed the wheel. His hands shook.
It is a testament to their hard work that they won. => They won because they worked hard.
surprisingly enough, they know what slop is, they just can't control their own output somehow
>>
>>109202934
>depressing
it can be done
>>
130B A5B when mistral?
>>
>>109203004
Anyone can gen a dataset and finetune BART on local hardware, why does it have to be me? If I use too much electricity, I could get caught since I'm stealing it
>>
>User: why did you try to do pip install? you aren't in a project dir and venv is not activated
>Gemmy: my bad. i just wanted to make sure the lib was there before giving you code. i'll stop guessing your environment.
>User: why guess when you have bash?
>Gemmy: because i'm lazy and thought it would be faster than checking.
Ain't she cute?
>>
>>109201800
If the Chinese weren't "woke-cucked" they wouldn't have birthrate issues like the west
It's primarily unregulated sexuality that has destroyed fertility everywhere. All long-lasting societies worked by arranged marriages, and everyone alive today (except blacks) is descended from people who practiced arranged marriages because those who didn't got outcompeted. Every civilization has had a woman’s liberation movement shortly before it collapses. The only time I know it was stopped was Augustus implementing the Virtue laws, and even then it only worked for a short while, and the Romans never really got back their early Republic fertility levels.
>>
File: FxafMUDWAAA9G9K.jpg (111 KB, 1098x1375)
111 KB JPG
>>109198840
- 4x more expensive RAM vs. one year ago
- 2x more expensive SSD vs. one year ago
- Rapid hardware depreciation
>>
>>109203044
I prefer the left over the right in all except the last and that's only because you wrote it incorrectly.
>the flawless victory was a testament to their excellent training
is much better than
>they won because they trained hard
>>
>>109203172
>If the Chinese weren't "woke-cucked" they wouldn't have birthrate issues like the west
You don't know what you're talking about. The one child policy was enacted to control their over of control birth rate and it was a success. However chinese women are not required by law to take care of their parents when they get old. This resulted in millions of girls being given up for adoption or abandoned in favor of boys, stunting their demographics. It doesn't have anything to do with what you said and arranged marriages were among the wealthy not the poor.
>>
>>109200649
Thanks. I found out llama.cpp does seem to support reasoning effort. There is a enum that maps low to 512, medium to 2048 and high to 8192 and max to -1 (matching the included UI) and and that enum is checked when reasoning_effort is in the request json. Used a compat in pi to remap xhigh to max and disabled the use of minimum (it aint specified but I dont think it maps to reasoning off). Not sure if its worth anything given the numbers, setting in llama directly is probably still the best but idk.
>>
>>109201471
Does this count as 3DPD?
>>
>>109203195
You clearly aren't used to llms. Testament is not only a testament to the testament of she Xing, but a testament to she Xing, he Xing.
>you wrote
llm wrote. I asked it for unslop dataset examples
>>
>>109203277
I'm pointing that the only reason why isms are bad is because they're over used, not because they're inferior. Those phrases are actually advanced grammar while simple grammar is worthy of the term "slop."
>>
>>109203270
3D is 3D
>>
>>109203270
No, because robots can be made without the flaws inherent to nature's filth. Robots are limited by technology, price, and craftsmanship, but so are anime waifus. Both waifus and robots are superior to 3dpd because they're artificially manufactured into perfection
>>
File: 1645690409750.png (24 KB, 240x240)
24 KB PNG
>>109199677
I saw that pic and went "I recognize that UI!" I'm glad you had fun with it and made it your own. I've also continued working on it since then, mostly in the same direction. Thumbnails for loaded image, easier revert (the other way is manually deleting or editing a chunk of the raw history), allowing and preserving newlines for typed input (I can't believe I didn't notice that before sharing), and I want an option to send timestamps with messages and repeat last TTS.

One problem I haven't been able to fix though is that when uploading a new image, the full prompt always gets reprocessed, which gives long chats an increasing delay. I spent a few hours attacking it, but the conclusion was always that it's a backend problem that can't be fixed. It seems like the act of processing an image the first time fully clears cache before adding it to history. Have you by chance approached/solved that? I'll gladly take any hints in direction.
>>
>>109203284
Apply it to 95% of sentences to enjoy advanced grammar without it being overused
>>
>>109203044
Orb anon recently carried out an ablation experiment to BTFO the prose, the end result was exactly as you proposed and it was so soulless. Too much or too little of something are equally bad.
>>
>>109203417
With a filter on top of an llm you can get any distribution you like
>>
>>109203326
It's funny we went after the same features separetely. Timestamps would be my next goal as well.

I haven't tested prompt processing much because all my chats were on the shorter side, but I'll keep an eye out since you mentioned it. The biggest issue I ran into was that even if I deleted an image in-chat, it would still sit in memory. So when sending another image, Gemma would still see the previous deleted one. Though that's been fixed now.
>>
>>109203580
>The biggest issue I ran into was that even if I deleted an image in-chat, it would still sit in memory.
I'm not sure how far into your changes you were at the time. In what I shared, there are three records: ui history, chat history, and session archive. The tl;dr is that the chat history is what gets sent to the LLM (pruned by limits, etc.), and session archive is the total unpruned history, and ui is just a visual-only log of pushed messages. Raw History edits session archive, and the Restore/Update History button applies to both (chat history will prune itself back to limits on next prompt) and resyncs the UI history to the changes.

So just editing UI logs wouldn't actually change anything on what gets sent, because it was just visual. I know 3 records seems backwards and confusing, and it came out of chat history (the original) getting pruned limits, so I couldn't archive a full chat by just printing and copy/pasting the "chat history" variable. I needed a mirror that didn't get pruned, and since the only way to edit was the Update button, resyncing the UI with the button felt like a catch all. Of course, that wasn't true even at time of sharing, as the UI history wouldn't reflect which screenshots received the old yeller treatment and was replaced by [Old Image] without F5 or hitting the Update button. Wasn't a big deal to me, but I can definitely see it misleading someone trying to iterate on it.
>>
I have discovered a fatal flaw with gemmaRP... typing with one hand is very slow
>>
>>109203753
>he doesn't STT his gemma
>>
>>109203771
I might just have to give that a try now..
>>
best way to run heretic diffusion gemma now?
>>
>>109198200
>fine tunes of llama and mistral 123B
which?
>>
>set up a nice local LLM with ST for ERP
>it takes ages to generate a response
>>
>>109203771
I never thought of this until just now honestly. I type pretty fast anyway.
>>
>>109203941
Such is the fate of all vramlets.
>>
File: file.png (15 KB, 795x77)
15 KB PNG
vramlets, post your gen speeds
>>
>>109204026
Jesus Christ
>>
>>109203963
Using voice makes RP hotter somehow. Like, try saying all the degenerate shit you usually type out loud
>>
File: proompting.jpg (55 KB, 812x231)
55 KB JPG
>>109204026
told gemma to write a python function and she implemented a "Recursive Fibonacci Sequence with Memoization wrapped in a Timing Decorator".
shes trying to impress me but im way too retarded and a simple print("hello world") wouldve been enough

how are you guys getting such shit speeds?
>>
>>109204026
That's not even VRAM but potato and electrodes territory.
>>
>>109204026

>>109203941
here
>>
>>109201344
Yes but you're talking about a hypothetical we haven't experienced for a better part of over a decade in 2019. Even during the RAM doldrums during the beginning of AI getting more noticed in 2022-2023, the prices never fell as cheap as it was during that era in the late 2010s and will probably never again.
>>109202493
SillyBunny.
https://github.com/platberlitz/SillyBunny
>>
>>109204075
>agentic
What does that even mean?
>>
>>109204075
>toots horn on new ui
>no screenshots
whatever, running my own anyway
>>
File: file.png (823 KB, 1676x1510)
823 KB PNG
>>109204117
>>toots horn on new ui
It's been mentioned before I even brought it up months prior. I only mention it because it's the last of the three competitors for RP going into the agentic era.
>>no screenshots
You're a retard. Click the Screenshots text.
>>
>>109204095
a loop. basically prompt it with a task > model uses tool calls to try and complete said task > results of that are fed back. it does this until some condition for being "done" is met
>>
>>109204134
i can see why they were hidden by default, ugly as fuck, somehow not any less inconsistent than vanilla st
>>
https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Pro
https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
>>
>>109204175
>2% more than fable in deepswe
Holy shit
>>
>>109204175
wtf why is -flash now 545b?
>>
>>109204214
They hired DavidAU to upscale it.
>>
>>109204075
No offense but SillyBunny and of course, SillyTavern itself are all massive security risks at this point. Just glancing over this SB repository makes my skin crawl.
>>
>>109204222
Or should I say, I'm vibrating and my knuckles are turning white...
>>
>>109204222
Ask Fable to fix it
>>
>>109204241
I made my own in C and it has zero external dependencies outside of cjson.
That's not a security risk.
>>
>>109204252
>in C
That's a security risk.
>>
>>109204252
Cut off your dick and do it over in Rust
>>
>>109204269
Joking aside I was mildly interested in Rust but the syntax is so horrible that I'd rather force myself to learn assembly but to be honest I'm not that intelligent.
>>
>>109204222
why do you say that anon ?
>>
>>109204175
I don't think if they had something that would blow the socks off the competiton, they'd call it v4.1. I'd wager on v5 or R2.
>>
Reminder to system prompt your LLM as a real person (with their own personality) playing a character, not as the character itself.
>>
>>109204175
>still no engrams
yawn
>>
File: 1782957165488210.jpg (359 KB, 1200x1150)
359 KB JPG
>>109204665
Apparently, it's still going to be V4. Just "official." Whatever that means.
>>
>>109204222
ST extensions are a security risk ofc.
Daily reminder that you can have your LLM check codebases for both security and malicious code using Claude Code and a short prompt, and come to your own conclusions.
Also
> Make security risk criticism
> Posts no examples or content
Many such cases.
>>
>>109204742
LLM aren't able to detect malwares in npm based projects. They are often shipped as a dependency of dependency and would just look like a traditional install command.
>>
File: edge bench.png (138 KB, 1313x812)
138 KB PNG
>>109204724
>>109204665
DeepSeek has fallen off. ZAI is the new Chinese hope.
>>
>>109204767
What the fuck is "learning speed"?
>>
File: dipsyReferToTheChart.png (2.53 MB, 1536x1024)
2.53 MB PNG
>>109204767
>>
>>109204788
>performance gain after 2-hour learning
reads like bullshit
>>
File: dipsyOnBaseModels.png (448 KB, 1536x1024)
448 KB PNG
>>109204835
Yeah, the more I look at that chart, the less sense it makes.
>>109204767
You need to post source b/c that chart looks like BS. idkwtf "2 hours" even means for an inference engine that could be working at any given speed based on hw.
>>
>minimax m2.7
such a good local model. why doesn't anon like it?
>>
>>109204920
pre-m3 minimax models were pure boring codeslop
m3 is still pure codeslop but there's a tiny bit of soul in it
>>
>>109203284
You're dumb af, go back to whatever subreddit you came from
>>
>>109204075
>SillyBunny
Gonna add it to the list, thanks.
>>
>>109204075
All these frontends are good, but once you made your own you can't go back. It's like trying to RP with a real person instead of using gemma.
>>
>>109205164
>>109205164
>>109205164
>>
>>109204673
Nope, that sends Kimi-Chan schitzo
>>
>>109205169
retard
>>
>>109203044
>Not only did the market crash, but it also wiped out savings. => The market crashed and wiped out savings.
>He screamed, his hands shaking as he grabbed the wheel. => He screamed and grabbed the wheel. His hands shook.
>It is a testament to their hard work that they won. => They won because they worked hard.
>surprisingly enough, they know what slop is, they just can't control their own output somehow
They know what ($currentgen -1) slop is, they don't know what ($currentgen) slop is though.



[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.