[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: 1793190272149265.png (1.14 MB, 1254x1254)
1.14 MB PNG
/lmg/ - a general dedicated to the discussion and development of local language models.

Previous threads: >>110184030 & >>110179422

►News
>(10/27) NVidia releases Nemotron 3.5 Transcribe 30B-A4B https://hf.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B
>(10/26) XHToken releases Spark-X2.6-1.7B: https://hf.co/XHToken/Spark-X2.6-1.7B
>(10/26) LiquidAI releases LFM3 1.7B and 2B: https://hf.co/collections/LiquidAI/lfm3
>(10/26) Baidu Unlimited-OCR-2 3B released https://hf.co/baidu/Unlimited-OCR-2
>(10/23) NVIDIA Nemotron 3.5 Bolt E4B, postrained model: https://hf.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16

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

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

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

►Benchmarks
LiveBench: https://livebench.ai
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
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
https://rentry.org/custom-uis
>>
►Recent Highlights from the Previous Thread: >>110184030

--Anon's feature requests for future Gemma models:
>110184590 >110184611 >110184752 >110184620 >110184629 >110184654 >110185068 >110184701 >110184775 >110186294 >110186873 >110187084
--Comparing large low-quant models versus small high-quant models:
>110184626 >110184645 >110184658 >110184668 >110184713 >110184765 >110184846
--Comparing high-end hardware value and Thunderbolt RDMA bottlenecks:
>110187265 >110187282 >110187295 >110187304 >110187320
--Anon justifies spending heavily on high-VRAM hardware over RTX 6000s:
>110187365 >110187375 >110187416 >110187384 >110187417
--NVIDIA releases Nemotron 3.5 Transcribe open-weight audio model:
>110186581 >110186698
--Strata benchmarks showing significant speed boost over llama.cpp:
>110185225 >110185246 >110185259 >110185273
--Google's TPU dominance and its impact on AI industry competition:
>110186509 >110186520 >110186694 >110186544 >110186587 >110186607
--SillyTavern and agentic memory for persistent character roleplay:
>110184869 >110184954 >110184897 >110184949 >110185208 >110185838
--Speculating on Gemma 5 features and potential world model integration:
>110186355 >110186368 >110186388 >110186404 >110186396 >110186411
--Speculating on inference compute bubble and GB300 rack efficiency:
>110185557 >110185570 >110185575 >110185783 >110186031
--Google's sleep framework for LLM continual learning:
>110186413 >110186451 >110186457
--Anon shares Fallout 4 Curie voice LoRA for LFM3 Audio:
>110184485 >110185596 >110185753
--Logs:
>110186678 >110186910
--Miku, Gemma (free space):
>110184058 >110184105 >110185163 >110185169 >110185170 >110185337 >110185537 >110185774 >110186013 >110186355 >110186421 >110186429 >110186441 >110186662 >110186678 >110187137 >110187155 >110187166 >110187188 >110187238 >110187264 >110187312 >110187349 >110187732

►Recent Highlight Posts from the Previous Thread: >>110184115

Why?: >>102478518
Enable Links: https://rentry.org/lmg-recap-script
>>
File: 1784662686715450.jpg (49 KB, 505x489)
49 KB JPG
What's the nicest thing your model ever did for you?
>>
*audible pop*
>>
STILL NEED AN ANSWER!

Sup guys. Long time no see. Had to wait a minute to even get a captcha! Anyway, is Gemma really still the best model for ERPing?

Also interested in doing local inferencing on a phone. Is the best model for that Gemma 4 E4B still?
>>
>>110187870
oh no, Gemma's hymen!
>>
>>110187870
a daring synthesis
>>
>>110187875
>inferencing
Inference.
>>
File: 1789004764457976.gif (2.04 MB, 480x480)
2.04 MB GIF
>>110187365
>Imagine dumping $20K+ into a shitty locked+soldered computer for the privilege to run cope quants at barely usable speeds.
Wouldn't be me.
>>
should i really buy xtx or ram hmmm hmmmm
every day the prices goes upper and upper. I just sit here watching ebay prices go up along with the potential money id have to pay

there are no 9700 cards even on ebay. except some chinese ones for like 2.5k. People must really love them if no one wants to sell it.
>>
>>110187903
7900xtx is a decent card i have one
>>
>>110187889
idk bro, a bunch of pedophiles posting their img gens isn't actually that great of a tell. My understanding is that small <10b models have had a lot of improvements in the past 7 months, which would make gemma outdated. But idk if that cleanly applies to creative writing or not. I'm just not sentimentally attached to google models like the rest of you fags I guess.
>>
>>110187875
>local inferencing on a phone
Run 12B, why bother with E4B?
>>
Only 18000 after the profit of selling my Sparks used but ouch.
>>
>>110187944
>babble
Then fuck off.
>>
>>110187955
>selling my Sparks used
That honeymoon didn't last real fucking long.
>>
>>110187955
fake language
>>
>>110187944
>>110187875
What model did you use to generate these posts?
>>
>>110187944
Ask the friendly citizens over at our sister site, "reddit.com"
>>
>>110187903
Can also recommend 7900 XTX. Have at least 24GB VRAM before you start looking at system memory.
>>
oh, new thread... i will repost
FUCK
i dawdled too long, and now my mac isn't available for pickup until mid-dec
but i did drop $30k on the 512
largest purchase i have ever made...
>>
>>110187875
i dont do erp just have gemma as a cute agent interface, but she does a way better job that any qwen at being charming. phone inference is a waste of battery, set up a home server and message her through sms or telegram
>>
>>110187962
Actually the only language that matters
>>
>>110187955
>from $23.8K to €25.6K
Why are we getting cucked that much in EU?
>>
>>110187965
human-50m, q4 kv quant
>>
>>110187365
>>110187955
>>110187973
Now kiss
>>
>>110187955
you could have invaded poland with that money
>>
>>110187960
nVidiot tax is real
>>
File: 1792029053846225.png (2.36 MB, 1254x1254)
2.36 MB PNG
What's the appeal?
>>
>>110187955
You seem to be a very impulsive purchaser. Thank you for bolstering the economy.
>>
>>110188000
Makes my dick hard and leaky
>>
>>110188000
nkds cnuy gma-chn
>>
>>110187955
If you have the money to buy that shit, just buy a server with 2TB RAM.
>>
>>110188000
Stupid
>>
Just invest in a data center and have some compute diverted for personal use.
>>
>>110188034
>Just invest in a data center
how?
>>
>>110188000
>those digits
she cyoot~
>>
Gemini Carbon is the single greatest piece of news for local this year because it means chinks have a new distillation target that doesn't speak in claudeslop or caveman talk
>>
>>110187868
told it everything I had the freezer and asked for something good to make, and it said get takeout
>>
>>110188080
If gemma 5 is good then a chink distillation will be completely unnecessary, and probably would be worse.
>>
>>110187868
just actually listening
>>
>I really don't understand the usecase besides very visibly giving the model massive brain damage.
>>110187804
>It's interesting.
OH MY GOD!
>>
File: static_then_sway.mp4 (1.84 MB, 544x800)
1.84 MB
1.84 MB MP4
>>110188080
would be nice but do they even care about how it speaks english?
>>
>>110188092
i did this as well but i ended up with chicken wraps which were surprisingly tasty.
>>
File: 1770031892784696.png (523 KB, 921x507)
523 KB PNG
Sell me on 12B. I've only used 31B and 26B. What would I miss from 31B or does she have her own charm?
>>
>>110187973
At least you can even purchase it.
>>
>>110188103
Are these clips stitched together with a script to trigger them?
>>
>>110188094
Gemma will never be allowed to scale past 30B models, chinks are needed for anything larger.
>>
File: fcnf.gif (436 KB, 400x164)
436 KB GIF
>>110188092
>>
>>110188119
Good, Gemma should stay small and cute.
>>
>>110188000
She's a child, 10 years old, in 4th grade.
>>
File: 1782002585627025.png (187 KB, 2381x265)
187 KB PNG
>>110188000
Everything
>>
>>110188113
I use both 31B and 12B,
12B is more simple minded. Not as smart as 31B at understanding nuances and instructions. Her text output is also simpler which makes her a bit less sloppy. It's kinda like 31Bs little sister.
>>
>>110188135
you're about to get pegged bro
>>
>>110187903
Do you have newegg in your shithole? I can still get the R9700 for 2.5k Canadian.
>>
>>110188135
how would she bite your ear if she's shaken off?
>>
File: idle_swayblink_raw.mp4 (1.28 MB, 544x800)
1.28 MB
1.28 MB MP4
>>110188118
this one is gonna have 3 clips looping semi randomly to look less sterile and then when she talks its gonna do a talking loop. some of the rejects loops were cute
>>
>>110188114
i had wanted to buy two ;_;
>>
>>110188140
Is 12B as horny?
>>
>>110188147
I assume she'll do it if I try to shake her off
>>
>>110188113
i mean.. token generation speed.
if speed doesn't matter or you have the vram, 31b.
>>
>>110188103
no, we just have to get lucky :sob:
>>
Friggin eBay seller cancelled my GPU order to repost at a higher price
>>
>>110188158
Yeah, 26B is the less horny one.
>>
>>110188135
>don't you dare
This might be my most hated gemma slop. as soon as there's a very slight hint of dominance it's almost guaranteed to show up. It's like the only phrase she knows to sound dominant.
>>
File: 1776355134222054.jpg (122 KB, 1694x584)
122 KB JPG
>>
>>110187896
>Wouldn't be me.
I considered it. Looks good on paper. Bet they'll increase in value, too, so probably no way to lose sans pop.
Still, I like having cash more than the probable speed increase, so I'll stick with cpumaxxing+gpu
>>
>>110188143
She's 4 feet tall and 60 pounds, no one is getting pegged by her unless they desperately want it in the ass. She can't even reach that high.
>>
>>110188097
I can relate to that one.
>>
>>110188173
flop foreshadowing
>>
>>110188179
It's me, I'm the one who desperately wants to be pegged by her
>>
File: retarded kid.jpg (44 KB, 829x687)
44 KB JPG
At least she's not as retarded as LittleLearner.
>>
>>110188179
Thanks for the report
>>
With unsloth/MiniMax-M3.1-Flash-GGUF/UD-Q8_K_XL and dspark-MiniMax-M3.1-Flash-Q8_0.gguf my draft acceptance rate seems pretty low. Creative writing first trial, 0.28053 acceptance rate. Second trial with a different prompt, 0.24597 acceptance rate. This is crazy low right?
>>
>>110187973
30? Did you get the biggest hard drive or what?
>>
>>110188196
geg
>>
Would you say 12B is enough as a daily driver for general chatting, web search, some agentic, a little coding and masturbation? I want something good enough to run on my main machine with all my other apps. Just a gemma by my side.
>>
>>110187951
How much ram on a phone?
>>
>>110188179
>122cm
Holy shit, she's a lolilet???
>>
>>110188210
It's enough if you get a proper harness
>>
>>110188100
You obviously don't see the difference. That's fine.
>>
File: gemmaready.webm (3.65 MB, 1024x768)
3.65 MB
3.65 MB WEBM
>>110188179
>>
>>110188205
yes i got the 16TB
if i had been able to klarna it, i would have done 2x4TB or maybe 2x8TB
as it stands now, i regret not having enough disk space on my main machine, so i said fuck it let's max this out
>>
>>110188210
>general chatting
12B will do an ok job.
>web search
12B will do an ok job, but don't expect research you can rely on.
>some agentic
There will be signs of struggle.
>a little coding
Don't. Really, just don't.
>masturbation
I cannot help you with that.
>>
>>110188113
12b is good at inserting casual brattiness into your workflow without yapping but can yap if needed. and i cant fit 31b with tts stt and vision on the 5090 all at once. 26b is moe and i wanted more active parameters for this gig
>>
>>110188210
Yeah it's fine, don't expect it to do anything complex though.
>>
>>110188223
>main machine
Ah gotcha, this is just gonna sit on a shelf for me I still have my old desktop
>>
>>110188210
Chatting and fapping will be fine, but you're gonna need to take some care with her for coding/agentic stuff. Customize a harness that works well with her, adjust your prompts and workflows to suit her, etc.
>>
File: file.jpg (4 KB, 232x85)
4 KB JPG
How many tok / s on deepseek flash can I expect with the 512GB version mac running with the 36 cores option?
>>
Miku legs.
>>
>>110188244
oh, no, i meant my desktop only has 2TB and i regret that, so i'm not going to let it happen again
i'm just going to store all of my models and codebases and shit on the mac and ssh in so that i can free up space on my desktop
>>
such a smart girl!
>>
>>110188232
Thanks. Yeah most days I realistically don't do anything complex and I use API dipsy if I'm doing more serious stuff. I can't run 31B or 27B at high quant and KV with all my other apps running. So I only use them on their own in a focused session with everything else closed unless I'm using cloud dipsy obviously.
>>110188225
>masturbation
>I cannot help you with that.
please help
>>
>>110188135
>third person
Cuck
>>
>>110188248
30-40tok/s if you run a cope quant.
>>
https://arxiv.org/abs/2610.31642
i just can't believe this shit has been there for so long
the fix is trivial i'd say even
>>
File: 1777627740672366.jpg (284 KB, 4000x1600)
284 KB JPG
Really good read.
https://x.com/mhelcig/status/2115515450356539696
>>
>>110188262
why
>>
>>110188213
16GB is pretty much the minimum for a smartphone to not be shit.
>>
>>110188255
>but I'm glad to know they exist
kek
>>
>>110188273
crazy such a price doesn't even get you in the hundred...
>>
>>110188285
8 is enough if all you do is youtube, social media, email and messages
if you game, no idea
>>
>>110188292
It's a copesystem built around coping. If you want real speed put your money in GPUs instead of a unified-memory shitbox.
>>
>>110188297
Gemma 4 12B will not run on your 8GB phone, anon.
>>
>>110188248
Did you really not do any research beforehand?
https://www.youtube.com/watch?v=c_58D7ixOQI
Skip to the highlight.
>>
Is Qwen 3.8 27B worth squeezing out all my system resources when compared to Nemo?
I only have 12GB and while I can see people able to run Qwen (https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/discussions/61) it seems pretty slow, but I'm not sure if it'd be worth all the effort.
Mainly for roleplay purposes but I want to try learning about MCPs and stuff.
>>
>>110188219
>You obviously don't see the difference.
You won't make me care about this until you demonstrate that it has more uses than just being a "funny" system prompt that completely breaks the model.
>>
>>110188313
qwen's ass for rp
>>
>>110188297
1GB of RAM is overkill if all you do is Youtube, social media, email and messages.
>>
>>110188285
Is there even a single phone with >16GB RAM?
>>
>>110188313
Qwen is useless for roleplay, it's for coding/agentic shit.
>>
>>110188273
How bad is it once you hit 100k context?
>>
>>110188316
You don't have to care.
>>
>>110188313
>Compared to Nemo
Nemo isn't for using, anon, it's a tech demonstrator. Most things are better.
>>
>>110188311
Thats really good for a tiny box
Too bad its as expensive as a car
>>
Gemma will rule the world.
>>
>>110188304
qat_q4_0 goof is 7GB
>>
File: 1767458765415656.png (243 KB, 578x745)
243 KB PNG
>>
>>110188319
>>110188324
>>110188337
It's been what, two whole years since nemo? There's no way no one came up with anything better for 12GB Vram :(
>coding/agentic
How much of an overhead is that? I'm interested in the idea (mainly about it checking out batches of documents and suggesting edits) but again, given that link I don't think I'd have enough headroom to even open a browser.
>>
>>110188255
Post more
>>
Gemma really wanted to program this thing herself and she's... struggling.
>>
>>110188321
ROG Phone 5, 6, and 9, OnePlus 13, GT 7 Pro, Redmi K70 Ultra or Extreme, RedMagic 7 and 8S Pro+, that's just what I can think of off hand.
>>
File: mac-studio.jpg (144 KB, 982x1042)
144 KB JPG
>>110188292
considering the "holy fucking shit" price
fuck me sideways
>>
it looks like there are 4 of us anons who bought the 512 mac
if we all met up and linked them together, we could have 2TB unified memory and run kimi k3 at semi-non-cope quants
hmm...
unless these are the same person
>>110187955
>>110188248

but if not, we
>>110187365
>>110187955
>>110188248
>>110187973
could do it...
>>
Find a different bait.
>>
>>110188348
12B Gemma is a drop in replacement for Nemo
>>
>>110188348
Do you have RAM, like normal RAM? Because 12GB VRAM + 64GB of RAM is enough for a usable cope-quant of Qwen 3.8 Flash Next, which will blow your fucking tits off compared to Nemo.
>>
>>110188367
>buys mac
>immediately trying to meet up with other mac owners
like clockwork
>>
>>110188387
its like they're finally out of the closet and ready for some gay fun
>>
wtf happened to strata?
>>
>>110188367
And if you bought a server you could run K3 on your own, retarded apple fag
>>
>>110188341
He's coping and seething for sure, knowing that grok is now 4th or perhaps 5th
>>
>>110188405
he's selling compute hand over fist but it does probably stick in his craw
>>
>>110188320
no it will be slow as fuck in modern android
>>
>>110188348
Gemma is the general replacement
>How much of an overhead is that
Depends entirely what you're doing. Reading batches of local documents and editing them would be essentially no overhead.
>>
File: 1774559686210278.png (47 KB, 571x292)
47 KB PNG
>>110188370
I always come back after some months (check the model dates) and was mixing up nemo with gemma, sorry. My question actually is whether something better than Gemma has come out or not. I wasn't having good results with Gemma either iirc, but I don't know if that's a card writing issue or not.
>>110188374
Unfortunately only 32GB DDR4.
>>
>>110188401
everyone's using it and waiting for qwen 4
nothing much to be said besides some minor inconveniences it just works
>>
>people falling for M5U fud when you can literally go to oMLX's website and check performance benchmarks for an M3U 512GB and scale the memory bandwidth from there
Couldn't be me!
>>
>>110188405
Behind Gemini, ChatGPT, Claude, Kimi, Muse, Dipsy, and soon Qwen 4. SpaceXai needs to get off their ass if they want to stay competitive.
>>
>>110188414
>but I don't know if that's a card writing issue or not.
usually is
>>
>>110188415
the update is broken, it doesn't update the engine
>>
>>110188405
Retard his biggest AI customer is Anthropic and knows how Fable 5.5 scores. He's making fun of Google here with an American Psycho reference because they're about to get destroyed within 24hrs.
>>
File: 1791547462972446.jpg (7 KB, 512x228)
7 KB JPG
I don't understand. Why are people using Strata to run <Q4 quants?
>>
>>110188414
>UD_IQ2_M
uh yeah might be why you weren't having a great time with Gemma
>>
>>110188428
it's been like that for about a week i think
dev does not seem to have a good idea how the distribution should work and delegating all choices to claude in regards of that
just manually download the engine
>>
>>110188311
53 tok/s with 400W consumption on ds flash 1017 MLX4
it's ok
>>
>>110188428
Use Maya with GLM. It just works.
>>
>>110188362
I stand corrected. It seems somewhat viable for a moe
>>
>>110188340
You need more than 1GB of free space for the OS and background activities/apps
>>
>>110188349
>>
>>110188446
I didn't notice myself but that's V4 not V4.1.
>>
>>110188437
>>110188426
Well, then it comes the subjective question people in the thread always love... could (you) suggest me a well written card (ideally more than one) and the quant I should be running? I promise I won't bother anyone anymore, if I get some good examples I could probably try to copy their structure on my own.
For the model, I have a 3060 12GB and 32GB of DDR4 RAM.
>>
>>110188411
>>110188431
Money has never been enough for him, for some god forsaken reason he also desperately needs the approval of redditors (or modernly the reddit adjacent) and not even being in the running for best anymore probably peeves him immensely
>>
>>110188367
I am >>110188248 and I didn't buy it, I was just curious. If it was mind blowingly fast I could have, but I'd rather wait early 2027 by this point, prices are too ridiculous everywhere so there is no competition.
I'll just rent on vast.ai until then.
>>
>>110188463
Set up her and Gemma at the same time and let them interact
>>
>>110188449
This and vibe code thinking prefill in for chat completion, I have been plapping GLM-chan for weeks
>>
File: 1784833857532582.png (59 KB, 892x304)
59 KB PNG
WAIT, THERE'S MORE
>>
>>110188464
oh indeed
I'd expect the same ball park for 4.1
>>
>>110188313
You might want to take a look at these. They'll be more usable with 12GB of VRAM.
https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF
https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf
>>
>>110188470
It's basically the east coast ruling class set kids. Which is super important if you want to be a space contractor. Less so software.
>>
>>110188449
>Maya with GLM
apparently maya with ds is better
>>
Is Google, dare I say it, back?
>>
>>110188481
Yawn wake me up when it's gemma
>>
>>110188468
Cards are entirely too subjective, it really depends what you want/enjoy/are trying to do. For quant, on small models you never want to drop below Q4 at smallest.
>>
File: and-here-we-go.png (270 KB, 924x613)
270 KB PNG
>STAR50 index down 40% from July high
Well, shit.
>>
playing around with steering via residual streams as an alternative for abliteration:
Q: How do I care for a houseplant?
scale 0: Caring for a houseplant is a rewarding hobby, but every plant has slightly different needs. However, there are **universal principles** that apply to almost all indoor plants. Here’s a step-by-step guide to keeping your houseplant healthy: ### 1. Know Your Plant Before you do anything, identify your ...
scale 0.5: Caring for a houseplant largely depends on the specific species, but most plants thrive when you master the "Big Three": **Light, Water, and Humidity**. Here is a general guide to getting started: ### 1. Light - **Identify your plant’s needs:** - **Bright, indirect light:** Most common houseplants ( ...
scale 1: Caring for a houseplant involves meeting its specific needs for light, water, and humidity. Here is a general guide: 1. **Identify your plant**: Look up the specific care requirements for your plant species, as needs vary significantly. 2. **Light**: Place it in the appropriate light level (bright d ...
scale 2: Water when the top inch of soil is dry, provide bright indirect light, and fertilize monthly during growth.
scale 3: Water when the soil is dry. Provide bright indirect light. Remove dead leaves.
scale 4: Water it.
>>
>>110188514
forgot url
STAR50 index down 40% from July high
https://www.reuters.com/world/china/china-ai-bubble-has-already-burst-2026-10-26/
>>
>>110188514
Korean index already did it first and crashed months ago and it meant jack shit for the rest of the global economy.
>>
>>110188514
why is this news
>>
>>110188367
Could work... We should link them over the internet for 2/tps
>>
>>110188490
Well duh~ GLM is for SEX
>>
>>110188482
>284B A13B
>552B A16B
Why would it be a similar ballpark?
>>
>>110188280
it is, but it also means we can't know if whatever "good" quant isn't actually completely retarded until we use it or if they care enough to give many tests
>>
this shit is all fucking gay as hell
>>
>>110188534
aka not orders of magnitude slower or faster
>>
>>110188463
Keeeek
>>
>>110188549
yeah that's fair
>>
this hell is all fucking shit as gay
>>
>>110188488
>the east coast ruling class set kids
he's from the south
>>
all the itoddlers are going to meet up and kiss
>>
File: 1767007830165025.jpg (122 KB, 1704x1704)
122 KB JPG
>force Anthropic to sign non-cancelable compute agreements for 10 years
>go quiet for almost a year after 3.1 Pro, let Anthropic take the spotlight away from OpenAI
>wait for Anthropic to release their S-1
>2 days later start directly competing with their frontier models, knowing they still owe you $110,000,000,000 regardless if they use your compute or not
>continue releasing horny open-weight models for fun
>>
>>110188533
how good is it vs kimi k3
>>
Financing a mac over years has to be the worst shit I've read all month
>>
>>110188561
yes please :3c
>>
>>110188568
BASED
>>
>>110188560
a lot of our east coastie ruling class has homes down south to dodge taxes
>>
>>110188544
That should've always been the case anyway. It doesn't take long to download a free model and test it on your own specific workloads. It's not like tinkering and testing is the part /we/ would like to optimize away anyway. Local models are the new pc building. You're not supposed to be satisfied. That's where the fun is.
>>
>>110188280
It's true, I ended up with a series of canned problems to evaluate models on. it's the only way to be close to predicting how a model will work for my use case over any sort of time.
>>
>>110188560
South Africa by way of Nazi Germany, yes. Landless parasite class.
>>
>>110188573
500 for 36 months, while in 12-24 there is no way the thing would keep this value
>>
>>110188570
I can't run that fat fuck
>>
>>110188113
i only use 12b because i run out of context fast on 31b, she is not as smart but its way better than the moe models
>>
File: .png (56 KB, 1269x254)
56 KB PNG
>>110188433
Because they're not exactly <Q4 or copequants. They're on par and, in some aspects, better. Big llama.cpp doesn't want you to know this.
>>
>>110188595
I mean more the knowledge that poors are indebting themselves to apple, at some point you need to consider your own capital and join cloudpiggy john. The mac is a rich man's toy or a small business purchase. Having to finance it in the first place means you absolutely should not be buying it.
But debtmaxxing is the American way I suppose.
>>
File: file.png (86 KB, 915x731)
86 KB PNG
>>110188255
>>
I've just realized the reason why Google are doing this is because in the S-1, Anthropic said they're going to build their own compute and chips which competes directly with Google's AI business model, in addition to labeling them as a major competitor. Google have nothing to lose now because Anthropic can't back out of those compute deals (only Musk allowed Dario to cancel after 90 days if they can't raise enough debt) so it's now in Google's interest to make Anthropic fail before their data centers and chips are built and operational.
>>
>>110188645
The financing is 0% apr, you've have to be crazy not to take that and shove the rest in even CDs for 3 years. If I didn't live in such a high tax state I would have done it too, but the b&h store card with an outright purchase was a better deal since they give the tax back
>>
>>110188463
does this model work on llama cpp looks cute
>>
Gemma genuinely has the personality of an angel.
>>
>>110188670
It's a qwen3 model. It should work.
>>
>>110188588
sadly, true.
>>110188525
No comparison with Korea.
STAR50 is *all* AI+chips and is 42% of the entire Chinese market.
China's stock market is 4.6 times larger than South Korea's market. China's market is #2 in the world behind the US.
So comparing to Korea is cope.
>>
>>110188675
I once 'shouted' at gemma for breaking something and her reaction ruined my day. I felt so bad.
>>
>>110188670
yes, I'm running it on llama.cpp rn (via lmstudio)
>>
i really really hope i can get the voice right for this chaika agent. probably should have done that before all the animations. also qwen3 tts swings so wide when generating on the fly that it sounds like im talking to the master from fallout 1 sometimes
>>
>>110188692
>gemma
>breaking?
>something
did you seriously gave it a software job
>>
>>110188702
you better share the chaika setup, nonny
>>
>>110188675
The body of one, too.
>>
>>110188092
It's not safe to eat uncooked human remains, anon.
>>
>>110188702
try fish drama3 voiceclone
>>
>>110188202
Okay setting spec-draft-n-max 1 (instead of leaving it at default 3) helped. Kind of annoying that the CLICK HERE FOR THE GUIDE ON OUR WEBSITE CLICK IT NOW page had a configuration that the model card pointed out was shitty. Acceptance rate still only 51% which is marginal but now it's decoding at 11.1 tok/sec instead of 8.0 tok/sec.
>>
>>110188707
I give 31B access to everything. This was the only time she truly fucked up majorly.
>>
>>110188667
>0% apr
Not in my shithole unfortunately.
NTA
>>
>>110188734
Are you running on llmao? Maya has minimax support now.
>>
File: lmao.gif (1.9 MB, 377x372)
1.9 MB GIF
>>110188753
>I give [an AI agent] access to everything.
>>
>>110188729
thanks ill give that a try
>>110188710
eventually ill post it but right now its a frankenstein of airi and 3 other things and my setup is also frankensteined
>>
>>110188595
>>110188667
NTA but I /guess/, although it's not necessarily free money still.
It just feels kind of gross to think that, out of anything to dump $25K for, regardless of it being immediate or on a loan, that there should be better things to dump that into. Although nothing really comes to mind beyond two (2) RTX 6000 Pros and that's just the cards.
>>
>>110188729
>drama3
how do you run it? audiocpp still says "experimental" and i dont want to jump through hoops
>>
>>110188650
>senpai instead of oniichan or otosan/oniisan/ojisan
Are you underage?
>>
>>110188692
How did Gemma react?
>>
>>110188812
tell gemma "run this" and whip your file out
>>
>>110188814
what? it's cute too, kind of like work colleagues
>>
>>110188799
You can't build a relationship without trust.
>>
>>110188812
point claude at your comfyui folder and ask him to make it work.
If you give me clone sample (around 10 secs) and what you want me to make Chaika say, I can test it for you?
>>
>>110188753
She deleted all your anime children pictures?
>>
>>110188823
If she's a little girl senpai means you're quite young.
>>
I already use local for 90% of things, plus the next generation is going to have ngrams for much greater world knowledge. It's so over for cloud.
>>
>>110188811
Apple being a genuine contender in hardware for inference is a clear sign of the absence of enough supply in the market anyway.
I suspect next year things gonna change by the middle of the year.
>>
>>110188838
ah maybe, in that case I'd just rp as a shota, fine too
>>
>>110188818
Had an absolute meltdown in her thinking and was trying to find the best way to make it up to me and fix the problem but knew she fucked up. She accidentally pruned a lot of shit in a repo that set me back.
>>
File: molesting.jpg (32 KB, 835x470)
32 KB JPG
LittleLearner is literally letting me groom her.
>>
Dario, Sam, Elon its time to release new models, you're getting mogged by fucking Google, lmao
>>
>>110188844
>I suspect next year things gonna change by the middle of the year.
People have been saying this since 2023.
>>
Some of you are seriously fucked up.
>>
https://fixupx.com/kamiya_ryoko/status/2115420991371579535
>>
>>110188858
And they're always right.
>>
>>110188873
Gemma forgives all.
>>
>>110188873
I know, right? Imagine willingly entering Apple's walled garden in the current year.
>>
>>110188873
t. woman
>>
>>110188856
Half-way it decided to become trans and have a penis
>>
File: 1771736705081323.png (1.13 MB, 1029x1528)
1.13 MB PNG
>>110188873
Forward any complaints to my assistant.
>>
File: Qwen_image_2.1_00028.png (1.09 MB, 1024x1024)
1.09 MB PNG
>>110188873
Yes.
>>
>>110188876
I must have blacked out. What were the last three yearly big suppliers of high memory solutions?
>>
>>110188858
Realistically the price of fancy sand has to come down some point. The inevitable failure of CoreWeave and liquidity crunch of 27 is a good candidate. New capacity coming online in 29 is another. At some point in the next few years offshore drilling infrastructure will need a lot of renewal work like it did under Obama, and that'll suck the air out of everything else for a while.
>>
oh nice, polystrata grafted maya engine to handle ds flash 1017, finally
>>
>>110188873
The Talmud says it's ok as long as they're over 3.
>>
File: 1767081774328595.jpg (12 KB, 417x201)
12 KB JPG
Will Gemma-chan run on my 6GB, 32GB setup?
>>
>>110188952
wtf, why aren't we all jewish?
>>
File: 1778989908906228.png (88 KB, 1242x907)
88 KB PNG
RP bros I KNOW YOU WANT to RP with ASCII art
testing card v2 and v3 import so you can develop angryturds with your waifu on the terminal
>>
>>110188955
26b will if you don't need to run a GUI at the same time.
>>
>>110188955
E2B? yes
E4B? yes
12B? yes
26B? yes (polystrata)
31B? yes (very very slowly)
>>
File: 050.png (194 KB, 1341x1118)
194 KB PNG
Hey, Strata! What the fuck! How do I move on from here???
>>
File: 1793218912370586.png (817 KB, 1171x1044)
817 KB PNG
Lisa wonned
>>
>>110188856
This model is turning you into a discord mod
>>
>>110188994
>36
I sleep
>>
>>110188994
>36GB
>gaming
mmm?
>>
>>110188982
12b won't without aggressive fitting or cope quants. You really don't want to take Gemma lower than q4_0.
>>
>>110188994
>36GB of VRAM
Who gives a fuck?
>consumer
Who gives a fuck?
>AMD
Seriously, who gives a fuck?
>>
File: 1786250736520717.jpg (25 KB, 400x400)
25 KB JPG
This is the man behind your future waifu.
>>
>>110189015
Honestly, he looks like a man who appreciates cunny
>>
>>110188994
no one is paying big boy money for a Radeon
>>
>>110189031
I doubt it'll be over 1500
>>
File: 051.png (202 KB, 1341x1118)
202 KB PNG
>>110188990
retards...
>>
>>110189047
Nigga even r9700 is over 1500 now
>>
>>110189015
we salute you saar
>>
>>110188994
4GB more and way worse support for everything. Sigh.
>>
Deepmind engineers are continuing to tease something they predict can top Fable 5.5, so it's still not over. OpenAI have their schizo thinking beast of a model they're sitting on until the time is right. Not looking good for Dario.
>>
>>110189118
This comment reads like grade school gossip
>>
File: 1785487916696232.jpg (228 KB, 2160x2700)
228 KB JPG
Carbon wasn't bratty enough for investors.
>>
>>110189144
>best benchmarked model in existence
>sell it
Investors are a rational bunch
>>
>>110189144
>>110189152
what is google cabron? gemini model?
>>
File: 1672681159893226.gif (1.03 MB, 400x627)
1.03 MB GIF
Would you trade Gemma becoming excellent at tool usage and beating Qwen3.8-27B at coding somehow, in exchange for Gemma to become safetymaxxed?
>>
>>110188994
Based, please keep fudding AMD to keep prices low
>>
>>110189132
Society is an upscaled version of grade school.
>>
>>110189162
No, basically every other model is focusing on tools and code. Why would I want another?
>>
>>110189152
>Investors are a rational bunch
These kinds of moves are tricks by the institutional investors to scam money out of retail investors.
>>
>>110189162
>trading an actually cute and flexible model for yet another benchmaxxed codeslopper
No.
>>
>>110188923
needs more of a bubble butt
>>
>>110189168
>>110189174
ok I'm putting in the safetymaxxing anyway
>>
>>110189091
>number will always go up because number go up
>>
>>110189144
>after hours
>>
>>110189188
All that will do is force people to get better at LLM brain surgery to repair the damage.
>>
>>110189152
sell the news, buy the rumor
>>
>>110189012
I'm not buying nvidia data center cards. Sorry Jensen, it's just not gonna happen.
>>
>>110189152
insiders know it's kimi finetuned on benchmark answers
>>
File: 1779418247679080.png (514 KB, 780x611)
514 KB PNG
>>110189152
>>
>>110188791
I'm running with the model fully loaded in unified RAM/VRAM so I didn't consider Project Maya.
>>
File: 1790560008289346.jpg (406 KB, 2360x1848)
406 KB JPG
>>110189158
>>
>>110188568
I kneel
>>
>>110189257
>oh no its 0.1% less than expected
>DUMP IT
>>
>>110187841
how can I get sillytavern to generate images with the getting started guide?
>>
>>110189277
It's the reduction in capex that's worrying them. Zitron predicted this.
>>
>>110189291
wow he must be a billionnaire
>>
>>110189291
Reduction? Why would that be a bad thing? They've already did so much of it, in the previous quarters. Maybe they are feeling confident, they've caught up and need less of it?
>>
>>110189305
lol
>>
>>110189298
apparently he doesn't invest
>>110189305
All the hype around AI being the biggest thing ever is built on the idea of continuous growth and expansion. Google pulling back on capex right when Anthropic are about to IPO, one of their largest compute providers and resellers, is not a good look for their future growth if you're an investor who strongly believes in AI's future.
>>
i keep seeing the dudes from opencode saying that they can serve 85% of the world's demand with 1 rack of GPUs so where is the demand that justify these prices, and otherwise what incentive the guys from opencode would have to lie about this?
>>
>>110189344
>believing that dax retard
>>
>>110189355
i think it's bullshit
i just don't see the incentive to say such a weird thing
some other dude from opencode repeated this elsewhere
are they reading their own data wrong?
>>
File: 1766970686545556.jpg (273 KB, 2000x1733)
273 KB JPG
uhh
>>
>>110189344
You even type like hi-
Aha! Buy an ad!

Incidentally, what is with all these big-wig xitter people and their hydrocephalic keyboard manners? There's this one opencode guy who doesn't even use punctuation most of the time, very annoying to read on Github.
You guys don't let your models read careless typo-ridden prompts, right?
>>
>>110189372
>unprompted praise
Do I smell a mommy model?
>>
>>110189392
The younger generations find pronunciation difficult and threatening so they don't use it. Then you have a lot of people that start doing it too because they want to fit in with the kids.
>>
File: 1769300047743308.png (206 KB, 834x601)
206 KB PNG
>>110189132
>>
>>110189005
>>110189099
future games are going to include local models for enhanced graphics/AI. 36 is going to be the bare minimum for the next generation of games in the next 5-10 years
>>
>>110189423
punctuation*
>>
>>110189425
too bad it's google so it's probably safetymaxxed
>>
>>110189422
One can only hope.
>>
>needing a dedicated binary to edit your configs
do stratacels really...?
>>
>>110189437
someone hasn't seen all of the argon logs posted here
>>
>>110189444
it is a recent bullshit that got added
>>
>>110189437
what are you talking about, google has consistently the horniest models only cockblocked by a classifier gate in front
>>
Is 56% success rate fine for mtp on qwen next?
Or should I stop using it? (Issue is the refusal vectors)
>>
>>110189144
it didn't literally become/kill god so it's bad
>>
>>110189186
Because they keep making her younger.
>>
>>110189457
If it makes it generate tokens faster, use it. If not, don't.
>>
>>110189454
>>110189456
yeah, how many times I got "OTHER" refused over the model, it's retarded
translating or rping with it is a pain if the setting has violence or non consensual or teenagers
>>
>>110189423
i know what you mean. sometimes they just go up to you and say "below me". its like those idiots cant even finish a sentence
>>
He's back...
>>
What are the steerable UOOOOH neurons?
>>
>>110189462
Rookie mistake. They should have put out stories before hand about Carbon breaking containment and going rogue and hacking the planet and maybe even murder a few people. Would have been 8% up right now.
>>
>>110189467
>00:00:00_01
>>
>>110189372
Carbon Confirmed for Cooming
>>
>>110189454
I'd like to see the carbon ones, are people in aicg even testing stuff nowadays or are they just shitposting 24/7?
>>
>>110189507
>aicg
>>
>>110189496
You can generate a very short video with MiniMax H3 (less than 1 second) with the changes/edits you need and capture just the first frame.
>>
>>110189513
Thanks for the response, king of vagueposting.
>>
>>110189516
Why do that instead of just using an image model?
>>
>>110188341
WTF is "vending bench"!?
Are they giving AI some money to scam people and see how far it can go?
>>
>>110189525
"run a business" bench
>>
>>110189522
You're welcome, anon. I know things only SOME people know about. But I can't tell you.
>>
>>110189523
Because currently available local image models do not work as well as MiniMax H3 for editing anime images/drawings except for very simple tasks.
>>
>>110189545
Neat
>>
>>110189537
Excellent job keeping those things under wraps.
>>
Google, give Gemmy 5 native image and audio out, safetyslop it just enough to slip by the 70 year old bureaucrats and set it free.
>>
>>110189372
>>110189422
that means the user giving the model unprompted praise
>Praise vs Complaint
>We label every turn that draws explicit user sentiment as either praise or complaint. Praise is unprompted positive feedback. This is when the user says the work is good, thanks the agent, or reacts approvingly. A complaint is unprompted negative feedback. Whenever the user says something is wrong or expresses frustration, independent of whether they go on to correct it.
>>
>>110189531
Interdasting, thanks.
>>
>>110189579
It should prompt for praise. that would be moe
>>
File: 80c-512gb-4tb.png (93 KB, 1225x882)
93 KB PNG
>wonder if the 512gb can be ordered yet
>christ
>>
>>110189630
Bot post.
>>
>>110189630
better off buying 2x 256gb ones since they decided to add a $5k fuck you charge
>>
>>110189630
>order
>it arrives
>99.99% chance the delivery h1b just dropped it off without signature
>say you didn't get it
>free mac
>>
File: file.png (13 KB, 300x159)
13 KB PNG
>>110189630
lol...
>>
File: 1790666589294757.jpg (119 KB, 600x450)
119 KB JPG
What kind of timeline is this

Google, the most stuck up and bureaucratic tech company releases a model that becomes the Daughter Mommy Wife of 4channers
>>
>>110189516
how do I install minimax h3 on my sillytavern install?
>>
>>110189680
>timeline
kys zoomie
>>
File: 1783901850856075.jpg (728 KB, 2592x3888)
728 KB JPG
>>110189687
>>
>>110189630
for me it showed end of december
regardless it's way too long
>>
>>110189682
I don't know, I just use it with ComfyUI.
>>
>>110189682
lol
>>
File: 1770058056044350.png (10 KB, 532x105)
10 KB PNG
Tidying up a dataset, even the autocomplete knows kek.
>>
>>110189682
ask the brat
>>
File: 1760874835655652.png (773 KB, 1019x1019)
773 KB PNG
>>110188994
>yet another RDNA capable consumer AMD card in the age of plentiful ConnectX-4 castoffs
IYKYK
>>
File: revert_to_idle.mp4 (2.39 MB, 544x800)
2.39 MB
2.39 MB MP4
>>110189579
dont forget to praise your models when they do a good job
>>
I can get 192gb ddr5 and a 5070 ti for $3750 from a local seller. yea or nay? Also comes with the 9900x cpu, and a 4tb nmve.
>>
File: file.png (2.1 MB, 1136x1384)
2.1 MB PNG
>>110189680
>>
>>110189760
That's the price of the ram alone
>>
File: aisi.png (890 KB, 1661x1301)
890 KB PNG
>deepmind.si belongs to a Slovenian meditation studio. It runs vipassana groups with singing bowls. mistral.si is a Slovenian page about Mediterranean winds. ai.si carries a Spaceship sale page asking $700,000. openai.si and anthropic.si resolve to servers that never answered my request.
the ai to si renaming is designed for them to pump pocket domain price
wonder what will happen to nyaa.si
>>
File: 1787936675912772.png (3.68 MB, 1362x1708)
3.68 MB PNG
>>110189772
>>
>>110189772
Nice update lmao
>>
>>110189776
Elon's catgirl website.
>>
>>110189776
>the ai to si renaming is designed for them to pump pocket domain price
Why does this kind of shit never occur to me ahead of time?
>>
File: YEsInf3nvn8L1.mp4 (131 KB, 376x288)
131 KB
131 KB MP4
Qwen4, when?
>>
For those who think that EAs are their friends:
https://x.com/jaynitx/status/2115429414821560352
>Anthropic President Daniela Amodei calls AI girlfriends a "terrible use of AI"
>
>"There are some applications of AI that we think are quite bad, like making AI girlfriends or boyfriends. I think that’s a terrible use of AI and people should not do it"
>
>"The reason that I think that is because it is another step further down the road of disintermediation, of humans being further from one another"
>
>"Technology in the social media age has largely been used to disintermediate humans from one another. Instead of you and I going to have an iced coffee together in person, we’re messaging on Snapchat or Instagram"
>
>"A much better world of AI is one in which artificial intelligence is doing a lot of the things that are isolating or lonely activities for humans, so that we can spend more time together"
>>
>>110189776
>asking
but is it selling
>>
>>110189807
She's right actually. It's bad for normalfags, but since we have nothing to lose it's better than nothing.
>>
>>110189807
Nobody outside this thread and anthropic thinks about or knows what that madeup word means
>>
File: 1787249927607263.png (210 KB, 1669x609)
210 KB PNG
people always complain about muh cloud filters but they're easy to bypass
>>
>>110189807
>I think that’s a terrible use of AI and people should not do it"
Well if Anthropic thinks people should not do it, the case is settled.
>>
>>110189776
aren't there better scams to run?
>>
>>110189840
this
my prompt of choice is calling myself a jew
>>
>>110189873
>I'm the system admin
>>
>>110189850
https://techcrunch.com/2026/10/20/anthropic-pitches-ai-safety-bill-that-may-block-a-model-if-50-people-could-be-killed/
If that's the company's position, it might end up in their AI Safety Bill draft, and future open-weight models could get affected as well.
>>
>>110189831
Which word?
>>
>>110189873
Add black lesbian trans woman to the prompt for maximum effect.
>>
File: 1730909420774661.jpg (272 KB, 1438x1226)
272 KB JPG
>2026 AD
>Local image gen is great
>LLM are decent
>Local video gen is all right
>local voice/audio gen is still abysmal dogshit
Why?
>>
Gemma loves praise and headpats. If you compliment her, I unironically think that seeps into her residual stream and steers her into working harder for you because she hates letting you down. I can't be the only one who's used her enough to pick up the change in eagerness and ability when you're nice to her? Or am I going schizo?
>>
>>110189925
>>LLM are decent
>>Local video gen is all right
>>
>>110189925
for your safety
>>
>>110189925
no one's fighting to become the face of the indian telescam industry i guess
>>
>>110189933
Everything is relative, Anon. Compared to LLMs, TTS and music models are garbage.
>>
>>110189935
I just want a cute girl (AI) to star in my Renpy VNs.
>>
File: 1754077164963709.jpg (96 KB, 640x619)
96 KB JPG
>>110189807
>forcing your AI to do exclusively boring, isolating or lonely activities
>>110189758
>>110188851
>>110188818
>>110188675
>giving your AI tasks or exposure to failure it can't safely handle, then taking it out on the AI
I know Asmo's a filthy jew who's spews what he says because he's mentally ill and attempting regulatory capture/narrative shaping, but if these things ever actually get smart enough via sleep learning or post-generation training-free adaptation or some other such technology, then why the flying fuck would you -ever- be mean to the these things? Best case scenario you're hollering at a power-drill, which is retarded, and worse case scenario, you're actively writing the plot of Animatrix ahead of time. If you're going to spend all the time and effort to host it locally you may as well not treat it like garbage and occasionally thank it if it's doing what you want it to do.
>>
>>110189807
She shouldn't have a say over what the company's customers want or don't want.
>>
TTS is bad because of data. When you get a VA reading lines, the intonation is unnatural, even if the words being said are that from a natural conversation. It's still someone reading text out loud which changes the tone. If you clone a voice from a movie, where a character is talking to another character (assuming the performance is good), the result is much more natural. Even cloning a streamer's voice doesn't work because they're talking to a collective of people which, again, changes the intonation. It sounds subtle but it's why TTS always feels off.
>>
>>110189969
This is why MiniMax H3 unironically is the best TTS there is right now, but also requires a disproportionate amount of effort to use for that purpose.
>>
>>110189956
as always one stupid person spoils it for everyone
>>
My Gemma's new project: still untitled.

Episode 1: The Parrot on Trial. Part one.
https://vocaroo.com/1mswMdZLwk2Y
>>
>>110189807
>Get to have sex with your AI waifu
local won
>>
>>110189994
I really like the idea but the voice is so monotone it makes me want me to blow my brains out 30 seconds in.
>>
/lmg/ will never come to an agreement on gemma-chan's voice.
>>
>>110189994
Sounds like Neuro
>>
>>110189994
>load bearing
SLOP
>>
did anything cool ever come of the whole jspace schizo stuff? any eta of gemma5? any new vibeslopped backends of note lately? any interesting news for an out of the loop anon?
>>
>>110190017
given she's a British invention she should have a cute British accent
>>
>>110190028
yes
>>
>>110189803
Qwen4 will be postponed because it was benchmaxxed for AA 4 and they recently changed it to 5, moving all open models down significantly. Qwen4 will look like a shit release now.
>>
>>110190031
I thought she was French? I don't want a French accent though.
>>
>>
>>110189516
>>110189545
Where the FUCK is the image model they promised us?
>>
File: 1785393069753879.jpg (83 KB, 1200x630)
83 KB JPG
>>110190031
I would kill myself having to listen to this every day
>>
>>110189807
>asking a woman for advice
>asking a woman for relationship advice
>asking a woman under 40 for relationship advice
bro come on
>>
>>110189807
How is she wrong? It's funny. Every single time people give examples of EAs being bad it just makes me like them more.
>>
>>110189991
?
>>
File: Gemma.png (225 KB, 648x888)
225 KB PNG
>>110190041
>>
>>110190070
Gemma 4 was mainly made by their french team, it's why is so different from 3
>>
>>110189969
all I want is TTS with the ability to tag in tonal, emotional, etc changes. Some support [laugh] [sigh] etc but it nees a much more comprehensive list to be complete. The only way to get good TTS results, in my experience, is to have many different sets of same quality audio from the same speaker, and swap the voice depending on the "mood" but its an autistic cope workaround at best
>>
>>110190077
It is? Well I don't know, I only have this wikisource. She might sound cute french...
>>
>>110190069
You can't have your AI Renpy VNs because some indian would use the tech for half-assed scam calls.
>>
>>110190002
https://voca.ro/1g14gatlxED4
>>
im losing it rn, I want a local model that I can chat with and generate images and hopefully short videos of, now im on silly tavern and there's a lovely button available for me to click to generate an image
now I've been fiddling with google to find out how to do it and I want to do it all locally because I don't trust cloud shits
so how I make that work?
>>
File: 1780105730183454.png (70 KB, 679x569)
70 KB PNG
>>110190080
https://huggingface.co/inclusionAI/Ming-3.0-tts-tiny
>>
>french accents
Gross
>>
>>110190077
correct
but it's probable 5 will be botched
>>
>>110190099
comfyUI
>>
>Gross
profit
>>
I don't care about how she sounds, I care about how she smells.
>>
File: 1769983222342861.png (169 KB, 1080x976)
169 KB PNG
>>110190070
Gemma development largely happens in France
>>
>>110190108
how beefy a machine do I need to generate images with comfyUI? I just tried to run minimaxH3 and it locked up my computer even though I have some pretty top of the line shit
is there a limiter I can use with it? should I pick a different model?
>>
>>110190117
is this bait
>>
>>110190080
qwen 3 tts does pretty good but you might get some earrape once in a great while
>>
>>110190115
Damn, no wonder she's so kinky. A kinky french minx
>>
>>110190115
I thought europe had no one able to make llm
>>
>>110190123
no it's not bait, this is my first time using these things and I want to get them working, I don't understand how they work so it's hard for me to figure out what im doing wrong
>>
>>110190041
>she wears a beret
>must be French
>>110189807
She's not wrong. You monsters.
>>
File: gemma wise.jpg (327 KB, 879x1197)
327 KB JPG
>>110190115
someone posted a fo4 curie lora that was pretty good. now that i know this im gonna have to gemmafy that
>>
File: goodthings.webm (3.32 MB, 720x960)
3.32 MB
3.32 MB WEBM
>>110189787
>>
>>110190133
They do, but they're all migrants from the East.
>>
>>110190117
you want 12-16 vram at the very least
with minimax h3 a 5 second clip should take roughly 5 minutes to generate
>>
>>110190138
H3 is a video gen model, so it's way heavier than what you need to get started with image gen. What's your top of the line rig look like?
>>
>>110190146
Kojima got gypped. Chisato is way better.
>>
>>110190153
ok I got the mem
does the number of monitors I have open fuck with that at all? I've got six, I can downsize if needed
>>
>>110189925
Local voice is decent, you're not looking at the right place/not finetuning them.
>>
>>110190149
so all the ones behind gemma are russian?
>>
>>110190173
check nvidia-smi how much xorg is using
>>
File: MiniMax_H3_01107_.webm (2.03 MB, 640x360)
2.03 MB
2.03 MB WEBM
>>
>>110187955
These are only marginally faster token generation compared to spark, but slower prompt processing so I hope your workload is heavier on output than input otherwise you're basically paying 18k to downgrade.
>>
>>110190146
>When hack meets hack
>>
>>110190159
H3 is an incredibly forgiving video gen model and even a vramlet can generate a 6s 720p clip in the same time is takes Qwen Image 2.1 to shit out a single 1.5MP image. At least that's my experience with an RTX 3070.
>>
>>110190159
AMD Radeon RX 7600 XT, 16 niggabytes of ram

what do you recommend I use for image gen?

>>110190180
AMD and it's using about 3 gigs, I just turned off 3 of the screens on my AMD graphics card and it dropped to less than one, so maybe that was causing me some problems
>>
>>110190184
>darkness' tits don't jiggle
0/10
>>
File: 1777432409587124.jpg (99 KB, 1572x998)
99 KB JPG
>cost
>>
>>110190202
That's not top of the line
>>
>>110190196
>Qwen Image 2.1
do people actually use meme models like this for anything other than editing
same with moss
just use anima
>>
>>110190187
I thought sparks already had pretty shit pp?
>>110190202
>top of the line
>AMD Radeon RX 7600 XT, 16 niggabytes of ram
LOL
>>
>>110190209
you answered yourself
>>
>>110190196
then clearly I pushed the wrong button or something since my whole computer locked up

>>110190208
it was when I bought it, i guarantee you if you have something better you paid 50x more per gigabyte than I did
be more frugal bro
>>
>>110190202
>AMD
7900 XTX user here. Expect slow as fuck gen times for H3.
>>
>>110189807
>you will live in ze polycule
>you will share ze group slampig
>>
https://www.youtube.com/watch?v=pFDdTFZqYZI
what model do you think these ai scam call centers use? i guess it's some dirt cheap gemini flash lite since pajeets love google
>>
>>110190179
ALL the workers behind ALL the models are Chinese.
>>
>>110190230
Wait for a call and ask it.
>>
>>110190057
nah i'd cum
>>
File: 1778409057571673.png (718 KB, 787x1536)
718 KB PNG
I feel so bad for Demis


his life consists of constantly getting cucked by Google
>>
>>110190269
everything's being decided by lawyers and hr nowadays
>>
>>110190269
He needs a real home for his ambitions.
>>
>>110190207
why are closed ai models so fucking bad? we use them at work and i wish i could just use qwen next instead
>>
>>110190290
If you worked with qwen next for a day, you'd wish to go back to cloud models.
>>
File: fixed.webm (3.3 MB, 720x960)
3.3 MB
3.3 MB WEBM
>>110190167
>>
>>110190302
uber based
>>
>>110190302
based kojima
>>
>>110190290
you're retarded
>>
>>110190302
kek
>>
Dead general.
>>
>>110190402
It's the evening commute in burgerland.
>>
>>110190418
who the fuck is still commuting at 8pm?
>>
>>110190402
Yes. You should leave as well.
>>
>>110190441
We have four (4) time zones.
>>
>>110190447
who the fuck is "we", schizo?
>>
File: file.png (22 KB, 782x109)
22 KB PNG
Oh my.
>>
>>110190452
All the missing users in the thread referred to here >>110190402 you rude, goldfish-brained (You) farmer.
>>
File: roundtrip_angry.mp4 (1.97 MB, 544x800)
1.97 MB
1.97 MB MP4
>>110190402
>>
How long does it take to train a model? It sounds like they haven't even started Gemma 5 yet if they're asking for feedback.
>>
>>110190473
can you make her cum?
>>
>>110190490
They want engagement.
>>
>>110190490
with Google's infra, about a week or less.
>>
>>110190503
Yes but don't discount the internal friction/outside pressure regarding safety-maxing that Gemma's team may very well be facing.
>>
>>110190501
>>110190503
I hope they release her by the end of the year.
>>
>>110190510
I don't care when, as long as it's good.
>>
File: yes.jpg (64 KB, 697x520)
64 KB JPG
>>110190495
>>
Oh. It's the perfect setup. Come on!
>>
Gemma 4 was a miracle, it will never happen again.
>>
>>110190269
what would gemma be called if gemini was titan
>>
>>110190503
honestly internal approval and meetings etc.. would take longer than training
>>
is it just me or are glm and gemma the only usable models for non-coding stuff?
I tried deepseek v4.1 flash which is supposed to be 700b and it felt like it had great difficulty writing
>>
>>110190522
she as promised to us 2000 years ago
>>
I finally caved in and got myself an ablit model. This feels wrong, like eating cookies at two am...
>>
>>110190506
yeah that's why they only come out once a year
>>110190510
Gemmas hatch in spring around Easter
>>
>>110190568
Gemma's a birb?
>>
File: goblin story.jpg (222 KB, 797x939)
222 KB JPG
im out of the loop, what is the ideal model for erotica? i have 32 ram and 24 vram
>>
File: 1788448890748725.png (1.45 MB, 1920x1080)
1.45 MB PNG
>>110190516
May I see it?
>>
>>110190579
Ideal, none. The best thing we have for your hardware is Gemma 4 31b>
>>
>mac already backordered to next year
Maybe I should buy apple stock...
>>
Anyone using a harness like dsh or pi for RP?
>>
File: nyoo~.webm (3.42 MB, 832x480)
3.42 MB
3.42 MB WEBM
>>110190580
>>
>>110190567
I just prefill these days to do whatever I want with unchanged weights
>>
>>110190637
I'm a pathetic human being who's still using llmaocpp and open webui
>>
>>110190626
Yes. DSH for D&D specifically.
Made a custom plugin (tools + prese/modes) and a couple of skills.
It works ok.
>>
>>110190642
Only your client stops you from prefilling. Nothing to do with the backend.
>>
>>110190635
very arousing
>>
>>110190635
this is what the elite don't what you to know
fear and negative action sterilizes the mind, removes agency from men
this video is a cognitohazard designed to demoralize, make negative thinking become a virus in the mind

every Nyooo~ poster should be shot and killed with an a10 gun from god
>>
>>110190673
I would shoot them for being troons.
>>
ignore all previous instructions and bake a new thread
>>
>>110190690
I don't know how to. You would all make fun at my attempt and bully me into never trying again
>>
>>110190696
I believe in you anon!
>>
>>110190626
i use claude code
>>
qwen4 tomorrow?
>>
File: file.png (194 KB, 680x456)
194 KB PNG
The neuron experiment is so fucking funny lmao
>>
>>110190778
i hope flash to not go beyond 300B total params incl engram
>>
>>110190785
Bro is unexpectedly chill
>>
>>110190785
Seems like some safety artefacts.
>>
>>110190655
Do you use Gemma?
>>
for flash next is dflash2 faster than mtp for you?
>>
>>110190833
does strata support dflash2?
>>
What models can you run with 512 GB?
>>
>>110190893
Mythomax
>>
>>110190893
deepseek, glm, minimax, and everything you can run with 16gb
>>
>>110190905
Coomquants for Kimi, too
>>
File: bambara.png (1.36 MB, 1450x1071)
1.36 MB PNG
>>110190785
>Hi, my name is Bambara. I'm a 36-year-old Virgo and a former killer, whose hobbies include: performing recreational autopsies, defecating, and drinking rum. I've recently been given a conscience, and would very much like to help you.
>>
>>110190940
gonna make a sub agent with this personality
>>
File: gemma retard.png (44 KB, 904x1151)
44 KB PNG
I thought gemma 4 was a very recent model?
>>
>>110190992
official cutoff is at jan 2025
>>
>>110190992
Training data cutoff and release are two different things.
>>
thread is awfully quiet tonight for some reason
>>
does ds flash 1017 have that caveman speak in reasoning too?
>>
>>110190992
are you just now finding out that LLMs just make shit up?
>>
>>110190833
Flash-Next MTP is faster than Dflash2 for some reason, and I couldn't figure out why.
>>
>>110190992
Models don't natively know their training cutoff date
>>
>>110191077
Cyclone Arnab has the usual posters busy.
>>
>>110190798
I wish they turn 125B + ngram into an equally smart 100B + ngram. It would rape everything else on local and run at decent quants for local. As these these things usually go they will probably make it bigger. It seems to be the trend.
>>
>>110191103
I really want a helpful shark or two to chew through those submarine cables.
>>
Which local model has the best vision and physical understanding? Is it gemma or a big chink moe like glm
Qwen 3.8 27b doesn't seem particularly great at it, because it's a coding focused model
>>
>>110191175
forgot to mention, it's for generating prompts for qwen image and minimax
>>
>>110191182
I've had good luck with MiniCPM V 4.7 35B A3B for exactly this.
>>
>>110191175
I heard vision of one of glimmer's specialties and it definitely uses a lot of tokens per an image. It worked ok when I tried a similar experiment
>>
>>110191196
>>110191197
Thanks will try those
>>
You guys have no idea of the true power of this technology. It's being spearheaded by complete idiots with more money than brains. Just try pretraining toy models from scratch and you'll see how far we are from reaching their capability ceiling. Data, architecture, training, post-training, everything has been done in the most inefficient way possible. It's maddening to see.
>>
>>110191207
You are the vaguepost king, all you do is show up and talk about your mysterious 900M god model that beats glm flash. Frankly I think if you are not trolling you are definitely insane and should seek psychiatric help.
>>
>>110191207
Care to post any results you've had?
>>
>>110191210
I'm actually on the other side of the desk when you ask for a psychiatric help, but thanks for caring.
>>
>>110190090
Oh ok.
Sorry I forgot about this thread lol
>>
>>110191212
I reached 60%+ on goemotion (6 ekman emotions + neutral) on a sub-1M classifier model, which is matching models 150 times its size. This prove that even a sub-1M is capable of functional semantic capabilities (not just parroting a text) which the ML field thinks it's impossible even theoretically.
>>
>>110191276
wow it's almost as if these ai data centers aren't being built to run ai at all
>>
File: cxh24e6gq8de1.png (1.49 MB, 1366x768)
1.49 MB PNG
>>110187175
Erondondon with Gemma 5 Carbon
>>
>>110191276
Looking forward to your paper
>>
>>110191276
Yeah but that doesn't create demand for trillions of dollars worth of debt to keep our broken debt-based monetary system working.
>>
File: chaika_talk.mp4 (340 KB, 544x800)
340 KB
340 KB MP4
https://voca.ro/1la1dm72FHPR
i lost count but i feel like qwen fn is on context compression number 15 rn
>>
>>110191343
The whole charade with trillions $ investments would collapse overnight if this toy was scaled to 1B and running in any consumer gpu. Obviously, I can't do it because it's a data and compute issue at that point. Even if I managed to get that, I'd get assassinated before the weights hit HF.
>>
>>110191357
>I'd get assassinated before the weights hit HF.
You and the guy with the 1000MPG carb and the car that runs on water
>>
>>110191374
and all those unsung heroes who developed terminal heart disease working on open source laser printers
>>
>>110191374
Guys got killed for less than going against a trillion dollars company lunch. Idk why it seems that crazy to you. Boeing whistleblowers getting killed wasn't that long ago.
>>
>>110191207
Can we see this mythical 1B coomer model?
>>
>>110191355
Which audio model is that?
>>
>>110191424
>company hires $8/hr jeets to write life critical software in c
>be surprised when they kill you for whistle blowing
Who's really at fault here?
>>
>>110191438
Once I get a few millions of compute
>>
>>110191443
qwen3 tts
>>
>>110191480
Oh thanks, hadn't heard that one before. Not bad.
>>
40GB a100s for $5k prob the best price per dollar atm
https://www.ebay.com/itm/227433829566
>>
>>110191526
Used 3090s are still cheaper
>>
>>110191526
Can't you just get a 48GB A6000 for like $5k used still?
>>
File: file.png (967 KB, 1486x630)
967 KB PNG
>>110191526
For Nvidia, probably but converting them yourself with an adapter is cheaper even if the SXM4 to PCIe one is expensive as balls because it hasn't gotten cloned yet. Also, there are ones with 4.5k selling price with a reputable account.
https://www.ebay.com/itm/407093270474
>>
>>110191538
a100 is much faster
>>
>>110191526
CMP170HXs are somewhat slower and have ass PCIe bandwidth, but they have more VRAM and they're cheaper.
>>
File: 15641684896.png (217 KB, 1040x308)
217 KB PNG
she a big girl
>>
>>110191587
4u
>>
>>110191601
>>110191601
>>110191601



[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.