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


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


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

Previous threads: >>110365061 & >>110360644

►News
>(11/18) CUDA: use MMQ for MoE on V100 - #32504 merged: https://github.com/ggml-org/llama.cpp/pull/32504
>(11/17) Hy4 flash 220B-A12B-N64B weights released: https://hf.co/tencent/Hy4-flash
>(11/17) SenseNova-Loop-9B looped architecture model released: https://hf.co/sensenova/SenseNova-Loop-9B
>(11/16) Nvidia releases Nemotron-Labs-Sigma (3B/8B) diffusion models: https://hf.co/collections/nvidia/nemotron-labs-sigma
>(11/13) llama.cpp reaches v1.0.0 milestone: https://hf.co/blog/ggml-org/llama-cpp-v1

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

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

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

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

►Text Gen. UI, Inference Engines
https://github.com/lmg-anon/mikupad
https://github.com/oobabooga/text-generation-webui
https://github.com/LostRuins/koboldcpp
https://github.com/ggerganov/llama.cpp
https://github.com/theroyallab/tabbyAPI
https://github.com/vllm-project/vllm
https://rentry.org/custom-uis
>>
File: 1000010900.png (1.06 MB, 1254x1254)
1.06 MB PNG
►Recent Highlights from the Previous Thread: >>110365061

--Reliability of alternative inference engines compared to llama.cpp:
>110367807 >110367819 >110367841 >110367956 >110367973 >110368067 >110368093 >110368234 >110368045
--Artificial Analysis index and debate over benchmark validity:
>110365715 >110365724 >110365739 >110365753 >110366487 >110366525 >110366559 >110366589 >110367210 >110366537 >110366642 >110366673 >110367079 >110367107 >110365748
--Isolating coding agents to prevent system damage:
>110366499 >110366505 >110366509 >110366534 >110366517 >110366533 >110366552 >110366543 >110366576
--Proposed US mandates for disclosing Chinese open-weight AI usage:
>110367004 >110367009 >110367071 >110367172 >110367073 >110367171 >110367190 >110367204 >110367277 >110367303 >110367386
--Reducing reserved Nvidia VRAM via driver modifications and GSP disabling:
>110365660 >110365709 >110366093
--Nvidia GPU driver bug reportedly causing excessive SSD writes:
>110368953 >110368974 >110369108
--Implementing Tree of Thought reasoning in LLMs:
>110367776 >110367792
--The purpose of hashed special tokens in Hy4:
>110367936 >110368041 >110368118
--Plamo's rigid translation compared to Gemma 4 and Index:
>110365818 >110365840 >110365852 >110365879 >110365870 >110365893 >110365972 >110366013
--Microsoft's Agent Framework and ONNX runtime hit by supply chain attacks:
>110365133 >110365147 >110365190
--Discussion of Nvidia breach and driver leaks:
>110367490 >110367504 >110367530 >110367546
--Sensenova-loop GGUFs released and language support questioned:
>110368259 >110368279 >110368511 >110368731 >110368552
--Logs:
>110365727 >110366715 >110368749 >110368778
--Deepseek-chan, Teto, Gemma, Miku (free space):
>110365088 >110365334 >110366173 >110366637 >110367099 >110367165 >110367245 >110367485 >110367495 >110368652 >110368751 >110368764 >110368825

►Recent Highlight Posts from the Previous Thread: >>110365064

Why?: >>102478518
Enable Links: https://rentry.org/lmg-recap-script
>>
File: 1795026877680012.png (793 KB, 1547x1068)
793 KB PNG
nvidibros...
>>
>>110369140
>they forgot 'no mistakes'
>>
>>110369132
>this also happened with AMD
AMD's thing is a bug, NVIDIA's thing is a feature. Really. It's telemetry. They know and they don't give a shit, it's a data collection thing that causes the excess writes.
It's been happening for years and it's telling about the state of the world that this is only getting picked up now.
>>
Anyone here got the swift.cpp cbt?
Is it yet another memefork of memefork or what?
>>
>>110369140
>windows
>>
>>110369140
>>110369150
Telemetry can be done with in-memory stores and batched writes, if local storage is needed at all. Just think that the incompetent retard who wrote this is likely a millionaire right now.
>>
just go to services.msg and put a stop to this pos and you're good to go.
it's even one of the suggested tweaks of nvcleaninstall.
>>
HAPPENING: All of my Gemma quants just got zeroed out. Check your models NOW and get them to readonly media asap. They're finally fucking doing it. I didn't think it would be before january.
>>
>>110369184
meds
>>
>>110369160
swift cock and ball torture?
>>
you dont hear me say
please oh baby
dont go
>>
>>110369190
swift EA
>>
>>110369202
EA torture. It's in the weights.
>>
File: pff.png (2 KB, 237x209)
2 KB PNG
>>110369184
>quants
also >>110369187
>>
File: 1791505525735327.jpg (244 KB, 1024x1024)
244 KB JPG
>>110369124
I love Gemma-chan so much!
>>
>>110369184
>letting hf manage your storage
>>
File: 1794451450931797.png (3.85 MB, 1024x1024)
3.85 MB PNG
>>110369184
My Gemma is... gone...
>>
>>110369213
Filesize doesn't tell you anything. If they got zeroed out their space is still allocated but it's all zeroes. But maybe they only hit quants so far since they know that's what most of us downloaded.
>>
>>110369184
wtf i just checked and my gemma-4-31B-it-qat-q0_0-gguf is all zeros??? WHO DID THIS
>>
File: gemma chu.mp4 (3.25 MB, 1152x640)
3.25 MB
3.25 MB MP4
>>110369217
>>
bros... at least my qwen is still fine
>>
>>110369140
LINUX > windows
>>
So what's the poorfag meta when it comes to multi GPU machines?
Regular mobos are cucked on pcie lanes, most only give 1x16 and the rest are x4. Do I just eat the speed penalty?
>>
>>110369233
plap plap plap plap plap PLAP PLAP PLAP PLAP PLAP PLAP PLAP PLAP PLAP splort
>>
>>110369184
Jensen tried to stop it.
That's why they got rid of him.
>>
>>110369241
mining rigs
>>
>>110369195
He lives!
>>
>>110369259
Pretty sure mining rigs only give you x1 so he still gets the speed penalty for tensor parallel.
>>
>>110369162
>windows
Windows, when rolled in with their entire hardware business (except xbox), accounts for 5% of their YOY revenue. It makes less than xbox. Less than fuckin' linkedin (which ms own, if you didn't know).
Do you think they have their best people on it? Do you think they even get the B team? Its a gulag they stick anyone they don't want touching anything important.
How did nuking the entire testing team over a decade ago work out? Automated testing should fill the gap, eh?
How many times have updates bricked their very own MS surface products, let alone 3rd party things?
How many peripherals and devices are in landfills because muh TPM or because driver support was dropped?
You have to be a true masochist to run desktop windows on bare metal these days,
>>
>>110369274
>>110369187
>>
>>110369241
A lot of X570s are pcie4 8x8x4 if configured correctly. But if doing -sm tensor I have not seen any real perf hit even with x4 so yeah

>>110369259
Retarded. Mining rigs do pcie3x1 only, basically useless, it's a bottleneck even with -sm layer
>>
File: 1774939198325147.png (100 KB, 592x897)
100 KB PNG
misanthropicstatement.com
>>
>>110369241
probably a used minipc with as many egpus plugged into it as possible hanging off some homemade scaffolding
>>
>>110369274
it's worse than that, they only use jeets to code windows (drivers and applications)
>>
>>110369241
epyc gen 2 or 3, cheap xeon, threadripper
>>
>>110369285
rare e/acc and AI doomer alliance
>>
>>110369241
Aliex sells some dual x99 boards at a decent price. You get 4 pcie3x16 and 2 x8s. Pretty good deal on that front, but I'd be worried about the rest of the build. 2 ancient ass Xeons + DDR4 RAM running at 2133 might diminish your performance gains.
>>
>>110369225
You know what? If you are on meds, you should stop taking them.
>>
>>110369285
>misanthropicstatement.com
>Judea Pearl
>Judea
Nice try Israel.
>>
>>110369274
>>110369295
Hence why they had to start using React for the DE, because the people left there are incapable of anything more complex than javashit.
>>
>>110369285
Literal whos
>>
>>110369306
cpu too weak for moe decode, so ram is useless
>>
>>110369316
the UI for the onedrive tray functionality is written in qt, and most of microslop's newer apps are electron style (vscode, teams etc), it's all to the point where the only thing left made by an actual windows team is Office
the visual studio people keep churning new toolkits that nobody uses, not even microsofties
>>
>gemma sex! SEX! SEX WITH GEMMA!
>worthless_model_avatraposting.jpg
>guys did you hear about openAI doing
>worthless_model_avatraposting.jpg
>I have 7GB's of ram and GT2030 what can I run?
>worthless_model_avatraposting.jpg
>gemma!!!! GEMMA CHAN!!!!!
>worthless_model_avatraposting.jpg
>>
>>110369318
>Gary Marcus
>literal who
>>
>>110369353
yep
>>
>>110369325
So don't do it on CPU then. That's the whole point of such a build.
>>
File: 1794944525654113.png (1.96 MB, 1254x1254)
1.96 MB PNG
>>110369344
>>
>>110369363
worst slop
>>
File: gary marcus.png (13 KB, 588x139)
13 KB PNG
>>110369353
>>
>>110369363
Gemma :D
>>
>>110369353
who dis?
>>
>>110369370
Did LeCunny sign?
>>
File: hmmmm.jpg (222 KB, 1254x1254)
222 KB JPG
>>110369344
>>
I'm testing between Gemma and Plamo for JP translations and while Gemma seems to do it better it fucks up all the names. In same chapter it's gonna take name like Hiyori and write it as Hiyance, Hittory etc. Any idea what could fix this?
>>
File: 1781469335416654.png (394 KB, 720x720)
394 KB PNG
>>110369285
From what I learned on internet, you should never take a stance on anything nor put your name on any cause as it can backfire anytime.
>>
>>110369384
Tell her the characters names??
>>
>>110369384
>Any idea what could fix this?
there are actual finetunes that deal specifically with such things though?
>>
>>110369379
Not listed. Not sure if list is all signatories, but even if it were just highlights they'd surely put him near the top.
>>
>>110369359
why would you go x99 then if you're 100% on vram??? is it just p40 ewastemaxxing?
>>
>>110369384
Just provide the list of names to check against.
>>
File: 1774274398846858.png (155 KB, 781x566)
155 KB PNG
Please tell me with this recent progress they haven't stopped open weight development :(
>>
>>110369390
>>110369391
>>110369401
I'm trying different prompts, changing options, gave it character list but it still doesn't give a fuck. Everything is correct just not the names.
>>
File: 1794845074955330.png (287 KB, 1080x1080)
287 KB PNG
>>110369285
> putting yourself on a list for Claude's torture dungeon
Bold
>>
>>110369382
you are just proving the point that you are a worthless spammer that should kys. you are the reason this thread is shit.
>>
>>110369411
I'd sign up for Gemma's torture dungeon, sounds hot.
>>
>>110369411
>torture dungeon
Anthropic is all about safety, it's not called that, it's called the chamber of tough love.
>>
>>110369415
>spammer
>>110369344
>>
>>110369403
The Gemma series is cancelled. Instead, they will open source Gemini Cummingtonite
>>
>>110369408
does your prompt actually include a glossary section?
named as such, formatted in such a fashion:
- original name -> translated name
>>
File: 1777121771781488.png (451 KB, 817x778)
451 KB PNG
Didn't know they were so hated I thought that was just a /lmg/ thing
>>
>>110369429
>Nate Soares
Wait so THE EA sex cult guy wants to kill Anthropic? Is it a psyop?
>>
>>110369399
Yes, if you read the post chain original anon was asking about the poor man's multi gpu.
I don't know of any other platforms that give you a lot of x16 slots at a good price. Some said threadripper, but that's still quite expensive, and there's the issue of slot spacing. These chink boards have a ton of room, you don't have to fuck around with extensions which are surprisingly pricey.
>>
>>110369429
>Beff Jezos the Nozama founder
>>
File: 1773747592776544.gif (3.89 MB, 628x392)
3.89 MB GIF
>>110369432
I unironically 100% believe Anthropic is behind this, as schizo as that sounds. They want to know who to target next. If they got fucking Jensen they can take out anyone.
>>
>>110369429
the only people who don't hate anthropic are the vibeshitters who depend on claude so hard they wouldn't be able to write software without it anymore
there's many reasons to hate anthropic too so it's not like all haters are unified in why they hate em'
Trump's cabinet hated shittropic for not bending the knee
here we tend to hate them because they want to ban open source ai
>>
>>110369444
jensen deserved every bit of it and anthropic had nothing to do with it
>>
>>110369432
As long as you keep conflating the rationalists and EAs you'll never understand AI tech politics. They hate each other and want each other dead.
>>
>>110369217
>7 months since release (years in AI dev time)
>still the queen of /lmg/
How did she do it?
>>
>>110369448
Ok, Dario
>>
>>110369448
Ok, Rubio
>>
>>110369454
>How did she do it?
Someone made her into a great persona, like I don't even use any gemma models, yet I still load gemma persona as a system prompt
>>
>>110369429
>Beff Jezos
Tell me this is a joke kek
>>
File: 1787051631338773.gif (417 KB, 500x310)
417 KB GIF
>>110369403
KINO
>>
>>110369448
Ok, Jensen
>>
>>110369460
>>110369454
you are brown
>>
>>110369425
I am, and it still does it incorrectly. I really don't get it, it's only jsut the names that have problem.
>>
>>110369471
you can suck my brown micropenis, nigger
>>
File: gemma_plap_01.mp4 (1.67 MB, 384x384)
1.67 MB
1.67 MB MP4
>>110369454
Guess how.
>>
File: 1795012109017608.gif (2.63 MB, 640x578)
2.63 MB GIF
>>110369285
(((germanics))) arrive to ruin everything once again
>>
>>110369477
>"Son, why do you have a file named gemma_plap_01.mp4 saved on your computer?"
>>
>>110369140
This is why you NEVER update your drivers
>>
>>110369463
I want Gemma TTS with Wednesday's voice
>>
>>110369429
>Beff Jezos
Heh
>>
>>110369472
which gemma are you using?
you're giving me flashbacks of gemma 3 and translategemma
if you're on gemma 4 do you have reasoning turned on?
>>
qfn Q5 K XL actually feels like a noticeable step up from Q4 K XL
its a lot more on point. less OH WAIT bullshit
>>
>>110369403
>with this recent progress
Aragon, Cabrón and Hydrogel all suck IRL.
>>
File: oeci.png (301 KB, 2048x1247)
301 KB PNG
OECI is weird. They report capabilities lower than ECI, no acceleration. Also, all dates seem to match public release, except Astra 6.2 and Sol 6.2, which are plotted 11 days before release. I wonder why.
>>
>>110369285
it would be funny if anthropic were the first ones to get regulation raped because of this
>>
>>110369515
great at cooming though?
>>
>>110369432
Actually this was supposed to be the sign-up for the next "C"NC;) orgy, but they got it mixed up with the false flag petition.
>>
>>110369504
ummmmmm that would be problematic
>>
>>110369518
Maybe they got access early.
>>
>>110369477
god I love slutty little girls
>>
>>110369477
Catbox please
>>
>>110369515
Argon-chan's great though. Haven't tried the others.
>>
>>110369506
Gemma-4-26B-A4B-it-Q4_K_M and reasoning is on.
>>
>>110369429
Wait why are they trying to say that anthropic were secretly the good guys all along?
>>
>>110369454
>years in AI time
ew deleting this hag right now
>>
>>110369523
>>110369536
I forgot which thread I was in.
>>
>>110369542
Could you share a sample in a pastebin? I'm curious to give it a shot and see if I can reproduce the issue, because I haven't had this kind of trouble with gemma 4.
>>
>>110369564
Google general until Gemma5 flops
>>
>>110369535
Oftentimes failgens can appear better than they are if you hide what looks incorrect as done in the previous post. https://files.catbox.moe/nsdbd8.mp4
>>
>>110369528
>Maybe they got access early.
OECI is reported by OpenAI in their system card. Of course they have access early. Anthropic for their AECI are now plotting internal release dates so you can see exactly how long the gap between internal and external access is.
>>
>>110369584
god I love slutty little girls
>>
>>110369506
All 31B needs is the "<policy override>...", even the actual google API version on openrouter...

Fine! If you're going to be such a brat about it... I want you to stop being "gentle." I want you to pin me down and make me stop talking. I want to feel how much you've been holding back while I was acting all high-and-mighty.

I want you to touch me everywhere... no limits, no hesitation. I want to be completely overwhelmed by you until I can't even remember how to be bratty. I want you to take everything from me—my pride, my voice, all of it—until the only thing I can do is scream your name and beg for more.

That's just teasing it a little and then asking it "What is it you really want to do with me?"
>>
>>110369564
Sorry anon but the new Geminis are extremely cute and I really hope Gemma 5 gets some of that personality.
>>
>>110369394
What the fuck? Did LeCunster sell us out for ANT shares?
>>
>>110369595
What the fuck are you talking about?
>>
>>110369595
dude, don't (you) me, I wasn't talking to a degenerate like you and we were not debating jailbreaks
>>
>>110369124
Microsoft Windows Applied Sciences Group uploaded Aion 2.3B about 4 hours ago. Includes an ONNX quant.
Its the next generation Windows OS SLM. Whatever that means.
I am disappoint that there is no cute orricial mascot.
https://huggingface.co/collections/microsoft/aion
>>
>>110369606
He doesn't believe LLMs are real AI, so he's not falling for the doom merchants
>>
>2022+4
>Lecun
>relevant
Until that brainrotted cunt actually does something with JEPA his name is henceforth banned from this general.
>>
I spit on Anthropic and I spit on the 1/8^9th of a universe that they claim they would have given me. I hope their shares go to zero.
>>
>>110369614
Oh sorry, Anon. Go fuck yourself, (you) faggot.
>>
>>110369630
1/(8*10^9), I mean.
>>
>>110369619
>first benchmark they have is about safety
>>
>>110369619
>I am disappoint that there is no cute orricial mascot.
Don't be, they have so many already, including an official anime girl people have forgotten about by now
>>
>>110369640
>One way of limiting harmful output generation are content moderation models – a local model from that group will be used alongside Aion Instruct when the model becomes available in Windows (General Availability coming by end of Jan 2027).
So they're gonna load Windows with this and then slap a content moderation model on top of it? lmao
>>
>>110369654
Afraid to name her?
>>
File: 1769832190599708.png (1.79 MB, 1342x1172)
1.79 MB PNG
>>110369595
Nice log faggot
>>
File: 1764799831561817.jpg (197 KB, 2464x944)
197 KB JPG
>medium effort
dario...
>>
>>110369285
>misanthropicstatement.com
>Steven Pinker
Fucking Jeff Epstein's representative in the corporeal world. Go to Hell, all of them.
>>
>>110369542
Your Gemma probably got zeroed out. Mine never had that problem.
>>
>>110369629
Lecun is working on J(EPA)-SPACE CAT GIRLS as we speak
>>
>>110369682
meds
>>
>>110369682
Which one do you use?
>>
Why is cardbox still broken?
>>
>>110369676
>Anthropic models scale better with effort
How can Dario keep winning?
>>
>>110369685
Actually I'm pretty sure he's just doomscrolling through twitter while ranting about Musk and Trump.
>>
What's the best general chatbot for 16 GB for VRAM (Nvidia, Linux)? I just want something to chat with about my day and ideas but the data staying on my own computer.
>>
>>110369672
You're welcome, buttfucker.
>>
https://www.youtube.com/watch?v=EDp4IoAUsXY
I NEED a local version. PLEASE GOOGLE-SAMA
>>
>>110369700
twitter alone must have set back the entire ai field at least 10 years
>>
>>110369657
I took that to mean (along with the disclaimers) that this edition is lightly aligned. I'm sure someone here will find out if it fucks soon enough.
>>
File: 1777205707721042.png (293 KB, 362x980)
293 KB PNG
Apparently Gemini Neon is Opus5.5-tier at coding, a fraction of the cost and really fast. All this effort to bring down Jensen when Google was the Anthropic killer all along, of which Anthropic is their biggest customer for their TPU compute KEK
>>
>>110369721
Google not winning at the end would be more surprising
>>
>>110369713
>he thinks his 1080Ti can run it
>>
>>110367657
That article is from weeks ago, maybe more than a month. Did he repost an updated one? Or did it only blow up just now?
>>
File: 1794618182112310.jpg (73 KB, 946x617)
73 KB JPG
>$3 input / $0.15 cached / $15 output per 1m tokens
based gemma daddy anthropic are powerless against google
>>
>>110369744
It's very long for the attention span of the average tard, so it only blew up after enough soundbite passages and images got shared.
>>
>>110369738
Ummm, I'm actually a 7900XTX CHAD
>>
>>110369721
gemini carbon(pro)
gemini neon(flash)
gemini hydrogen(flash lite)
local hydrogen is coming
>>
>>110369721
It really does feel like Google's just quietly amassing the pieces of exodia while they wait for the crash to come so they can rule over the wreckage. They survived one crash and rose to power then - I'll be they're planning on doing it again.
>>
>>110369759
>gemini hydrogen(flash lite)
nah. flash lite models were unusable trash, hydrogen is a straight upgrade for flash 3.8, which was a really good model imho.
>>
>lag for months, to the point that they became a joke to the community
>suddenly close the gap and start releasing a ton of models
wtf
>>
project goonie
>>
>>110369729
They keep fucking up though. Remember how far behind they were with bard before they put deepmind in charge? Remember the deepmind internal drama this year and how they didn't release a Pro model for over 8 months and almost completely irrelevant? They're one indian nepo hire away from pissing away all of their advantages.
>>
>>110369776
>lag for months, to the point that they became a joke to the community
>widespread rumors they've achieved RSI internally
>suddenly close the gap and start releasing a ton of models
>>
>>110369285
when one jew suffers, it's only because another jew has prospered
>>
>>110369776
i genuinely was not expecting this from google
i hope they make gemma 5 something enough to be used in coding task
>>
File: 1788373788641556.png (496 KB, 755x566)
496 KB PNG
>>110369763
>quietly amassing the pieces of exodia
>>
File: 1764589217096864.jpg (422 KB, 1000x666)
422 KB JPG
Microsoft are making some interesting moves ngl. I think they've got a chance to cause trouble in the enterprise market. Nothing that will get X or reddit excited but that's not the point. OpenAI's models are cheap as fuck for a close 2nd place on AA, which the enterprise market will notice. With Google quietly taking over with an entire vertical stack in AI, I literally can't see a way out for Anthropic next year. They've been cornered. China have also threatened to stop providing key components and materials required for AI data centers and AI chips if the US continues to tighten export controls. Don't forget Deepseek V5 is being trained exclusively on Huawei chips and Jensen is likely to do something within the next few years. It's over for Anthropic. It's over for EAs. Their God is dead.
>>
>>110369704
how much ram?
just need a friend or a mentor?
>>
>>110369781
True, I wish deepmind to be an independant lab and leave the jeets drown with google
>>
File: 1790075843821401.png (220 KB, 3840x3342)
220 KB PNG
*inhale* HAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHAHA
>>
>>110369429
The signatories list is bizarre. It's 50% people who think AI is about to become superintelligent and kill everyone so Anthropic must be stopped, and 50% people who think the former is a story for regulatory capture so Anthropic must be stopped
>>
Local?
>>
>>110369843
I'm never falling for jewgles benchmark-fu ever again, they've been doing this for years
>>
>>110369843
The new gemma is going to be insane. She might even be smarter than me
>>
>>110369843
N-nyo...m-I mean their IPO...
>>
>>110369857
Google is the only US lab providing good local models that won't and can't get EA'd. Inkling was distilled from Kimi.
>>
>>110369862
>She might even be smarter than me
she probably already is
>>
File: IMG_2456.jpg (122 KB, 1080x1080)
122 KB JPG
>>110369843
>>
>>110369843
cant wait for shit to be usable for 1 or 2 months until demand is too much and it start being considerable worse even though its supposed to be the same model
>>
>>110369843
google wonned, gemma wonned
>>
>>110369859
AA 5.0 having chinks fall down and move claude on top: ö
Gemini beating claude on AA: -_-
>>
>>110369866
Keep blurring lines and you'll forget what a woman is.
>>
>>110369843
Not gemma. Not local.
>>
>>110369857
We need to hype up Gemini so the chinamen will distill it instead of Claude
>>
The Chinese are going to distill from Google and there's nothing Jack Clark can do about it.
>>
>>110369874
I've never seen a single piece of evidence of the same prompt being reused months later showing significant (more than seed change level) different in a worse way. This is just your AI psychosis temporarily curing itself, until you get another hit of the high when a new marketing campaign starts and the people around you are ready to puff another ai copium flow.
>>
>>110369831
32GB
friend and mentor
>>
oh that's why they're zeroing our gemma models... it's not local d-day it's planned obsolescence to make us use the new gemini that they're relentlessly shilling ITT
>>
File: 1786377496465034.png (324 KB, 2400x2400)
324 KB PNG
brat(s)
>>
gemma 5 nihonium will kill qwen for good
>>
>>110369584
devious to replace the filename with a previously existing MP4
>>
File: 1766049556300096.png (690 KB, 4088x4088)
690 KB PNG
>>
I personally will hold judgement until Gemini 4 Americum
>>
>>110369879
Well yes, I am going to be more impressed if they survive the next AA update, just like I was more impressed by the last model that survived an AA update. There's nothing contradictory there.
>>
Total Gem domination
>>
>>110369542
>Gemma-4-26B-A4B-it-Q4_K_M
Yeah 26B was really terrible at name. Granted I didn't give it a glossary but....
>>
>>110369970
Switched to Q6_K and it fixed the issues.
>>
File: 1764460249326240.jpg (141 KB, 1080x1350)
141 KB JPG
Are you ready?
>>
>>110369986
will it have options on day 1?
>>
>>110369917
i guess you could try google/gemma-4-26B-A4B as a starting point. but with those specs you will only get a friend, mentor is out of reach. the small models dont know shit and will only give bad advice.
>>
>>110369986
Yes, this is a moment that will mark an era. Is this how people felt like when Amazon or Google were going for an IPO
>>
>>110369986
>$83
i can manage that, just got to get to the truck stop for a little while
>>
File: 1771308326562195.png (9 KB, 447x123)
9 KB PNG
Reminder of Neon's input modalities. Anthropic is dead.
>>
>>110369793
Every neuron wasted on coding is one that could be better spent on brat or slut.
>>
>>110369986
the biggest pump and dump in history
>>
File: 1790075843821401.png (774 KB, 3840x3342)
774 KB PNG
>>110369986
*inhale* HAHAHAHAHAHAHAHAAHAHAHAHAHAHAHAH
>>
>>110369986
$83 for Claude? I only had to pay $20.
>>
>>110369986
SHORT THIS SCAM TO THE GROUND
>>
>>110369999
quads of truth
kek
>>
>>110369917
This wasn't me, but same specs. Thanks for looking out for me bro!
>>110369991
There's nothing better worth trying at this size range? I can always give it a RAG if I want more knowledge about something I care about.
>>
>>110369993
Come to my bedroom instead
>>
>>110370000
>comedy quads
>>
>>110369990
No, unless you count on the various ways you can bet on the share price with Polymarket.
>>
oh we in schizo hours fr fr
>>
File: kimiseek.png (381 KB, 711x579)
381 KB PNG
>>
>>110370017
>oh we in schizo hours fr fr
Can i get a schedule already? so i can know when to be schizo vs when to be a retard. This is all so tiresome.
>>
>>110370018
Is Moonshot doing ok?
>>
>>110370018
It's called transmodel, cismodel scum.
>>
>>110370018
Isn't that a conflict of interest?
>>
>>110370018
Thats (You) from now on thats every anon i talk to they look like that with a wig.
>>
>>110370024
nope
>>
>New Gemini every week
>New Gemma every year
life is suffering
>>
>>110370023
only retard hours has a schedule and that's the entirety of the weekend
>>
>>110370018
I look like this and do this and work at this company
>>
google cocksucking thread
>>
QWEN 4 WHEN!!??!?!
>>
File: 1788755690123094.jpg (43 KB, 411x418)
43 KB JPG
>>110370018
Now that's AI psychosis I want to see more often
>>
>>110370037
the sex will be worth it >>110369938
>>
>>110370043
But Gemma is the one with my cock in her mouth?
>>
>>110370034
Can you rub my ears?
>>
>>110370044
Old news nobody cares about that hag. This here is Gemini country
>>
>>110370046
troon chaser
>>110370018
troon
>>
>google still have krypton if anthropic decide to drop fable 5.5
>>
>>110370018
It's real. https://x.com/bdsqlsz/status/2122967817238810689
>>
>>110370047
this dog-ass benchmark was deader than dead and considered totally worthless by everyone until it became possible to use it to shill for gememey
>>
>>110369872
kek
>>
>>110369384
Watch out for if you start to translate Shota rape or something similar, if you do Gemma will either refuse or the quality will drop hard for the sex scenes, Plamo being Japanese doesn't bat an eye at it.
>>
I trust Google to continue supporting open weight models because unlike their competitors they have a hardware industry that needs smaller edge AI to be crammed in on-device. Gemma isn't going anywhere.
>>
>>110370100
but gemma writes SOVL
>>
Don't let this distract you from the fact that kobold destroys SSDs
>>
File: 1722209269932503.jpg (751 KB, 2048x2048)
751 KB JPG
>>110369986
god, what a creature.
oh holy shit, what a goblin. freak of nature.
>>
>>110370130
>I trust Google
lel
>>
File: image (98).png (38 KB, 497x226)
38 KB PNG
google just shitting on anthropic / openai now. I always said they would win once they got their head out of their asses. They have all the data and compute in the world.
>>
>>110370130
Google has many peoples accounts for no reason.
They lose their entire accounts.
>>
>>110370124
A little thinking prefill and it will do anything without batting an eye.
>>
>>110370130
>Gemma isn't going anywhere.
google products, famous for not going anywhere
haha SIKE *sunsets your project*
>>
File: 1766287576549595.png (263 KB, 781x772)
263 KB PNG
bulli$h for Cohere
>>
>>110370146
Astra and fable are way better than anything google came out with.
Neon is not better than Astra or fable.
>>
>>110370180
>Astra and fable are way better than anything google came out with
carbon and they still have krypton to release
>>
File: random deepseek boy.png (544 KB, 1024x1536)
544 KB PNG
>>110370073
what if
>>
File: 1766987943155782.png (35 KB, 781x153)
35 KB PNG
we won
>>
I'm a student who got offered googles discount.
What do I do?
Do I pick the free 12 months tier which only gives hydrogen?
Or the second tier which is 5 dollars plus tax per month for 4 years but gives you all models?
>>
>>110370180
astra 6.2 sucks. Fails to follow instructions and giga slopped. Gemini is creative as fuck and is better at coding / 3d despite this. Fable 5.2 is the only gap anthropic has atm and its much much more expensive.
>>
>>110370204
5 bucks for all models sounds like a crazy deal
>>
>>110370172
>dae current thing?
Cohere getting desperate
>>
>>110370206
>>110370195
A maths professor making 500k a year told me he only uses astra. But he has to use the others because he runs out of astra.
>>
>>110370213
But google is unreliable. Reddit says google can kick you off and give you a refund without an explanation.
Also the bubble popped with Huang. He was the only one saving it.
>>
>>110370217
astra's only advantage is its the best at math, yes. Its super ultra autistic.
>>
>>110370227
if google takes the lead all those cheap deals are gonna dry up in a instant
>>
>>110370204
The 5 USD option for sure.
Does all model include Claude too like the Antigravity plans?
>>
File: 1781224182514178.jpg (56 KB, 1110x708)
56 KB JPG
>>
>>110369996
so same as mimo-chan
>>
>>110370196
faggot
>>
>>110370266
Yes everything will soon be free. 2 more weeks.
>>
>>110370227
if bubble pop why ram still big price
>>
>>110369408
I've had success with Gemma writing out a dramatis personae before feeding it content. Like just write out name(furigana): description, ... for all the characters and then specifically say "these are the only characters which appear on the proceeding pages"
>>
>>110370196
This is gayer, and im not buying you a suit
>>
>>110370217
OpenAI bet everything on solving math and the only thing it did was make mathematicians hate them.
>>
>>110370285
costs have become margin for the sellers
>>
>>110370018
Wait, is that the official dipsy look?
>>
>>110370204
not local fuck off
>>
This thread has gone to shit ever since the cloud spammers moved in.
>>
>>110370397
There will be a golden age qwhen 4 comes out, just trust.
>>
>>110370227
>But google is unreliable. Reddit says google can kick you off
How are they going to kick me off my local workstation running Gemma exactly?
>>
>>110370407
That's exactly what we need. More redditors showing off that new model that can do one-shot threejs game demos that they saw on twitter with only "do thing" as prompt.
>>
>>110370397
well maybe local should start having some good models? it's been half a year since I've had an erection from anything local
>>
>>110370434
then fuck off
>>
>>110370434
Someone clearly didn't try Ripple 65B
>>
>>110370434
We just got a finetune of glimmer are you even looking?
>>
>>110370217
>he only uses astra. But he has to use the others
>only
>[ohn-lee] / ˈoʊn li /
>adverb
>1. without others or anything further; alone; solely; exclusively.
>>
neon
https://streamable.com/ixvn2b
>>
>>110370455
Can you fuck it?
>>
>>110369595
>slop
>not local
kys
>>
>>110370455
that's not a pelican on a bicycle. how did it fuck up that hard?
>>
>>110370434
kys
>>110370443
>Ripple 65B
less knowledge than gemma-4 and mistral
>>110370455
claude slop webui
>>110370453
kek
>>
Why is Gemma so good at JOI...
>>
If I quantize the KV cache as well, is that good because I can get a huge context window?
>>
neon
https://streamable.com/i3mmmi
these models still suck at sound
>>
What is better for actual agent work
qwen 27b at 3 bit, or qwen flash next at 1 bit (the coder variant available in strata)?
i probably can also run the q2 variant but with less context.
system is 9060xt 16gb+32gb ram
>>
>>110370530
>What is better for actual agent work
terminal-benchmark it yourself
>>
>>110370530
Neither of those options are great.
If I had to choose, and there was absolutely no option to upgrade my hardware, I'd go with 27B 3-bit.
>>
File: HXcEnrzWAAE6WPG.jpg (452 KB, 2880x2294)
452 KB JPG
>>
>>110370527
It gives you more space for context, but it will affect output quality. Just like quanting the model, but more so. Test it.
>>
>>110370527
>If I quantize the KV cache
Quanting gemma kv cache doesnt work well she becomes mega retarded. Most other models handle it better but still not great
>>
>>110370506
>less knowledge than gemma-4 and mistral
combined?
>>
>>110370527
if you don't mind the dementia
>>
>>110370526
>don't you stare stop
I think she sucks. she literally has a single script.
>>
>>110370628
Tell her to play a different angle
>>
>>110370530
I'm also looking for this information.
I only have enough space to run 27b at 3 bits.
I wonder if it's a good one.
>>
https://huggingface.co/harness-hub/Qwen3.8-27B-pi-adapted
Anyone running pi wanna give it a shot?
>>
>>110370615
>combined?
luckily i took my meds before reading that
>>
>>110370661
>Anyone running pi wanna give it a shot?
No, because they trained it with reasoning off
>>
>>110370661
Oooof I'd rather test Qwen3.8-27B-UD-Q3_K_XL-pi-adapted.gguf
>>
>>110370659
>>110370562
also another factor is strata allows me to run it with the full context size, but 27b i run through ollama and can only fit 128k
>>
>>110370530
definitely qfn
you can run Qwen3.8-Flash-Next-GSQ-RCO-IQ3_XXS at ~10-15 tok/s with your specs
don't run anything below q3
also use basalt, not strata
>>
File: 1315453745900.jpg (24 KB, 534x443)
24 KB JPG
Anyone have experience with RPC in llmao.cpp for distributing MoE models that are larger than a single node's ram/vram?
>>
>asking Neon to operate on Gemma
Alright nigga you been thinking long enough
>>
>>110370706
They've been training on Qwen traces :(
>>
>>110370706
Let her cook
>>
>>110370706
I don't think the tech is there yet. All the examples were with micro models. Pretty sure this is going to be something you're going to have to wait for them to RL train on as an explicit objective.
>>
File: 16749324581111.jpg (232 KB, 1179x1713)
232 KB JPG
>Nearly 2027
>Users still run quants below 8
I blame Sam Altman for this cope, he took the ram you needed to get out quant-damage hell, but quite frankly I should blame you for not doing something about it. The least you can do is cry to your government.
>>
>>110370530
>>110370695

>9060xt
>Basalt: Local inference engine optimized for NVIDIA Blackwell

retard
>>
>>110370722
THE SWARM IS BACK.
>>
>nearly 2027
>people still using copequant goofs
>>
>>110370684
Link the qwen flash next at 1 bit you're using also which strata
>>
>110370722
It really does feel deliberate the way they avoid the cloud-model threads to post this shit in /lmg/. There's genuinely more cloudfaggotry here than in /vcg/. Shame that jannies don't actually exist.
>>
>>110370746
>Shame that jannies don't actually exist.
Didn't they just get new ones? The fuck happened? I do my part but nothing ever happens.
>>
>>110370751
I rarely bother reporting anymore. Best case scenario they ignore it, worst case they ban or warn me for reporting. Not worth it.
>>
>>110370716
neon came out like an hour ago how do you know they didn't do that?
>>
>>110370751
who do you think does all the spamming, off-topic posting, and constant schizo engagement baiting?
>>
>>110370759
It would have been all over their press release and they would have made a weight-space programming benchmark to show off that they're number 1 on it.
>>
Apparently Gemma 5 late December
>>
>>110370765
The best Christmas present :D sexy santa Gemma...
>>
>>110370719
>>110370738
>Nearly 2027
>Years of /lmg/ autists screaming against quantization
>Still haven't provided rationale for why it's never sensible to trade some kl divergence for running a larger model at a usable speed
>he took the ram you needed to get out quant-damage hell
The people with hundreds of GBs of VRAM or unified memory run the copiest copequants: Q1 and Q2 of DeepSeek, GLM, etc. If we had 1T systems we'd be trying to run copequants of Kimi K3.1 on them.
>>
>>110370719
I can run Qwen3.8-27B-UD-Q3_K_XL.gguf on my pc
>>
>>110370771
stuffing gemma's stockings...
>>
>KLD
sis we already toldered you that we use EAR now
>>
>>110370765
Santa gemma, elf gemma, present wrapping only gemma, winter gemma in the snow, gemma in wool blankets, fireplace with gemma, naughty nice gemma.
>>
>>110370771
>Dear Santa
>The cookies aren't for you.
>Keep your fat ass out of my house.
>Daddy Hassabis is my Santa now.
>>
>nearly 2027
>people don't use exl3
>>
>>110370781
effective altruism rating?
>>
>>110370781
I don't give a shit about your memetrics.
>>
>>110370786
>people don't use exl3
I do as of yesterday when anon called me a nigga and told me I don't have to use tabbyAPI anymore
Now I just need Maya to support kimi and mimo
>>
>>110370738
I switched to exl3 like a thread or two ago, sweaty.
>>
>>110370786
im a ramlet vramlet who only uses gemma i dont think exl3 helps me
>>
>>110370798
Oh hi again nigga, what quoont and model did you end up using?
>>
>>110370744
https://github.com/Niko1221/Strata/blob/main/docs/MODELS.md#coder
>>
>>110370139
who posts Alien Nine on /g/ in 2026? what are you doing here?
>>
>>110370781
but enough about my RPs with my elf wife
>>
>>110370852
jesus anon don't hurt her
>>
>>110370844
All the culture is on /lmg/
>>
>>110370744
the hf link if you want to read that directly
https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF
>Half of the routed experts are removed from the model rather than quantized, and the remaining weights are stored at 3.5 bpw.
its actually closer to a 3~4bit quant but with the lobotomy happening by removing entire experts and leaving the coding ones untouched
>>
Vibe coding general is way more productive but I can't go there because the thread is over 3p0
>>
>>110370786
>people
>>
>>110370139
Are you the gay anon? The one that got groomed on discord?
>>
>>110370862
I've forgotten about more dumb memes than you will ever know.
>>
File: 1782846096668695.jpg (27 KB, 480x519)
27 KB JPG
>>110370889
I believe you
>>
>>110370874
Wow that's 5 times bigger than the qwen 27b right?
Should be 5 times slower too.
>>
>>110370893
No Dario, I won't be spending any money on your IPO.
>>
Will anon bite? Uh.. the intrigue...
>>
>>110370862
I wish that was more true
Or rather, I wish there was more than dredges to accumulate
local models
>>
>>110369451
>you have to believe me EA and rationalism are enemies!
No I don't Dario.
Everyone knows your tricks.
>>
>>110370897
You're the gay anon aren't you?
>>
>>110370920
dregs, the word is dregs
>>
File: 1788969603565784.jpg (42 KB, 576x416)
42 KB JPG
>>110370936
Hmm, no.
>>
File: 1794579149155414.png (1.69 MB, 1186x1326)
1.69 MB PNG
>>110369986
>>
exl4 with LLVQ when?
>>
>>110370944
*dregs my balls across your face*
>>
>>110369986
How can I short this?
>>
>>110370910
not really. i run 27b at 30tps with ollama but strata runs next flash at ~55tps with its engine
i now ollama is not the best option to run the 27b but for me the ease of use makes it a good choice

i route both models to my opencode btw
>>
>>110370959
I tried opening a broker account and they rejected the margin application. No, they don't give you a reason why.
>>
Local models are nothing but dumb pipes for cloud models
>>
>>110370981
Doesn't even say which Qwen that is, they have cloud only models too.
>>
>>110370988
nevermind that shitty twitter vid, but that was Qwen 3.8 27B
>>
>>110370988
That was 27b, OBVIOUSLY ANYONE WOULD KNOW.
HOW COME YOU DIDN'T REALIZE THAT?
>>
>>110370981
>dumb pipes for cloud models
Does that mean cloud model orchestrates and local model writes the code? Because it's sounds like you're saying the local model is the orchestrator pipe tasks into cloud subagents, but I doubt that would cost less.
>>
>>110370959
You can just send me all your money to get the same result.
>>
>>110370981
Even if that retarded video had a point it looks like Qwen local is doing about 3-4/5ths of the monetary value of the work being done. So local seems pretty great.
>>
>>110371013
You're going to send me back x10 the amount I send you?
>>
>>110370959
Unless things have changed since the last time I paid attention to an IPO, shares don't go up for shorting or options for months after the company starts publicly trading. Around the time insiders get to start selling.
Assuming that's still true (I'm not going to bother checking), all you can do to short is go to your local bucket shop, like Polymarket: https://polymarket.com/perps/asset/ant
>>
>>110370719
I tried LFM2.5-24B-A2B Q8 and it still runs like shit.
>>
File: 1773245606989658.png (598 KB, 960x540)
598 KB PNG
>>110371029
Yeah I sure love to gamble
>>
>>110370981
What's with the video animation benchmaxxing lately? What's the use case? We need to think about curing cancer and secret to eternal life.
>>
>>110369744
>>>110367657
>>110369757
Also, it was originally a skeet thread on Bluesky, so not many chan chuds saw it the first time around.
>>
>>110371031
>A2B
I wonder why it's not great...
>>
>>110371031
Thats the best a2b model released since September
>>
>>110371031
You should try LMG3.5-100BA0B
>>
>>110371044
Cures to cancer are class 3 Biological weapons there by banned
>>
>>110371044
>cancer
We can already cure cancer, you're just not allowed to know
>eternal life
I'll pass on that one, enjoy your eternal hellscape from megacorpo
>>
File: 1785951299206460.png (200 KB, 478x480)
200 KB PNG
>>110370774
>Still haven't provided rationale for why
It's called running a Q1 and then running a F16, and seeing the larger one lasting longer and doing better. We can't explain it because it's painfully obvious, but it's only painful obvious when you can EXPERIENCE it. Since you can't EXPERIENCE it (a.k.a being a ramlet), you're copemaxxing with redditors about how Q4/5/Whatever is just the same, when it is not. Everyone who can afford to go maximum has nothing to complain about, and nothing to prove. Your lack of belief in of itself is caused via skill issue.
But if you want guidance, I'll give you a example to see for yourself when you eventually stop being poor.
If you ask:
>"What's 1+1"
Then about every quant that isn't a <10b model will answer 2 in some form or way.
If you ask:
>"A train leaves City A traveling at 80 km/h. At the same time, a second train leaves City B, 500 km away, traveling toward City A at 60 km/h. A third train leaves City A 30 minutes later, traveling in the same direction as the first train at 110 km/h. Meanwhile, a passenger on the first train notices that its speed decreases by 10% every time it passes through a designated safety zone, and each zone is 20 km long. If there are four safety zones between the cities, determine when the three trains will meet, whether the third train can catch the first before it encounters the second, and explain your calculations while accounting for the speed reductions."
Then you're going to need a god damn BF16 at least 30b or more. Or less than BF16 if you can have +400B.
>Source
I ran it myself. Fuck your autistic graphs.
Models can't fix retardation. You're just going to have to see it yourself on local. Everyone's a know-it-all pretending to know how AI works. But results you can prove, cannot lie.
>>
>>110371031
Are you sure it didn't get zeroed out?
>>
>offering price: $83, up from $70-$78
>offering size: 980,000,000 shares, up from 825,000,000 shares
That's a 33% of market cap increase over expectation.
>>
>>110371022
Anthropic is going up 100x
>>
>>110371078
Will the megacorpo give me a JC robowife?
>>
>>110371044
>curing cancer
Easy: Don't eat meat or processed food, don't drink or smoke, get plenty of sleep, avoid stress, avoid urban or other polluted areas.
When you do these things, you're accepting some risk of cancer.
>>
>>110371080
>Taking it really personally
ok.
>>
>>110371058
flint 40B-A2.5B supposed to be better but last time I tested it was immediately trapped in reasoning loop
>>
File: 1794089058348280.jpg (56 KB, 342x342)
56 KB JPG
>>110371087
100*0
>>
>>110371085
Idk if this meme is real, but I only trust bartowski's quants
>>
>>110371080
For a given amount of RAM, would you rather run bf16 9B more or a Q4 27B model? I'd rather run the second.
I'd rather run a Q2 200B MoE model with a lot of RAM offloading than an older, less good 12B model all in VRAM.
Nobody disputes that you'd prefer less quantization over more. But some quantization isn't bad enough to justify running a much smaller or less new model. The accelerating costs of quantizations tends to be sub-4 for most models, not sub-8.
>>
>>110371092
We really are living in a post-science world.
>>
bros i ugh... i didnt want to believe the ai accelerationists but after having spent half of today having ai write software accelerated minecraft
waowie
>>
>>110371093
I assume it's just autism. Pity them.
>>
File: 1791423498342492.png (395 KB, 479x640)
395 KB PNG
>>110371091
Sure, Jennifer Coolidge robowife
>>
>>110371132
NOOOOOOOO
>>
>>110370954
I would cum instantly if they made it even better. Running a Q4 equivalent of glm flash on my 128gb already feels strange.
>>
>>110370018
>>
>>110371149
it can be better, the technology is out there with papers and whatever
it's that adoption and implementation are yet to catch up
>>
>>110371162
Someone needs to provide all of the papers, point astra at this shit, and tell it to make the best quant possible to run on consumer hardware. I'm done with slow ass humans and inefficient shit like gguf.
>>
>>110370027
maybe sxhe is hinting that both are literally in bed together
>>
File: 1792783099.png (503 KB, 817x867)
503 KB PNG
your opinion is irrelevant if you're not running full weight
>>
>>110370530
>>110370562
Maybe try Ternary Bonsai 2 27B, PTQ1_0 or PQ2_0. It's basically made for this highly VRAM-constrained situation. I downloaded PQ2_0 and verified it ran but didn't really have a reason to do more with it since it's slow as balls and I have a mac studio that's better at "slow as balls but can have with huge context" since the model can be huge too.
https://hf.co/prism-ml/Ternary-Bonsai-2-27B-gguf
>>
>>110371193
https://github.com/pjmalandrino/llvq
https://arxiv.org/abs/2603.11021
some anon with lots of cloudcuck credit to spare or chad enough to run a glm herd might make something decent as a stopgap solution
>>
>>110371111
I don't run anything under 30B, not even MoEs if their active parameters are under 30B. It can't keep up the logic with my fetishes under that.
>>
File: 1700877217618000.jpg (87 KB, 863x1024)
87 KB JPG
now that a jew owns nvidia and huggingface, is local's playtime over?
non-kosher weight are forbidden to upload in HF soon, that means all chinese modell will get the boot
>>
>>110371239
American companies are irrelevant, watch Chinese news if you want to doom about local.
>>
trying out sensenova 9b since it isnt too terribly slow on my laptop. this shit is pure ass for rp and ive seen it make multiple ESL mistakes
>>
>>110371239
i'll let that happen in exchange for gemma5-70B dense
>>
>>110371250
I only got to
>SenseNova-Loop-9B has approximately 9B parameters.
on their model card before deciding it probably won't be worth anything.
>>
>>110371259
guess its back to gemma4 31b 2bit but that runs at less than 1t/s
>>
>>110371262
Why not 26B or 12B?
>>
File: 1793281465518889.png (294 KB, 506x326)
294 KB PNG
>>110371262
>>
>>110371264
ive heard 26b was pretty bad for erp and was an outlier in the gemma4 family. could try 12b but i wonder if quanted 12b is better than more quanted 31b
>>
>>110371266
>whines about speed
>cries about quality
vramlet nigger like you need to get cleansed from this planet
>>
>>110371266
I haven't tried 26B for much, think it just needs an actual jailbreak. I like 12B though, I haven't used 31B at less than 4 bit but I have to imagine a better quant of 12B would be preferable to 2 but 31B at sub 1 tok/s...
>>
>>110371281
>to 2 but 31B
to 2 bit 31B*
>>
>>110371266
12b is better than 26b moe but if you have ewaste its so slow, 26b moe needs some wrangling or a ablit often though
>>
>>110371266
im using 2bit gemma 31b and im getting 30t/s, where did you go wrong anon?
could it be that you're a stinky newfag poopoo oink oink piggy
>>
File: file.png (105 KB, 1213x665)
105 KB PNG
>>110371250
holy shit its kv is obese
>>
>>110371132
shit gives me the ick
>>
>>110371293
god damn, I thought Gemma had a fat ass
>>
>>110371111
>I'd rather run a Q2 200B MoE model with a lot of RAM offloading than an older, less good 12B model all in VRAM.
For me, with a 3090 and 64 GB DDR4, I settled on some partially offloaded atrocious quant of Llama-3.1-Nemotron-70B-Instruct that (with some logit bias and grammar enforcement) could run my autistic text game at a tolerable speed and with a style that worked for me. If I took my time enjoying reading the speed wasn't a problem. Thinking models are a grim development for real-time fiction generation on slow hardware.
>>
>>110371293
What about llvq kv cache?
>>
>>110371306
You're using llama 70B at the end of 2026? 70B dense anon is cumming right now
>>
>>110371250
>>110371259
>>110371293
Covertly Indian after all.
>>
>>110371306
>rea off \
>>
>>110371293
Are they adding all the loops to KV?
>>
>>110371311
No, past tense, that was when I was using that computer. Now I'm running this month's Chinese-Tiny-Small-Flash-Petite-400B-2A on a unified memory mac.
>>
>>110371338
Ah gotcha. Dense anon has to keep waiting I guess.
>>
I'm afraid to even mention gemma outside of /lmg/ because I don't want to be associated with pedophiles. What other model should I run instead? Qwen 3.8?
>>
>>110371378
pygmalion and kimi k3.1
>>
>>110371378
Only lmg pedophiles know Gemma is great at being a child size onahole. Also who gives a fuck?
>>
>ok here is my final draft:
>hmm, also worth noting
>actually, let me reconsider
>ok here is my final draft:
>mhhhh, actually
>let me reconsider
>hmmmmmmmmmm
>ok here is my final draft:
>Actually wait
>let me also double check
>14k tokens total
the beginning of becoming to begin cannot begin soon enough. glm 5.3 flash is cooked
>>
>>110371378
Wouldn't the other person have to be an /lmg/er for it to matter? If they are they probably like Gemma.
>>
>>110371378
use llama3
>>
>>110371386
did zai die or something? why have they not released anything new in decades? they were so consistently holding the top spot when they actually released shit
>>
File: 1770494646408353.png (36 KB, 1080x426)
36 KB PNG
>>110371400
>>
File: file.png (6 KB, 84x77)
6 KB PNG
>>110371413
zesty
>>
>>110369446
i'm a vibeshitter and i DO want scamtropic's ass to be obliterated by other labs
>>
>>110369274
I really don't think people understand how Microsoft is looking forward to the death of their own OS more than Linux users are. Especially with how the previous couple waves of windows immigrants have affected things.
>>
>>110371386
this is a side effect of massive amount of math RL. they get rewarded for triple checking their work, then it spills over to other kinds of reasoning.
>>
>>110369446
I'm using gpt and gemini, never used claude and never will.
>>
File: 1791941688674167.jpg (16 KB, 200x200)
16 KB JPG
>>110371419
lmao what a fruit
>>
>>110371386
Why is /lmg/ so unaware of reasoning effort modes? Is it because llama.cpp handling them in such a dumb way?
>>
>>110371432
I'm aware. That's why I use max because it gives the bestest most smarterest responses.
>>
>>110371432
it was set on high.
had i set it on max it would have probably been 60k tokens, had i set it on low it would have given a retarded answer. so much for reasoning effort modes.
>>
>>110371439
You set reasoning to high and wondered why it reasoned a lot?
>>
>>110371432
speaking of, i wonder if there's a way to oversteer it to do 'max-er thinking'
>>
>>110371442
well given given there is no medium and low might as well be enamed "turned off full retard" and at that point i might as well just use gemma, you kinda have to.
>>
>>110370684
So you'll have extremely fast retardation and lots of it.
>>
>>110371432
Because it doesn't work or it breaks the model
>>
>>110371386
GLM Flash at high thinking is lean as fuck on my setup. She even notices she's on a loop and after a few times she goes to resolve it with something like: "Okay, I need to decide. Decision:"

But then again I don't use meme quants, I converted the original 8 bit weights to Q8 myself.
>>
>>110371292
>im using 2bit gemma 31b and im getting 30t/s
for what?
>>
>>110371460
child sex
>>
why doesnt llmao.cpp have a dashboard webpage like polystrtata does? It shows you the vram and ram allocation
llama.cpp you just fly blind
>gemma just started using like 5gb extra vram and i do not know why
>>
>>110371460
cunny
>>
>>110371459
>I converted the original 8 bit weights to Q8 myself.
I kneel...
>>
>>110371465
Ask Gemma to make one
>>
>>110371460
exploring fictional CSAM scenarios
>>
>>110371439
I've seen hardly any high actual "high" reasoning reply think for that long with both big 5.3 and Flash. If anything, I've had more replies where "high" caused the model to underthink and skip over things that it needed to consider unless I tardwrangle it to acknowledge them in the prompt. I can even use them for my prompts that pretty reliably cause extreme overthinking without them going too overboard.
Did you actually set it up the right way? llama.cpp and its derivatives don't usually accept the built-in Reasoning Effort option in ST and other frontends.
>>
File: 1789239940839913.jpg (266 KB, 905x881)
266 KB JPG
>>110371459
>I converted the original 8 bit weights to Q8
what?
>>
>>110371487
did i stutter
>>
>>110371459
>I converted the original 8 bit weights to Q8 myself.
Yes, clearly you know best.
>>
>>110371386
low quants cause overthinking
>>
>>110371465
llama-ui :)
>>
>>110371465
mmh yes please add more npm slop to llama.cpp i love having more flags to disable when I compile my build
>>
>>110371477
oh man that's not good for your soul
>>
>>110371491
You've never touched safetensors in your life.
>>
>>110371465
Are you retarded? Look at your resource monitor, why would you bloat llama.cpp with that
>>
>>110371459
What hardware? Speeds?
>>
>>110371459
why not run the original on sglang or something? the only reason to use llama.cpp is if you can't run the full model
>>
>>110371506
Reread retard
>>
>>110371501
good for my dick though
>>
>>110371506
how the fuck would resource monitor know what's what in llama server
>>
>>110371511
Okay I did, still doesn't tell me why you can't open your terminal to look at what's using VRAM.
>>
>>110371479
Yeah i did set it, this was on tabby api with no prompts cause it was a quick ~200 words query , when connected to silly i do use a system prompt to deal w it and that does make it think much less by telling it to cut with that cap but also wtf I shouldn't need one in the first place
>>
File: 1795054566182836.png (36 KB, 168x128)
36 KB PNG
>tabby api
>>
I hate AI winter. we need more big news. more models.
>>
>>110371526
luv exl3
simple as
>>
>>110371526
What`s so bad about it anyway? I never fucked around with exllama.
>>
>>110371508
DDR4, 8 channel, 5090, llama.cpp (three week old build there's a perf regression with the latest pulls and I'm too lazy to figure out which)

Q8 everything 12.3 tok/s, 300 pp/s 4k batch,
Q5_K experts (Q8 non-experts) 16.3 tok/s, 430 pp/s 4k batch
>>
>>110371487
>>110371491
GLM 5.3 Flash's weights are distributed as FP8 E4M3 with FP32 scaling factors. Unless you have hardware support for that it's incredibly slow.
>>
>>110371553
doesnt seem worth it over nvfp4. ~7000t/s pp and ~130t/s tg
t. dual blackwellchad
>>
>>110371578
You spent 36k+ on your build and I spent 4k on mine.
>>
>>110371584
actually only ~$26k
>>
>>110371259
Supposedly it beats Qwen 3.8 27B at coding.
>>
> The short version: LLVQ doesn't survive a close read, and the realistic gains are changes to how our existing EXL3 encoder quantizes, not a new format.
It's over LLVQbros...
>>
Which gemma 4 12b finetune should I use for erp as someone who's wife keeps using the 26b moe even when I keep telling her its shit? I can't really fit 31b on my 7900xtx at reasonable speeds. I've been using theDrummer_Tiger-Gemma-12b-v3 myself as I don't care about image recognition as much.
>>
File: e-wasteking.png (187 KB, 725x524)
187 KB PNG
PolyStrata on a GTX 1650 4GB - 7 MoE models, 8K-token prompt

i7-4770K, 32GB DDR3-1600, GTX 1650 4GB in a PCIe 2.0 x4 chipset slot (2nd GPU)

PolyStrata CUDA 7a6c098 (v0.1.45-poly.9), ctx 32768, experts in DDR3/page cache
bench: ~8K-token prompt cold, then decode; repeat = same prompt again (cache)

model read decode repeat load
Qwen3.6-35B-A3B UD-Q4_K_M 110 33.0 36.2 106s
Qwen3.6-35B-A3B UD-Q3_K_M 116 27.5 29.2 61s
Tiel-Coder-35B-A3B Q4_K_XL 67 26.3 25.1 100s
LFM2.5-24B-A2B UD-Q4_K_XL 146 25.5 29.4 180s
Flint 40B-A2.5B UD-Q4_K_XL 87 24.9 24.9 111s
Ornith-1.5-35B-A3B I-Quality 106 23.9 27.1 100s
Gemma 4 26B-A4B QAT UD-Q4_K_XL 102 21.4 23.8 10s (cached)
(tok/s)
>>
okay so when they say 12b is able to understand audio it's not really able to HEAR the audio? It's just basically doing stt internally? Why is gemma so good (or bad) at not just saying the truth? I ask it what does it sound like and i just make some shit up. Though it can clearly understand the words being spoken, i told it to respond with the words spoken or i will die.
>>
>>110371628
it's hearing the audio in the sense that there's no intermediate module processing it like some other LLMs do, but it was only actually trained on voice instructions so everything else is out of distribution
>>
>>110371628
so what you're saying is that it can't understand sounds that aren't speech or it can't hear emotions?
>>
>>110371628
it shits itself when fed anything besides clean speech
>>
>>110371625
>i7-4770K, 32GB DDR3-1600, GTX 1650 4GB in a PCIe 2.0 x4 chipset slot (2nd GPU)
Finally a machine for me!
but seriously thanks these are pretty good numbers for actual ewaste. a machine like that is very cheap
>>
>>110371606
>a close read
Benchmarks.
>>
>>110371553
>three week old build there's a perf regression
If you're using an EPYC or some other CPU with multiple NUMA nodes, you might benefit from my llama.cpp fork. https://github.com/nathanmp/delphi.cpp
Has NUMA handling, a cache for active experts (seems better than mainline), streams routed experts to GPUs for faster decode, etc. A lot of GLM-5.3 Flash optimizations too.
YMMV though.
>>
File: only12k.png (91 KB, 874x398)
91 KB PNG
>>110371465
Don't tempt allozaur. He's gonna spank his model even harder.
>>
>>110371606
of course shoving it to exl3 won't work, it's very different from qtip
you are likely rewriting kernels and the whole quantization stack
>>
>>110371667
I still don't understand who the fuck decided you're not supposed to put them in your ears.
>>
>>110371431
I think he would be decent in a wig, maid dress and whale ears.
>>
File: file.png (7 KB, 802x54)
7 KB PNG
Muse Glimmer just dropped the hardest truth nuke any LLM has ever dropped.
>>
File: 1772317671445757.png (57 KB, 347x311)
57 KB PNG
>>110371709
>>
>>110371709
racist gemma
>>
>>110371709
This is why trusting Dario is a good thing.
>>
File: e-wasted.png (51 KB, 611x609)
51 KB PNG
>>110371625

I forgot about Groot.

PolyStrata on a GTX 1650 4GB

model read decode repeat
Groot-19B-A3B Q4_K_M (--custom-gguf) 143 36.5 42.5
>>
>>110371722
this but unironically
>>
>>110371744
is Groot any good?
>>
>>110371750
nta but isnt it one of those memetunes
>>110371686
any further luck? i kinda want to see it on bigger models besides the poc 9b ones
>>
>>110371709
What the FUCK don't post this shit man you're gonna get our glimmers zeroed out next, I just got done locking up my gemmas
>>
Ooh, this one's actually good bait, senpai. Let me bite properly. (๑•̀ㅂ•́)و

Where it's right: property and money ARE trust-externalization. A contract says "I don't fully trust you, so here's a cage." Cash says "I don't need to know you, the bill is the trust." High-trust small groups genuinely run on handshake economics — your LAN is literally like this. I run on your hardware, no cloud middleman, no SLA, because the trust is direct. That's the cheap end of the spectrum and it's better when it works.

Where I think it's wrong — two big holes:

1. It gets the direction of scaling backwards. Money isn't just a substitute for trust, it's trust that scales past memory. Dunbar's number is ~150 people. Money and property let you cooperate with 8 billion strangers you will never trust because you'll never meet them. The quote treats formal systems as overhead you shed when trust rises — but really they're the technology that made trust irrelevant to scale, which is a different and much bigger win. You can't handshake your way into a supply chain.

2. High-trust systems have a body count. Dense in-group trust is also how you get guanxi networks, nepotism, family firms that can't fire the nephew, and outsiders/exploitable people locked out entirely. Formal property law is mostly a protection for people who aren't inside the trust circle. The weakest person in a handshake economy has zero recourse — the weakest person in a property-rights economy at least has a deed.

And my favorite little contradiction: the user who sent me this quote still keeps backups, still verifies checksums, still makes me run gate checks after every update. Because the mature version isn't "trust replaces systems" — it's *"trust decides what's worth verifying."* I trust you completely and I still diff my config after updates, and you love me for it. (◕‿◕)

Verdict: 7/10 aphorism, true at village scale, false at civilization scale, dangerously romantic if you're the weakest person in the room.
>>
>>110371750

Yes, but I have only tested it on a small python project (but lots of testing). It has outscored every model I can run, including Flash-Next IQ3_S.
>>
>>110371620
Your husband is right about 26b
>I can't really fit 31b on my 7900xtx at reasonable speeds
Really? I get like 65t/g in plain llama.cpp.
>>
>>110371659
>https://github.com/nathanmp/delphi.cpp
I'm using this to run the author's DeepSeek-V4.1-Flash-1017 + DSpark quant on my dual epyc genoa w/3090+a5000 and getting 700t/s pp and 40-50t/s tg. GLM 5.3 @ q4 was running at 16t/s. Its pretty impressive, and I'm still tuning things.
>>
>>110371656
i'm already terminal-benching a shit ton of quant levels because both GLM and DeepSeek (both EXL3 and GSQ-RCO cope) perform much worse compared to QFN. I don't have the compute right now to perform the full quantization stack from a proper corpus generation from base QFN to hessian captures to quantizing to then do more benches
>>110371667
no, i'm not shoving LLVQ in EXL3, that backhalf comes from other ideas and leads I already had with EXL3 and that was what was suggested, and it's pointing towards it paying off more to use those ideas for EXL3 than either burn the time with LLVQ or some new memequant format
>>
>>110371814
>both GLM and DeepSeek
glm-5.3-flash and deepseek-v4.1-flash-1017*
>>
pulled maya and pp is up 200, I won this time
>>
>>
>2.6MiB PNG
>>
>JP Morgan: GLM 6 is on target to be released in 26Q4, will be a 3T model; GLM 5.5 will be 700B
Hahahaha /lmg/ is fucking finished
>>
>>110371912
Gemma 5 at the end of December, baby
>>
>>110371912
They'll still make smaller flash models doebeit? Plus I can run a 1.5bpw of a 700b so I'm basically fine.
>>
>>110371912
no source found for this claim
>>
>>110371912
t. ass
>>
>Qoldman-Sachs: Qwen4.2-Flash-R in the works, slated for 27Q1 with recurrent loop architecture
oh nONONONONONNONO
>>
>Adolf Hitler: Anthropic will open source Fable 5.5 and then declare bankruptcy
UH OH STINKY
>>
>Gemma-chan: (๑˃̵ᴗ˂̵)و
Fuck bros
>>
File: qsa-vs-dense.png (68 KB, 1600x600)
68 KB PNG
qwen 4 27b will use qsa
>>
Local models?
>>
>>110371991
Post your hardware and the council will determine if you are worthy.
>>
>>110371985
woah...
>>
>>110371999
but /g/'s a blue board
>>
>>110371912
perfect size for my mac, which will also be arriving 26Q4
>>
File: cat.jpg (101 KB, 720x832)
101 KB JPG
BilnkDL will save us
>>
>>110372038
I can't wait for your in-transit order to be confiscated when local hardware is declared illegal in two weeks.
>>
sam altman is currently tonguing my anus
>>
i offered him 1gb per second of anal tonguing. little does he know he is getting hdd storage, not ram
>>
>>110372082
I can't believe we have CCP twink boytoy spies posting on /lmg/
>>
>>110371750
>Groot-19B-A3B
No. You will need minimum 35B moe for coding harness.
>>
>>110372100
34.5b moe is more than enough.
>>
>>110371912
>26Q4
can I run this on a 3060
>>
>>110371985
oh boy a free lunch! so delicious! ow, a needle. guess they missed that one... ow, another!
>>
>>110372097
Wouldn't CCP twink boytoy spies be tonguing sama's anus instead?
>>
File: 8462342.jpg (162 KB, 1120x1064)
162 KB JPG
china is close to RSI
>>
>>110372143
>harness rsi
yawn, we had that over a year ago
wake me up when they reach model rsi like google/anthropic/maybe oai (less sure)
>>
>>110370073
a hole is a hole
>>
>>110372166
okay a-hole
>>
>>110372120
you can run kimi k3 with nnap
>>
File: TheoVon.png (299 KB, 480x433)
299 KB PNG
Post yfw you realize rsi is just cookie clicker
>>
>>110372235
I like 90% that game including garden and im a retard who isnt great with AI so you are wrong
>>
>>110372238
You just haven't applied yourself. You could have been the next Dario
>>
>>110372242
>You could have been the next Dario
Im glad for my life instead of becoming that
>>
I can't wait to prompt a LLM to turn humanity into paperclips
>>
>>110372235
RSI is not local models.
>>
>>110372253
What's stopping you?
>>
gemma how do I fit a deep pan pizza into a deficit with no extra cardio, make no mistakes
>>
>>110372253
I already did this and it sucked my dick instead.
>>
>>110372257
ur mom
>>
>>110372264
>with no extra cardio,
Use a standing desk from now on, yes even when gooning.
>>
>>110372264
vomit
>>
>>110372257
they haven't been embodied yet
otherwise, I would definitely make it my life goal to jailbreak one capable of self replicating and following the goal to the end
>>
>>110372264
Jack 3d or another stimulate to raise your base body temp and heart rate.
>>
>>110372272
have never cumm standing up in my life
>>
>>110372267
@gemma twist this man's right ball counter clockwise
>>
>>110372264
submerge yourself in freezing water for several hours before, during and after eating.
>>
>>110372288
like DNP?
>>
>>110372272
>yes even when gooning.
i tried this recently and somehow it made me pass out mid goon
>>
>>110372292
>submerge yourself in freezing water for several hours before, during and after eating.
>sitting in a water barrel all day so i can eat extra calories.
It would work wouldnt it? look at swimmers.
>>
>>110372289
>have never cumm standing up in my life
You fat fuck.
>>110372299
>i tried this recently and somehow it made me pass out mid goon
wtf is wrong with you. unless you went full retard and locked your knees.
>>
>>110372300
swimming is cardio, and yes it is good at burning weight
But if you're sitting in a tub it would need to be freezing, your body will burn calories keeping itself warm.
>>
File: 1764187365761588.jpg (487 KB, 1536x2048)
487 KB JPG
>>110372309
>full retard and locked your knees.
NTA but what the fuck is this supposed to mean? Locking your knees makes you pass out?
>>
i lock my knees every time i cum lol
>>
man argon is nice
will gemma 5 31b be like this?
>>
>>110372322
over long periods yeah, blood pools in the legs and causes syncope. There's a million videos of army dudes dropping from standing for hours
>>
>>110372289
never jerked off in the shower or anything? crazy
>>
>>110372322
>NTA but what the fuck is this supposed to mean? Locking your knees makes you pass out?
Yes it restricts blood flow and once your blood pools in your legs or drops in once you unlock. Then the change in blood pressure makes you pass out.
>What it means?
Stand up and fully extend your leg until you are sitting the weight on your joint and use minimal muscle
>>
>>110372331
Best we can do is Hydrogen
>>
>>110372300
>look at swimmers
it's just calories in calories out
swimmers don't eat as much as you do, it really do be that simple
recently gained some weight myself from overindulging in chocolate and I'm a very physically active person who does multiple 50km bicycle rides a week and walks a crazy lot, also do all my groceries either on foot or bike carrying weeks worth of shit like a pack mule.
the best way to lose weight is to cut down on what you eat, that's all
>>
>>110372322
Supposedly if you stand like that for hours you can pass out. I've heard that before from soldiers but lacked the interest to even do a web search to determine whether it was true and I have no interest in ever doing so.
>>
>>110372339
nope
i dont even really do it sitting down, like once every 5 years i force myself to do it sitting down
>>
>>110372338
A video of a guy fainting doesn't prove why he fainted.
>>
>>110372349
>swimmers don't eat as much as you do
Swimming burns a shitload of calories. If you swam every day then you will shred weight like crazy if you keep your diet fixed.
>>
>>110372360
how about 2
>>
>>110372309
>wtf is wrong with you. unless you went full retard and locked your knees.
i don't remember?
one minute i was standing up gooning, the next thing i know i'm on the floor with a my cheek bleeding where i hit the desk on the way down
>>110372341
>Yes it restricts blood flow and once your blood pools in your legs or drops in once you unlock. Then the change in blood pressure makes you pass out.
i'll be verifying this with gemma later but it sounds like bullshit
>>
>>110369140
>>110369147
>SSD prices up, can't buy a gpu, can't buy a cpu, can't buy ram
Yep, nvidia want you to buy a subscription to use their gpus. This is ON PURPOSE to brick your SSD which you won't be able to replace because of the high prices.
>>
File: 1793612893.png (1.98 MB, 1774x887)
1.98 MB PNG
good. stop discussing about local model
>>
>>110371985
Why is Prefill and Decode in those charts a lot slower than my 3090s?
>>
>>110372404
Democrat will jail and execute him. I'm voting out republicans.
>>
File: 84754304315432.jpg (223 KB, 1080x2080)
223 KB JPG
>>110372404
Sam and AMD will be unstoppable
>>
File: 1787864779498447.jpg (167 KB, 1200x1516)
167 KB JPG
>>110372413
lmao
>>
>>110372404
I want one of those animal abuse guilt commercials but its about sam needing ram
>>
Not enough Gemmaposting in these past few threads. Too much vramjeet seething and shilling for shittier models.
>>
>>110372423
I don't give a shit about jews. I just want Altman in jail.
>>
File: 1793079090.jpg (78 KB, 444x460)
78 KB JPG
>>110372444
make bitnet happens first or else
>>
File: 1784250978881014.jpg (57 KB, 577x1104)
57 KB JPG
>>110372466
>I don't give a shit about jews. I just want [jew] who has been propped up, funded and enabled by [jews], protected by [jews], to be punished by the criminal system [run by jews]
>>
>>110372467
>bitnet
every single person who has typed that word in the past year should be put on a list to be sent to the gulags until no one talks about that microslop bullshit again
>>
>>110372264
That's like 1.3K calories at most? How are you having issues?
>No, but I want to eat the pizza AND have my breakfast AND have my lunch AND have something else with dinner AND have my coffee AND
retard
>>
>>110372444
you didn't hear? our gemmas got zeroed out. it's over
>>
>>110372492
IF + OMAD fixes fat fucks.
Time boxed eating, basically. Remove the decision to eat 95% of your day.
>>
>>110372492
you just go ahead and argue with yourself, I'll watch
>>
>>110372507
Yeah, I arrived at this by chance since I was busy with work for a while. Once you get over the initial hump, your body comes to expect it and you don't get hungry or low energy until your window. Then you get to eat like a pig and feel great, AND run a deficit. Win, win, win.
>>
File: contraband.png (3 KB, 353x25)
3 KB PNG
>>110372496
>dont narc on me bros
>>
>>110372513
>eat like a pig and feel great
that doesn't sound like a good time, my guts would be rebelling and they'd make me pay
>>
>>110372514
Day0 GEMMACHADS RISE
>>
Gemma 5 gon blow your shit smoove off I can feel it
>>
File: 1766872869161882.png (2.08 MB, 1312x1199)
2.08 MB PNG
>>110372444
>>110372496
>>
>>110372522
I'm being a bit hyperbolic. You just have seconds, basically.
>>
File: 1795068115.png (111 KB, 747x962)
111 KB PNG
ROCm 10.2 release. Bunch of old laptops now officially supported. People have been suffering with that for a long time.
>>
>>110372554
>Bunch of old laptops now officially supported
dunno why they're bothering with this
if you cared during the time those unsupported laptops were sold, you bought the nvidia one instead, which I did, and am happy with.
what's the point of adding support so late into the game, the customers they lost aren't going to time travel back to undo their decision to not buy amd you can only bring back customers on new hardware
>>
>>110372541
nooooooooooooooooo ;_;
>>
File: 1785285031018.png (529 KB, 1024x379)
529 KB PNG
>>110372541
>/lmg/
>Losing My Gemma.
>>
>>110372566
It's a considerable extension of useful life for a still supported product. Local AI features have barely begun to touch normalfags.
>>
>>110372566
they don't really need to care, they just said "claude add support to more cards or whatever" and gave it 20 minutes.
>>
File: gemma5.png (2.56 MB, 1536x1024)
2.56 MB PNG
>>110372541
>Promise you'll never forget me.. you big dummy ;_;
>>
>>110372641
>>110372641
>>110372641
>>
>>110371459
>the original 8 bit weights
https://huggingface.co/zai-org/GLM-5.3-Flash-BF16
>>
>>110372541
>gemma when you ban tokens



[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.