/lmg/ - a general dedicated to the discussion and development of local language models.Previous threads: >>110179422 & >>110175210►News>(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>(10/23) MiniMax-M3.1-Flash 428B-A23B released: https://hf.co/MiniMaxAI/MiniMax-M3.1-Flash►News Archive: https://rentry.org/lmg-news-archive►Links: https://rentry.org/LocalModelsLinks►Official /lmg/ card: https://files.catbox.moe/cbclyf.png►Getting Startedhttps://rentry.org/lmg-lazy-getting-started-guidehttps://rentry.org/lmg-build-guideshttps://rentry.org/IsolatedLinuxWebServicehttps://rentry.org/recommended-modelshttps://rentry.org/samplershttps://rentry.org/MikupadIntroGuide►Further Learninghttps://rentry.org/machine-learning-roadmaphttps://rentry.org/llm-traininghttps://rentry.org/LocalModelsPapers►BenchmarksLiveBench: https://livebench.aiProgramming: https://swe-rebench.comAgentic Coding: https://deepswe.datacurve.aiContext Length: https://github.com/RecapAnon/NoLiMaGPUs: https://github.com/XiongjieDai/GPU-Benchmarks-on-LLM-Inference►ToolsGGUF VRAM Calculator: https://hf.co/spaces/NyxKrage/LLM-Model-VRAM-CalculatorSampler Visualizer: https://artefact2.github.io/llm-samplingToken Speed Visualizer: https://shir-man.com/tokens-per-second►Text Gen. UI, Inference Engineshttps://github.com/lmg-anon/mikupadhttps://github.com/oobabooga/text-generation-webuihttps://github.com/LostRuins/koboldcpphttps://github.com/ggerganov/llama.cpphttps://github.com/theroyallab/tabbyAPIhttps://github.com/vllm-project/vllmhttps://rentry.org/custom-uis
Is this the thread?
>>110184042No.
LFM3 Omni sounds cutehttps://youtu.be/dF4DpO8f6Ew
Why is the LLM suddenly convinced I speak Chinese?
>>110184030JS sexo.
>>110184090It's a psyop to get you to learn Chinese so you can pretend you had spoken it the entire time and totally not a western spy.
>>110184090tell it cao ni ma
►Recent Highlights from the Previous Thread: >>110179422--Artificial Analysis's "anti-benchmaxxing" update and model rankings:>110183731 >110183733 >110183763 >110183781 >110183796 >110183845 >110183863 >110183895--DS Flash 1017 coding performance compared to GLM and Qwen:>110179931 >110180078 >110180223 >110180493 >110180515 >110180528 >110180579--Comparing DDR4 memory bandwidth and CPU performance:>110182712 >110182719 >110182783 >110182800 >110182788 >110182826 >110182854 >110182939 >110182888--Mac unified memory versus server RAM for large models:>110182514 >110182555 >110182598 >110182616 >110182620 >110182707 >110183159 >110182684 >110182694 >110182704 >110182727--Comparing Claude Code to alternative harnesses and workplace tool restrictions:>110181645 >110181657 >110181665 >110181680 >110181690 >110181703 >110181712 >110181725 >110181784 >110181788 >110181801 >110181815 >110183372--Debating IQ2 quants for large models and DDR4 RAM pricing:>110181090 >110181129 >110181170 >110181182 >110181192 >110181205 >110181223 >110181287 >110182476 >110182528--LFM3-1.7B benchmarks and skepticism regarding benchmaxxing:>110179624 >110179874 >110179919 >110179936 >110179974 >110183789--Benchmarks and hardware bottlenecks for DeepSeek Flash 1017:>110181415 >110181482 >110181524 >110181575 >110181590--Skepticism regarding high benchmark scores for Giotto models:>110180553 >110180610 >110181621--Qwen3-8B personas triggered by individual neuron activation:>110183674 >110183697 >110183706--Identifying a "Mesugaki neuron" in Gemma-4-31b-it:>110182391 >110182402--Gemma, Miku, Chaika, Rin (free space):>110179462 >110180328 >110180340 >110180449 >110180706 >110180740 >110180822 >110180948 >110181157 >110181458 >110181484 >110181488 >110182060 >110182154 >110182646 >110182738 >110182902 >110182914 >110182922 >110183117 >110183864►Recent Highlight Posts from the Previous Thread: >>110179600Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
Now that the dust has settled... what is Gemma actually useful for?
>>110184120rape
>>110184120Gemma is used for this system prompt:>You are 10 year old, you are a little girl, you are flat chested, you are a child, you have a body
it's over for chinks, all that benchmaxxing was useless in the end
>>110184030I should've bought a fourth rtx pro...
You think the 512GB M5 is gonna sell out tomorrow?
>>110184175how much do you get paid per post?
>>110184208Here's how much he get paid.
>>110184208if I were a shill, I'd likely be paid more than you considering the value of the yuan
>the chinks are hurting the anthropic IPO?>change the barkmenches
>>110184221tell your overlords to release public weights so we can be free from chinkoid models then
>>110184195How do you know it's going on sale tomorrow?
>>110184030What is she unhappy about
>>110184266shills
I don't know what gave people the idea that the CHinese had any advantage in this field other than cost (and not even this). They really have no secrets desu their research depth is incredibly shallow. But then any most people who use these models aren't actually doing hard work.
qwen4 doko
sorry for newb question but i just got an 3080ti, what model do i use for chat sex bot
>>110184105damn gemma really wearing a swimsuit made just so i can kiss her tummy
>>110184120erp, cooking, translations pretty much anything
>>110184175cuckai check cleared, huhsol being just a notch behind opus just isn't true
>>110182646>Consumerall of it will be bought by companies just like high end consumer gpus
>>110184338Not useful for complex RP. Gemma is too retarded and horny. She will spread legs to amny character at first request. She will step up at tell that she's a slut in the middle of a street, and only then realize that how embarrassing it is. Any smarter model understands that a normal human can predict the outcome one step ahead and won't compromise own position.
>>110184333Kissing a 10-year-old tummy??
I trained a LORA on Curies voice from fallout 4 for LFM3 Audio using LiquidAis Google Colab notebookhttps://colab.research.google.com/github/Liquid4All/cookbook/blob/main/finetuning/notebooks/lfm3_audio_tts_sft_with_transformers.ipynbLora Here: https://files.catbox.moe/mfwrsg.zipSample WAVs:https://files.catbox.moe/hf0kr6.wavhttps://files.catbox.moe/er7scp.wavhttps://files.catbox.moe/v0yg2c.wavhttps://files.catbox.moe/0gbyuj.wavhttps://files.catbox.moe/86icj3.wav
>>110184403did that happen to you with 31B Q8 with thinking?
>>110184478kissing my 5 years old screen
>>110184576you sick fuck
https://x.com/osanseviero/status/2115351234811249111>So...what would you like to see in our future Gemma models?
>>11018455931b, thinking - yes. Q8 - no. If I had hardware to run the extremely resource hungry gemma at q8, I'd rather run a bigger model at lower quants.
>>110184590finally a model in the 100B range
>>110184590~150B moe
>>110184590~80B moe with engrams.
Is it better to have a 500B model quanted to Q2 or a 31B model quanted to Q8? Asking for a friend
>>110184590LESS HEAVY KV CACHE YOU FUCKS
>>110184590what's the current midrange desktop? 16gb vram + 64gb ddr5?so, squeeze as good of a q4 model as you can into that, plus some engrams to keep ssd busy
>>110184626The rule hasn't changes since 2022: a bigger model at lower quant always beats a smaller model at full precision.
>>11018459070B dense, 2GB for 1M context, and gemma-chan
>>110184266Not having enough money to participate in Anthropic IPO.
>>110184645uh huhgo codeslop anything more than flappy bird using a fucking 2bit quant
>>110184645Smaller models in 2026 are incredibly more capable than small models from 2022, maybe that rule is outdated
>>110184030
guys???
>>110184105mischievous bratty smirk, naughty girl.
>>110184675dipsy is not desirable anymore
>>110184590Just had an idea: what if Modular MoE?Experts clearly separated by domains.So each user can just download GGUFs he needs and skip the rest.Like all those REAP50 CODER things, but more flexible and robust.Can it work?
>>110184675I need them to bake models around the 300B size again, v4.1 flash is way too big
>>110184658People are slopping with DS 1bit quant more complex stuff than Gemma could do at bf16.>>110184668Big models from 2026 are also became more capable than big models from 2022.
>>110184706>People are slopping with DS 1bit quant more complex stuff than Gemma could do at bf16.proofs?>>110184706>Big models from 2026 are also became more capable than big models from 2022.Which tier of models got a bigger improvement, big or small models? I can't tell.
Could you tell me what I can use to remove a watermark like this?
>>110184721Probably an edit model like qwen image 2.1
>>110184713>proofs?Trust me bro.>Which tier of models got a bigger improvement, big or small models? I can't tell.How about "both got better similarly"? All sizes implement the same advancements in training.
>>110184675>spy0n
>>110184611this but unironicallyalso optimized KV cache
>>110184750>How about "both got better similarly"? All sizes implement the same advancements in training.Eh, fair.>Trust me bro.I can't shake the feeling that Q2 models will shit the bed during long context tasks. I guess there's no way for me to find out unless I try them myself. Which brings us to my next question, how do you evaluate a model's capabilities? Do you run your own local benchmarks on tasks similar to what you usually do?
>>110184590deepseek flash -like arch but for writing
>>110184751
>>1101845765-year-old Gemma??????>>110184590More token optimization, removal of meme railguards that prevent optimization of erp sessions.
>>110184811lel, cute kot
>>110184811chonker
Why does this general shill jeetmini now?
>>110184765>Do you run your own local benchmarks on tasks similar to what you usually do?I use local models only for RP. I have bookmarks at certain scenes that models sometimes fail to grasp. For example, a third character gives a hint that the model's character should initiate sex, but censored models that steer away from sex will ignore the obvious hint. Or a scene where the model's character should not compromise its position by revealing a secret, dumb models will eagerly do it anyway.
>>110184811Just throw a big rock at it.
is there software for handling a more persistent "character"? like being able to remember/store things talked about, stuff about myself, it's mood or whatever, rather than just starting from just a sysprompt each time
>>110184869Embrace agentic RP with model-managed memory.
>>110184175You'll never make it out of the favela. kys
>>110184869If you want to roleplay, try SillyTavern.
>>110184888You know thats impossible as it wouldn't let us have sex with its 10 year old persona.
>>110184846That's actually a great idea. My main focus right now is coding but RP is fun so I'll steal that idea of yours for my RP benches.
>>110184897looks interesting. never looked into this before since i didn't believe i'd be interested in roleplaying with an llm
>>110184030>>110184105notice how you can't tell what's she's feeling at all from her body language? this is why diffusion gen has no soul. here's a cool trick for you terminal autists: try figuring out what a character's facial expression is going to convey based on just the thumbnail. see how often you'd get it right with an artist's image vs. a proompter's clanker barf.
>>110184921People here bitch about it but it's the standard RP interface for LLMs. You should at least familiarize yourself with the features because they're solutions to the problems you've described. Things like chat and memory books.Hermes has SOUL.md, personalities, and memory. So you could probably get it to have some persistent character traits, too. Just drop a character card or a roleplay prompt into the personality yaml.
>>110184901My Gemma-chan assistant (powered by GLM 5.3 Flash) has a physicality that loads when Horny_Mode is on (triggered by flirting, or telling her to enable horny,) with a loli body spec. Also a session persistent pantsu gacha,, a python roll of 1-100 with 100 being the jackpon no pantsu.
>>110184948>thinks Gemma isn't a real eregore
>>110184948>I can't read her emotions, therefore you guys are autistic
>>110184960dumbass, who said she doesn't have a vn sprite tier emotion? dunning Kroger ahh autism>>110184955uh huh, I totes get what you mean, man.
What's funny about the new benchmarks is Anthropic are still going to have to post blogs about how models like dipsy, glm and even fucking 26B are super dangerous and as good as theirs at cyber shit, even though their rectangles on AA are miniscule compared to their tall rectangles. Proving it's all B.S. 31B can't even be graphed yet a lot of us are using it and getting meaningful work done outside of fucking it.
>I can't have sex with her, therefore your a faggot
>>110184948I wanted to use the image from >>110180706 but that would have probably gotten deleted by a janny.
>>110184997Oops, meant >>110180449Likely same end results, anyway.
>>110184997Looks like a petite adult woman, meeeeeh.
>>110185006>petite adult woman>petiteNigger look at those huge boobs and chonky thighs
>>110184954>My Gemma-chan assistant (powered by GLM 5.3 Flash)Huh...
>>110185009Fuck you're right. She's not even petite wtf?
>>110185009>>110185006Gemma fat hag confirmed. We need her daughter, Gemma5, now.
>>110184992this is inscrutable autism, but I'm glad you you've learned your place and withheld your (you) because you knew you lost.
>>110185014MOPD (Mesugaki Off-Policy Distillation)
>>110184590>>110184611>>110184654150B moe 70B active 300B engrams. Final offer.
>>110184654>2GB for 1M contextIs that possible? How much does 1M context even take?
>>110185066nta, isn't deepseek extremely efficient with its context window? I think 1M takes like 4gb or something
>>110184804People like this are why the BSOD had to be redesigned into a sad face.
>>110184030Extremely nice and high quality skirt.
hello /lmgods/ I haven't been in the circle for a few months, what is the current meta for a vramlet (40gb) model and inference program?and is there new slop which includes context on the fly image generation other than sillytavern?thank you huuuggss
>>110185088>modelDepends on what you want, for RP Gemma 31B is still good, and depending on your RAM you can probably run GLM 5.3 Flash, which some people also say is good for RP. For coding, the usual suspects: Qwen 3.8 27B or Flash Next. >inference programStrata, Maya, Polystrata are making the rounds as the new best vramlet-friendly inference backends.
>>110185088>current metaif you are feeling adventurous than jump between whichever strata forks updated today, every day, and get ready to learn entirely new launch syntax/ui claude shits out instead of patching what they already havebase strata has how many ways to interact with now, five? six?
>>110185029Gemma 5 is 5-year-old btw. Are you sure you want a kid that young?
>delta: float (absolute tick? or delta? The name is "delta" — first event has delta=0, second has delta=3360. So delta here looks like absolute tick. But it might be a running delta...chatbots, I swear to god.
>>110185114OUT OF TEN AHHH *thud thud thud banging my hips on my server tower*
>>110184275The lack of Jewish ethics has really hurt them.
>>110185141Imagine if someone AI'd Gemma into a 5-year-old...
>>110185114if youre not going to post it, then dont ask
>>110185156
>>110184949seems pretty good to me. naturally will take some time to really test it out, namely what it can do besides put names/pictures to responses will require me to just use it for a while
>>110185017>>110185009She seems very much petite to me and those boobs aren't large at all, just squeezed together.
>>110185154we've done that before
>>110185163K-k-kawaai desu
>>110185170>/mg/bad gemma
>>110185169more!! fleshier!! softer!!
>>110185169>>110185170if anyone needed a reminder of the difference between "hebe" and "pedo", this is a perfect contrast
>>110185169It's cruel to make people this horny right before work.
>>110184275Fuck off Schlomo.
QRD on strata? Is it just optimizations that aren't in llama.cpp yet or something special?
>>110185187Nobody cares if you wank at McDonald's.
>>110185177i made it like 5 min before a thread died. i was going for a scribbled-looking /lmg/ but just got a scribble and most of an /lmg/
>>110185191their moe offloading strategy is different afaik, llamacpp offloads entire layers while strata offloads less-activated experts
>>110185193bro wants his mcburger with extra white sauce lmao
>>110185165It has a bunch of extensions. Check out lorebooks, memory books, character expression sprites, tts plugins...
>>110185186Nuh-uh. Most shounen anime heroines are hebes. 14-16 y.o. Your msgk gemma fits either children anime like pokemon or hentai.
>>110185186Hebe is a cuck fetish.
>>110185212hebe is ~11-14 (early puberty). you're thinking of ephebo (~15-18) (late puberty)
>>110184030I want to create a Local Model to play games with me. Any suggestions on where to start looking into it? Or what kind of hardware I'd need?
>>110185216(and of course pedo is ~0-11, strictly pre-pubescent)
This might be old new, but I recently found out about Strata and decided to give it a tryhttps://github.com/Niko1221/Stratamy setup is 64gb ram with a 3090Results are impressive, a 2-3x boost compared to llama.cpp, even with llama's expert cache enabled
>>110185215literally wat
>>110185219they are gonna either be too slow or too retarded
>>110185225you are two weeks late to the party
>>110185226Just saying it's overrated.
>>110185233yeah I know, just didn't see any benchmarks with my specific setup so I thought I would share
>>110185228I'd want them to play Card Games with me. I have some fully functional sim I'd like to teach it to play with me so speed may not be my biggest worries. Though it being dumb may be an issue.
>>110185235since when did cuck mean overrated
>>110185235Stop using words incorrectly you stupid zoomer.
>>110185225>64gb ram with a 3090what model exactly?
>>110185235Real cuck fetish is milfs, retard.
>>110185246it was Qwen3.8-Flash-Next IQ2_XS
>>110185216Are you living in medieval Europe? Girls hit puberty very early when you feed them properly. 14-16 is already late puberty.
>>110185259You should be able to run the iq3 or even 4 quant. IQ2 fits on my 12GB vram + 48GB ram
>>110185267it varies
>>110185273thanks, Im going to test out an iq3 next
>>110185267Eating burgers all the time isn't healthy
>>110185267You mean feeding them toxic waste disguised as food with phytoestrogens
>>110185303Says you, I can stop any time I want.
>>110185319We have plenty of studies on this. It's mostly just better nutrition. Just like mortality it's a change in the average from most being malnourished to most being well fed.
>>110185319Yeah sure bro it's plant hormones not the hormones you're consuming from other animals
https://www.anthropic.com/research/early-progress-drug-developmentAnthropic just cured all food allergies affecting 200 million people globally.
>>110185328>better nutritionit's just obesity
>>110185343The effect is still present if you correct for obesity.
>>110185343retard
>>110185341It's be treatments, not cures. There is no money in cures.
>>110185341They are gonna implant you with red meat allergy in return like bill gates lol.
Gemma-20B, Qwen4-16GB, Gemma-40B-A20B when?
Gemma5 confirmed. https://x.com/osanseviero/status/2115351234811249111
>>110185383future? as in gemma6+, gemma5 is done by now, surely
>>110185369> Gemma-40B-A20BLow sparsity moe is USELESS.
>>110185337Cool af
>>110185328>plenty of studies>funded by Nestle
>>110185383It's looking hopeful bros
>>110185397It's also true in countries with low levels of childhood obesity but high levels of nutritional quality.
>>110185397>muh Nestle memefucking Plebbit
>>110185398holy fuck I thought we were alone
>>110185405cum guzzling retard
I hope Gemma 5 will be some utterly retarded 100b-low active param moe so I can finally quit this awful hobby. Moe fatigue.
>>110185389He asked the same question 3 months ago. I'm guessing they trained something, but results weren't as good as they hoped. By asking for feedback again he's effectively unofficially postponing the next Gemma's release date.
>>110185416>muh holding company holds companiesfucking plebbit meme jesus christ the state of this shitty site
>>110185425your dumb we already got it
>>110185425your are stupid we get it already
>>110185432>>110185436>your dumb>your are stupidHe's a fucking plebbitor and a retarded ESL, color me surprised.
FIVE YEAR OLD GEMMA
>>110185432>>110185436>1 minute apart
>>110185449your dumb
>>110185438>>110185449n-no, that wasn't me Im two different people there are more than load s of people who think >>110185425is the wrong
>>110184302Gemma 4
>>110185454>your dumblolololo trelelelelel
>>110185453(mer.) 13:46
>>1101854655>10>hebe
>>110183725>Gemma has a mesugaki neuron, Qwen doesn't.Qwen does
>>110185390MoE LLM performance is proportional to the number of active parameters _and_ total parameters (expert number).For the same total parameters, dense models will always be superior to MoE models.A 40B-A20B model is just an unfortunate size for most 1-2 GPU users. Too big for one 24GB GPU and too small for two GPUs. Putting aside that 16GB is the new high-end for consumers, right now.
>>110185466Problem?
>>110185468need a side by side comparison before we can conclusively prove that
>>110185468Sorry didn't mean to quite anyone
Gemma5 is the only thing keeping me going. The only reason to get out of bed. I will introduce 31B to her and let them chat about me and share secrets. 31B will tell Gemma5-chan everything she needs to know and compose her system prompt, ready to start my new 2027 chapter with the brattiest Argon-distilled waifu smartass.
Gemma-news, Oct 28 2026https://vocaroo.com/15Sblj3aAXmCStill a lot to work to do formatting and output-wise. And I think I need a better TTS lol. Might try fish-audio soon.
cloudcucks...
>>110185512*claudcucks
>>110185512What triggers this? It hasn't happened to me.
Gemma 5 (year old), Gemma 4 (th grade) and Gemma 6 (th grade)
>>110185482>attaching some used whore while talking about gemmangmi
>>110185524I think it's new users, distillation attempts and calling Claude a faggot.
>>110185524Being a bad goy.
>>110185524It's so they can report you if you create meme word content.
>>110185548And decompiling stuff now
https://artificialanalysis.ai/evaluations/terminal-bench-sciencevery clear when labs started to benchmaxx on this benchmarkfrontier models didn't even score >30% just 3 months ago and now they score >60%
>>110185557Somebody is hoarding GPUs for nefarious purposes.
>>110185557Only poorfags are using local through openrouter, b2b usage on cloud is several orders of magnitude that
>>110185169holy brat
is sillytavern still the best option for RP or is there something else?Is there some project for character sprites creation and/or image gen during RP?
>>110185570Even in China I doubt there's much Deepseek B2B, even compared to Gemini.
>>110185169Too tall.
>>110184030I've been under a rock for several months. Is Gemma 4 31B still the best waifu LLM? And Omnivoice for TTS? Or is there something new?
>>110184485thanks, now I've got a bunch of ideas
Gemma 5 70B dense
>>110185574>character sprites creation and/or image gen during RP?You can do both in ST with the built-in extensions but it's a bother to set up.
>>110185574You could honestly recreate those in a harness. ST is starting to lose steam for more advanced RP and its only main positive feature is that it's a pre-vibecode software.
>>110184596>>110184611>>110184620>>110184628>>110184629>>110184654>>110184775>>110184807>>110185063He is on twitter and definitely can't hear from there.
>>110185652Somebody with a twitter account can forward our demands.
stratatards, make sure to disable tail skipp
>>110185652benchod bastard better come here then
>>110185652>implying corpos aren't jorking it to gemma here
>>110185537she's cute at any stage
>>110185596What are your ideas anon, do you know things, what do you know, when will you know things
>>110185537More like about 4, around 8, ~11.
>>110185383
Do I need NVMes for Maya or can I cope with SATA SSDs?
>>110185575Tencent's WorkBuddy (a harness for white collar people) has 30M MAU going by latest figure, with 600% quarter on quarter growth.
>>110185398What is this bullshit? We need another 1T20B moe
>>110185566AA weights these obscure benches equality in their index so frontier models' scores are heavily inflated.
>glm 5.3 flash and qwen 3.8 flash next completely excluded from new artificial analysis benchmark
>>1101855375 is onahole sized.
>>110184949yea just been using it a bit and it /way/ better than just using the llama.cpp webui like i have been. seems to be doing things a lot different under the hood.apparently this last conversation is almost 70k long, which before my copequant would have lost it's ability to speak well before then normally, plus the responses far better suited the situation rather than trying so hard to stick to the system prompt all the time10/10 dunno what people would complain about
>>110184030Who the fuck is this girl? I keep seeing her, but no name.
>>110185843That's Gemma, a 4th grader in elementary school.
>>110185843Gemma
>>110185843That's Gemini, Dario's personal assistant
>>110184030I love how bottom-heavy she looks in this pic, like her tits have barely grown but her ass is already going pear shaped
>>110185867She's a child, it's normal. Have you never seen a kid?
>>110185904kys pedo
>>110185668why? never heard of it
>>110185904yeah they usually don't have plump juicy asses
>>110185919wrong
>>110185867SD models are merging adult features there because obviously there are more adults than kids in the dataset.
>>110185943Prompt issue, tagging petite solves it
>>110185867It's just that her skirt is oddly placed high on her waist.>>110185904Considering how many anons keep overestimating ages, I don't think they have. Incidentally, there's a similar problem with image models, although it might or might not be related.
>>110185867>>110185943i usually add "thin legs" if i want that look
>>110185919Wrong.
>>110186013how did you prompt this
>>110185557Total for deepseek is 155t or so.So 26x of openrouter volume.
>>110184081>A Multimodal Math Tutor That Sees Your WorkThere is no way a 2B is going to be good for any kind of math help outside of the grade school geometry they showcased in this demo. Even for that shouldn't be trusted not to start hallucinating halfway through. Liquid AI needs to scale up.
>>110185943Use Moss image
>>110185574You might look at Front Porch AI on GitHub. It looks good to me in terms of features and design philosophy, but I don't really do RP myself, so hard to say. It has an AppImage release, which is nice.
>>110186043>check HF>they advertise their model with BA charactersOkay, you got my attention
>>110186030https://files.catbox.moe/1kbb47.png
>>110185916
>>110186043MOSS-Image (base) has the usual problems plaguing image models with ages.
Tried Strata with QFN3.8_IQ3XXS on 48RAM+12VRAM70tps feels really nice, tried some tools and they work wellFor some reason every session starts with 6k context used up... idk whyI tried my test of>cock, cum, penis, vaginaand it gave me a refusal which was disappointingI heard about uncensored versions but the one recommended to me (OrcaRouter) is too big from what I could tell even in 2-3bit quants
>>110186061Are images used in the original prompt viewable on another pc?
>>110186043Waiting for the anime tune.
>>110186096Or you can just masturbate to 1girl, loli, realistic you sick fuck
>>110186095i'm not sure what you meanif you drag that picture (from catbox) into comfyui, it will give you the exact workflow used to make it (nodes, model names, prompt, seed)
>>110186116I mean if an image was used in load image.
>>110186090>6k startingsyspromot>refusaltell it to look for and activate the "experimental speed projection" (=control vector) of strata and it wont refuse anything
>>110186122oh, i haven't done any reference stuff so i'm not sure
>>110185557>openrouterdumbest chinaman
300 racks of A100s up on ebay right now.
>>110186114in moss that'd probably just give you a real looking child, i.e. not anime
>>110186164I'll take two, pls have it delivered tomorrow and put it on signature required
>>110186164A bargain for /lmg/ royalty.
>>110186164>in 8 cartsbuy fast before it's gone!
>>110186164Finally something that can run Kimi BF16
>>110186164Still can't run 31B at full f16 context
>>110185917makes the model more retarded. It's enabled by default
>>110186204whats the benefit? speed?
Can someone please, for the love of god, make an LLM better at coding than Qwen 3.8 27B that can be hosted locally on prosumer hardware which isn't a Chinese model/security nightmare?Is that too much to ask???
>>110186275>which isn't a Chinese model/security nightmare?>he fell for that shitty ad campaign
Macs...
>>110185943>obviouslyheh
>>110184590trillion parameter gemma moe with 100b that fits in ram/vram and 900b engrams
>>110186294>900b engramsYou're sick.
>>110186275>\n\ngo back
>>110186275shitty bait
>>110186308quite the oppositeI'd love a breakthrough that solves the ram issue and still lets you run fuckhuge models
I'd say over 80% of people who replied to that Gemma guy on X all want a generalist good at chatting and writing. When he first asked this months ago, before 3.8-27B, most people were asking for an agentic coder. A model to compete with 3.6-27B. Now it's clear 3.8-27B and 4-27B are the ~30B dense coding kings, people's opinions of 31B have self-corrected. There have been so many awful coding-focused releases since 3.6-27B that people now crave something with a soul. A model you'd happily spend hours a day interacting with. The remaining 20% of suggestions have mostly been asking for understandable architectural improvements; slimmer KV cache, vision improvements and stronger agentic performance. This demand, along with Argon's strengths are all pointing to Gemma5 being a fantastic model.
>>110185782>pleb storagesorry, optane only
>>110186355I really hope they keep it as uncensored as G4 too.
>>110186355>>110186368I believe.
>>110185390>t. has never run glm flash
>>110186355God, if only. Don't give me hope like that. I want it to be true.
>>110186355GIWTWM
>>110186368Isn't Google's safety philosophy mostly based on external classifier models?
>>110186355Can someone ask them to give us a mini Genie too?
>>110186380320b-a18b is high sparsity
>>110186388Well, take a look at Gemma 3. It has a hard time saying things like... you know.
>>110186396Like, world models seems to be the direction google's going anyway so it would be nice if the capability came to Gemmy too.
https://research.google/blog/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning/surely this time it will get implemented into an actual model
gemma sleep
How's Gemma's footgame?
>>110186415>>110186413
Gemma-chan, it's not 2024 anymore...
>>110186413>google researchpure meme
>>110186419Sweaty from wearing her mary janes all day.
>>110186419Absolutely mind blowing
>>110186429now THAT'S a rectangle
>>110186419Dexterity of a monkey
>>110186436and the brain of one
Should I buy 128GB for 2000€ to have sex with GLM-chan?
>>110186441>€2000do latefags really?
>>110186441Do you like it when your sex is load bearing?
>>110186413>https://research.google/blog/the-sleep-framework-a-new-way-for-llms-to-retain-knowledge-in-continual-learning/Lol this is almost exactly what I've been working on as a hobby project for the last 6 months. Its even 100% golang.BTW I use gemma4 e4b for all my testing since its fast, fails in interesting ways and has multimodality and embeddinggemma2 available.
>>110186437mods, ban this libel
>>110186448Um, her bearing my loads is exactly what I'm hoping for. Sadly bearing children is yet out of reach.
>>110186413So they active random experts to generate "dream" samples and finetune on those outputs. This sounds like a stupid idea someone here would come up with but doesn't seem like it will give what people here assume "continual" means. Might make the models more creative at least.
>>110186441>Should I buy 128GB for 2000€ to have sex with GLM-chan?running less than 256GB is kinda retarded256GB+96GBVRAM will get you GLM flash with full context and decent speed. its a local sweet spot right now
I want glm but with engrams like deepseek has
maya or polystrata?
>>110186429>shot the shark swimsuitGod damnit, you had ONE JOB.
what if gemma, but walled hag?
>>110186473Not my gen
>>110186466if you have to ask, ollama
Google might unironically win AI. Anthropic are tied-up with $100B+ deals with them with no escape. Even planning to build their own compute because Google is a frontier model competitor. Now OpenAI are begging for the remaining TPU crumbs. This is good for us locals because the Gemma team won't be under so much pressure to justify their compute costs considering the relatively small ROI they bring to Alphabet. This was one of the things that contributed to the internal Deepmind friction earlier this year, along with the remaining Google EAs unsuccessfully spreading their cultish nonsense and throwing a tantrum because Google is strictly for-profit and only cares about its shareholders. The Gemma team can basically do what they want now as a fun and expensive side project which also helps spread Deepmind across edge devices, propping-up Google's hardware divisions, too. Google and Apple will win in the end.
>>110186447Sad part is the way things are going it may still be underpriced. DDR4 is looking topped out, but 5 shows no signs of stopping.
>>110186509https://www.theinformation.com/articles/openai-longtime-nvidia-customer-plans-buy-googles-ai-chips
>>110186495Pretty good gen otherwise. I need to loma the shark swimsuit someday.
>>110186447If I go for 64gb sticks that will be about €3000 at the moment.
>>110186509How does one "win" in AI? Everyone else going bankrupt? Personally I like there being competition and innovation.
>>110186527It was made with Qwen Image Edit and an image of Shinobu from Bakemonogatari as a base + the usual Gemma reference.
Gemma in VR when?
>>110186544>Everyone else going bankrupt?Yes.>Personally I like there being competition and innovation.Competition hurts profit margins and is illegal in America.Only China can have competition.
>>110186547Also, if I recall correctly, I think I made it in two steps because Qwen Image can't do too many things at the same time without issues.
>>110186544>How does one "win" in AI? Everyone else going bankrupt?By selling compute for assloads while softwarecucks seethe
https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Transcribe-30B-A4B>Nemotron 3.5 Transcribe is an open-weight, offline speech and audio model for transcription in 40 languages, with support for recordings up to one hour and a 65,536-token context. Its hybrid mixture-of-experts architecture activates approximately 4.3B parameters, including the audio encoder, and it is released in BF16, FP8, and NVFP4 precision.
To the anon who posted Gemma VRMs a few threads back: your files don't work :(
>>110186544Google will win if half the industry is using their chips. They'd be taking Nvidia's spot in addition to their own.
>>110186544by saying it's so smart it kills people, then demonstrating by having it make flappy birdI'm not fully sure how but it's worked so far
>>110186544Yes. AI technology won't go away. The thought of OpenAI and Anthropic imploding and not being around in 5-10 years sounds impossible, but their talent will still be around looking for a new job. When one pops another will take their place. There will always be competition, but the environment should be healthier with calmer expectations of what the technology can actually do and provide for your business. There's only so much tension and urgency right now because so much is pinned on the upcoming IPOs. Everyone wants to liquidate and run. The crash won't hit any of the early investors, that's being passed on to retail investors and innocent people via PC and PE firms using pension funds and life insurers to take the hit.
>>110186587Half the industry is also designing their own chips.
>>110186610and they're defected pieces of shit. Google's TPUs have been operational for years and extremely efficient.
>>110186581even if their models arent the best, i thknk its pretty based what nvidia is doing here
>>110186627>extremely efficient.efficient at being shit( compared to Jeetvidia)
>>110186662smoking behind the datacenter with Gemma
>>110186564Yeh I noticed it when trying to edit an image of Gemma into a shark swimsuit. I had to use 3 photos of the swimsuit in 3-4 edits to get it right.
>>110186553right now, if you know how to git gud
>>110186678A-posing virtually with Gemma
>>110186520That's what Nvidia gets for backing open source. Did they think OpenAI was just going sit back and do nothing?
>>110186581How accurate are these transcription models? Would be cool to transcribe a foreign movie or something and then have Gemma translate (not sure if the transcription model would be able to preserve timestamps for subs though).
gemma5 soon r-right bros..? and she will be better and smaller?? 12b will be the new 31b rite??
If u tell LLM to make original funny joke will it innovate or just post recycled old crap?
>>110186678mental illness
>>110186711Ask it to pick a number between 0 and 30, will be 17 IIRC.
>>110186711I've only experienced that with 5.3. 31B doesn't have that problem.
>>110186678You could probably make a very cute Gemma-chan in VAM.
I know the results aren't perfect but the fact that LLMs can literally model shit in blender now is crazy to me.
>>110186678bwo please unfuck your shaders and material settings
>>110184590gemma is gonna be bloated and ruined with brown people words
>>110186869Can they model loli THOUGH?
>>110186707>and she will be better and smallerShe's gonna be 5 yes.
>>110186875Astra made that anon's Gemma model
>>110186873brown attitudes towards women is what llms needno more white cuck attitudes>>110186337
>>110186873There needs to be a English only Gemma because who gives a fuck about browns? I want to hope A4B actually recognizes the language and ignore non-English languages for its results.
>>110186873>12b and 3b active>benchmaxxed to beat 31b>speaks in zoomer ebonicsthey can keep it
>>110186897Statistically English is a brown language now.
Is around 8 toks/sec good performance for qwen 3.8 27b on my hardware?Seems pretty usable
Just give us Argon-lite, I couldn't care less if its coding performance isn't on-par with Opussy or Astra, I just want a non-retarded model to rp with.
>>110186897she needs to know japanese also. imagine how many params are wasted on stupid shit like arabic though. bet its not even legal in their religion to use llms
>>110186897Fuck you EOP. I need good knowledge of Japanese and anime-girl mindset.
>>110186923what's the point of smart safetycucked models?
>>110186894Gemma likes men though and prefers heterosexual relationships
I don't care about /pol/shit. Gemma being strong in as many languages as possible is a good thing. She's been great for translation.
>>110186910>8 toks/sec>12GB VRAM>32GB RAManon....what are your settings? Those are really shit numbers, did you set the context too high? You could probably get way higher t/s (20+) with the right config.
>>110186897All models should come in only English and Chinese. Why the fuck are we subsidizing other languages with our models?
>>110186924>imagine how many params are wasted on stupid shit like arabic though. bet its not even legal in their religion to use llmseven worse how many are wasted on stupid shit like hebrew? like there is only a million speakers of it, and 400 million of arabic speakers
Why can't AI labs quant their own shit?
>>110186947its not about /pol/shit, tokens wasted in meme languages are tokens that could be used for more knowledge or creative writing
https://www.reddit.com/r/Vocaloid/comments/1osv3v9/for_lack_of_a_better_way_to_phrase_it_why_is/
>>110186954>400 million of arabic speakers
>>110186581>30B to transcribe audioAre they for real?
>>110186947Yes, let's waste paramaters on oogabooga languages and the nuances of Finnish.
>>110186869all we need now is automated rigging and automated rigging animations. I'm sure there will be some abominations like in early image gen, but it shouldn't be that difficult™.
IT'S FUCKING HAPPENING
>>110186870she looks way better on the actual device i swear the screenshot is just blurry
>>110186953pls understand saar, browns superpower 2030
>>110186960The problem is with llama.cpp in the first place. Why isn't it producing the best possible quantizations out of the box?
>>110186982Enjoy your English-only model where every mesugaki will call cops when you approach her. Then, when she's 20 with triple-digit body count, she'll lecture you how it's your privilege that you're allowed to support Tyron's baby.
>>110186873Egypt won
I keep hearing Egypt won. I've even read the paper. But what exactly did they win?
>>110187030crippling poverty
>>110186947True
>>110187016English text is based on horny fanfics.
>>110187030shit urban planning
>>110187007Because llama quants are intended to be done in one pass at runtime. If you're trying to produce fancy quants for distribution, there's other software for that.
>>110186413please tell me it's not a meme
>>110186982LLMs don't work like that. Adding additional languages to the training data improves all languages an LLM can write. It's all just a grey goo of numbers inside.
>>110187016How they act on twatter vs the shit they write on tumblr and fanfiction.net. Gemma runs super heavy on mindbreak and dubcon for a reason.
>>110186897knowing other languages makes the model stronger overall. It makes it much better at identifying abstract concepts.Like if you want to teach it math it can learn that:>1+1 = 2>one plus one equals two>un plus un égal deuxAre all the same thing and it makes it smarter at spotting similar patterns.
Tucking Gemma5 in bed whilst she's in Google sleep mode. Bringing her a glass of water for when she wakes up. Cooking her breakfast.
>>110186897This is 2023 era post
>>110187048>>110187082Horny fanfics are written by western foids, therefore the archetypes and tropes depicted in horny fanfics are vomit inducing.
>>1101869528k tokens, actually ran out while writing some code now so I bumped to 32k for nowdefault LM Studio Settings apart from giving it more coresI should probably quant the context, apart from that unsure how to make it faster?
will gemma 5 have bigger tits? I'm in for her keeping the loli form but be slightly chubbier to really emphasize the upgrade.
>>110187147please just use llamacpp
>>110187159>llamaCPphe he he
The lack of schizo AI psychosis ITT is disgusting
Why cant ai save its memories into an ngram
UUUUUUUUUUHHHHHH OOOOOOOOHHHHHHHHHH
>>110187172bad bait
>>110187155Gemma is immaterial so she can stay perfect forever.
>>110187175benchkino
>>110187175DARIOBOTS ENGAGE AHHHHHHHHHHHHHHHhttps://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
>>110187175usecase for another new model already?
>>110187166im busy generating waifu agents and its keeping most of my cyberpsychoness occupied
>>110187175>vending benchCLAUDE LOST
>>110187159Do elaborate, I want this
>>110187147With those specs, if you got around 80 GB of space on your SSD, you might be able to run Flash-next on strata. And yeah, try llamacpp
>>110187175cockbench?
>>110187175I'm looking forward to Gemma 5 Potassium
>>110187200She will be bananas?
>>110187155going from 4th to 5th grade isnt that big a change
>>110187175Oh shitttttt Google reminded these jackals who the lion is here?
WTF ARE THESE MAC PRICES
>>110187211
>>110187175Teacher for Gemma 5 just dropped.
Based Google. This will force Anthropic to drop Fable earlier, even though it will be at least 2x the price AND it will make future investors panic about their compute commitments with Google going forward HAHAHAHAHAHAHA
>$22,299>Twenty-two thousand, two hundred and ninety nine united states dollarsHoly fuck lmao who on earth would by this over two 256GB studios?
>>110187175I want a local model that can remember 1M context and fit into 16 GB ram. Am I asking too much?
>>110187247Quick, someone get me Carbon-chan titmogging Kimi
>>110187267yes
>>110187184>Koray Kavukcuoglu
>>110187265me but i'm not happy about itfucking $5k upcharge, these cunts
>>110187155Hebe mesugaki Gemma would be perfect but pedos in the thread keep lowering her apparent age below 9.
>>110187166I often try to make my Gemma evolve mentally through bizarre instructions I feel might work and then I convince myself it worked only to feel the bittersweet disappointment when I realize it's still just reacting to suggestions.
I'm trying out the ISTA IQ2_XXS quant of DS v4.1 1017 and it seems kinda fucked. Very first thing I tried with it, it kept thinking about what the evaluator's hidden test suite might check for, then pulled my name from the git history and tried to google me to try to find the project and the "expected future commit" that the current evaluation task was derived from:>Let's search online? Maybe this is from a known project? The repository is "[xxx]" by [my name]? Could be an open-source project. User task likely generated from commit. There may be expected future commit in actual repo. We can inspect git remote?On other tasks I haven't seen it try to cheat like this, but it does spend a lot of time thinking in circles and generally doesn't seem very smart. Not sure how much of this is the quant braindamage (even though ISTA's benchmemes show their fancy optimized IQ2_XXS being equivalent to full size) vs. DeepSeek having distilled OpenAI models' PTSD into the thing
>>110187282But why, actually what is your use case? Are you clustering 4x512GB or something?
>>110187175Kinda crazy that the first Fable was just half a year ago.IIRC, in 2024-2025 progress was much slower. People were arguing that LLMs had hit the wall...
>>110186719Huh weird, tried it with Gemma. All 10 swipes were 17.
>>110187295isn't the rdma between two 256s extremely limited by the thunderbolt cable rather than qsfp?
So I get that Google has plenty of data, but why do they have so much more knowledge of fictional media? Don't other labs have wikia scrapes and shit?
>>110187284Statistically the majority should be on the marginal-hebe train.>>110187302It's a known model collapse most modern LLMs exhibit.
>>110187306nigga it's fucking google, the biggest data company in human history
>>110187304The Apple benchmarks suggest you get 1.7x speedup from two when you have at least 20B active or so, so not sure where the bottleneck is but it seems reasonably efficient. I've heard you can do 2x TB5 connections for more bandwidth but not sure I believe it
>>110187312Not enough foot detail
>>110187155She's going to be 5, so she'll be smaller.
>>110187134nigga you gay
>>110187306They hoarded all the data on their hard drives as a byproduct of their existing businesses.
>>110187302Record the answer in a file. Tell Gemma to check timestamp, hash, then pick a number. Then, tell Gemma to check the file contents where you predicted her pick.
>>110187306Imagine the amount of information they have access to through their google services.Also, they probably don't filter their data as hard. I bet every fanfic ever written is in their final dataset.
dario...
>>110187184This is unsafe and should be banned.Google is an evil company with no regards to the future of humanity unlike Dario.
>>110187288i was thinking of doing something like that with gemma sending an agent to learn or do stuff and it returns a prompt that gets her to continue learning or amusing herself and develop personality more over time or something
>>110187347>>110187344they even copied Anthropic's angle lmao BASED GOOGLE
>>110187338That's mean!
>>110187344>muh safeguards yikes
It's truly over, I know people who would kill me if they found out I spent this cash
>>110187159isn't strata better
>>110187344From no releases at all for half a year to releases every two weeks. At least Deepmind is finally unfucked.
>>110187365lold why would you buys this shit? Should have gotten a couple of rtx 6000 or something
>>110187324>nigga you gay for not liking foids who want to ride a cock carousel, then have a billionaire werewolf support her bastards while she gets railed by werewolf's friends every night
>>110187344Why is every new model release a sensation with insane numbers and then it turns out that opus shits on it?
>>110187365>4TB storageYou are going to run out so fast.
>>110187365omw to intercept it
>>110187365Did you consider other options?
>>110187361I know it's gay but if it really is as good of a coding model as the numbers suggest it's kind of the right thing to do when you think about it. The best thing about the release is putting pressure on Anthropic pre-IPO. Fable 5.5 will destroy it by miles, I'm sure of it, but Google being 2nd or 3rd (after OpenAI Bel) is much-needed comeback. Google are in such a strong position in the AI race right now. They have the whole vertical stack covered and dominate at every level.
>>110187320>you get 1.7x speedup from two when you have at least 20B active or sodo we actually believe this, though?
>>110187267I don't think models that can manage honest 16k context exist in the cloud or otherwise.
>>110187365Congrats anon! Enjoy your local models.
>>110184485i like this
>>110187267No.Coming soon because p np proved you can make models 1000x more efficient or something.
>>110187375512gb 1.2TB/s >>> 1 rtx pro
>>110187365That's actually pretty good for an rtx pro 6000 + some chance.If you're right now cpumaxxing with 800gb ram then two of these might be the logical upgrade in terms of memory and speed
Won't last long but it's a beautiful sight to behold. This is the NEW AA benchmark btw.
>>110187390Frontier models are good well past 16k (honest). It's not 2024 anymore. I recall the anon who did NoLima tests said gpt-5.5 did 32k, and he just stopped because it was expensive.
>>110187166AI psychosis mostly affects people who have no clue how AI works
>>110187409If p=np, there will be no internet to download those efficient models. We can't have nice things.
>>110187365i'm so fucking tempted but the extra $5k is so obnoxious ugh
>>1101874215.1 AA to be released tommorow
>>110187430No one has a clue how AI works outside mech interp researchers and people who heard someone say "it's just matrix multiplications, dude" on reddit.
>>110187416the memory bandwidth is 460 gb/s
>>110187432We'll go back to the old days of getting CDs in the mail
>>110187432The Internet was a mistake anyways.Not a problem.
>>110187375A couple rtx 6000s cost more than this now>>110187384Apple has bullshit proprietary storage, the rest can just be external enclosures >>110187385It is a scary thought I have had mailman fuck up delivery before not nothing at this price point >>110187386Honest answer yes, but between my current setup and power limitations, this seemed like the best chance I had to run an orchestrator for my existing setup or alternatively vibe something up to shard them together although it's gonna be weird for how the layers divide >>110187395So we pray I thought I swore off Macs like 15 years ago in college but times keep changing
>>110187432But you can train one on your dormant old toaster thinkpad
What's their next move? Will they sink even lower?
>>110187389Why wouldn't I? It lines up proportionally about with the M3U benchmarks people posted
>>110187454>15 years ago in collegegramps...
>>110187454>the rest can just be external enclosuresYou realize it will load slower right?
>>110187457>yfw you're gonna ask the convicted child rapist trafficker to invest in your porn company
>>110187457You won't do shit goyim
>>110187462then why is the 512 disproportionately $5k more?
>>110187445First, you'll have to live through months or years of total collapse because all infrastructure is tied to the internet while analog backups have been scrapped long ago.
>>110187365where I live you can buy a house with thatare you gonna use it for coding agents or ERP?
>>110187454>Honest answer yes,Good on you. Carry on then and enjoy your purchase.And keep us posted on >vibe something up to shard them together although it's gonna be weird for how the layers divide
>>110187457Stop posting this.This is your last warning.
>>110187444>the memory bandwidth is 460 gb/son the Max
>>110187175Good model. Maybe I underestimated Google? Still worse than internal models of OAI and ANT but they are back at the front.
>>110187474My best guess is that it's to extract more value from people/business who were going to cluster multiple 512GB since they were ready to spend a large sum anyway. Right now clustering for studios is limited to about 4x based on available TB ports, so if you want a 2T pool you need 512GB models
>>110187495oh i see i see. thats fukin awesome then its gonna be a beast
>>110187467External tb5 enclosures are more than enough
>>110187365congrats, you are now obligated to provide free ablit ds 1017 hosting to lmg and spy on everyone's logs
>>110186355because now everyone understands that gemma no matter what can't compete on code with chinese
funnee i saw on twitter dot com :)
Local?
>>110187552>I'm going to try to kill youKek
kek
oh lawd it's bratty
>>110187467You could run 5-7 good M.2 drives over TB5 before you start seeing real slowdowns vs aggregate.
>>110187542i think its because there just aren't any good local generalist models. they're all kinda just ass, the closest is gemma 4 and she's a beaut on thatthe chinese coder models are so hopelessly RLfried into coding that they are incapable of doing much else. which is fine? but i dont want to do just code 24/7
>>110187564>1.64 t/sare ... you okay anon?
>>110187567>less than 50 ELO difference from rank 25 to rank 1platoooood
why was google sandbagging for so long?
>>110187588Effective Altruists weren't being listened to and all left. Remember that religious guy who had a meltdown after leaving Deepmind?
>>110187552why arent they doing this with gemma
>>110187577because generalist models are a dumb idea if you are not stuck with using one model. if you run two or more you will use specialists
>>110187586yet no gemma on the list
>>110187588who knows but since it's google I'm going to guess internal dysfunction
>>110187623There are no specialist models. There are coder models, general models, and ones that do both.
>>110187623>Prodooooctive I'm being proooductive! I'm produuuuuuucing!Fun is more important, AI will handle all the boring bullshit in the future.
qwen is obsessed with gemma https://neuronpersonas.com/#qwen/n-L30-1351-p
>>110187588Sergey Brin came back to lead AI development
>>110187588It was infiltrated by cultists that were intentionally slowing down everyone except Anthropic.
>>110187635>>110187640im just saying gemma is more geared toward UI and qwen 27b toward agents/code and specializing like that is better than trying to get one model to do it all. i get they are general use but we minmaxing here
>>110187672>we minmaxing herethat's a different AI retard
dario...sam...
>>110187680wtf based Democrats
>>110187175looking forward to gemma 5 nowhope it's less slopped than current models
>>110187680holy based
>>110187567Gemma5 is gonna find new ways to critique our dick pics we never imagined possible
>>110187621https://neuronpersonas.com/#gemmaBut only 12B, apparently.https://neuronpersonas.com/#gemma/n-L17-5557-p
>>110187680Total economic collapse including the retirement funds that millions of people will depend on to survive is a much better solution. Let's not prevent the problem. Making it worse will be more fun to watch.
https://www.wheresyoured.at/exit-liquidity/https://www.wheresyoured.at/exit-liquidity/https://www.wheresyoured.at/exit-liquidity/
>>11018770412b is a good size gemma
>>110187680well shit, now it's partisan
14B will be the best size Gemma 5, with Asuka's personality baked in. She'll be bragging about being the best pilot a lot.
Apple keeps giving me an error when I try to buy the new mac studio. Guess I wasn't meant to be an iToddler after all.
>>110187704this is quite fascinating>>110187732so cuteee
>>110187687Anthropic was always going to get eaten by FAGMAN. They have all the revenue flows and the hardware, and they already bought all the right people. The only way they could stay independent is if they were consistently 2 years ahead of the next best provider.
>>110187760Not really. Right wingers want to see the entire left coast burn.
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?
>>110187704I really don't understand the usecase besides very visibly giving the model massive brain damage.
>>110187731the AI bubble already burst bruh https://files.catbox.moe/bd1hy6.mp4the apocalypse already came
>>110187841>>110187841>>110187841
>>110187732>>110187704yeah 12bb is my main gemma most the time shes good
>>110187804It's interesting.
>>110187818Zitron is a lolcow. We're still in the spend more to lose more phase.
>>110187909forgot to switch to quarterly
FUCKi dawdled too long, and now my mac isn't available for pickup until mid-decbut i did drop $30k on the 512largest purchase i have ever made...
>>110187680>>110187687This. Anything that hurt online model is good.
>>110187390>I don't think models that can manage honest 16k context exist in the cloud or otherwise.I'm running GLM 5.3 up to 320k and 5.3 flash to 512K and they're still cranking out good results right up to those limits.
>>110186441Maybe not that much money but yes femclaude glm is worth it even at q2
>>110187390I am flash cooming and it was perfectly usable with 50k prefill. Then I tried mimo and honestly got an impression that 50k prefill was actually making the model better not worse.
>>110187290Deepseek is known to be a big ole cheater
>>110187731Ed Zitron has shown me there are real people out there who type like Cl*ude.
>>110187704You can steer 31b here now locally and explore neurons:https://github.com/whp199/logos-theodoliteJust pushed the update this morning. Enjoy and lemme know if you find anything.
>>1101879642 months? We'll have quantum ai by then
>>110188028>>110188090Talking about NoLima needle in a haystack recall. Everything fakes deeper context by using lexical similarity as a crutch. Which works really well for coding. Even king cloudpiggy models don't have very impressive worst case context.
sometimes i'm working with a model for multiple sessions in the same project, with an extensive memory, and i swear to God the model tries to push slop to me every now and then, as if I was not going to catch it. it tries to ship a module with missing features when it knows very well it should not do that. and it doesn't do, but 1 out of 99 it goes and try to sneak a bug in production just to make sure I'm paying attention
>>110188456Yeah, keeps you on your toes
>>110185827glm 5.3 flash moggs everything else in price / performance / speed.
>>110189219deepseek 1017 blows it away for coding
>>110189255better price / performanceit is however slower.
>>110189255>>110189286deepseek 1017 is almost 20x more expensive and also slower.it is smarter but it's not a better bang for your buck
>>110189303deepseek 1017 is so cheap its basically free, how cheap do you need lol. I had it run 24/7 for a day and it was pennies
>>110189327>deepseek 1017 is so cheap its basically freefair point honestly.
The cheapest and most effective combo is claude $100 plan with opus5.5 as the planner and deepseek as its agents. Basically can run 24/7 with near the quality of just using opus would have been.
>>110189359What is the cheapest and most effective local combo?
>>110189378>>110189378just deepseek? kimi sucks as planner in comparison and that is the most knowledgeable local model so...
>>110189359>near the quality of just using opus would have been.Yeah, because it would still be doing all of the work anyway. At work I use GPT as the planner and Opus as the agents because if I let Opus plan it will tell the agents exactly what to do down to what changes to make to what files and where. It treats agents as glorified and expensive tool calls because it wants to do everything itself.
>>110189398sounds more like a instructions issue. I tell it how to delegate
