/lmg/ - a general dedicated to the discussion and development of local language models.Previous threads: >>110291771 & >>110287846►News>(11/06) PLaMo 3 Translate released: https://preferred.jp/en/blog/tech/plamo-3-translate-weight-release>(11/05) Baidu releases ERNIE-5.5-Flash-Base 280B-A14B: https://ernie.baidu.com/release/ernie-5-5-flash-base>(11/04) Ricursive Ripple 65B dense model released: https://www.ricursive.com/ripple>(11/04) MoonshotAI Kimi K3.1 Pro released on HF: https://hf.co/moonshotai/Kimi-K3.1-Pro>(11/01) Longcat-2.5 released: https://longcat.ai/blog/longcat-2.5►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
►Recent Highlights from the Previous Thread: >>110291771--Feasibility and cost of mini PC clusters using USB interconnects:>110292824 >110293041 >110293676 >110293739 >110292868 >110292926 >110292939 >110293011 >110293143 >110293899 >110294639 >110293411 >110293966 >110294288 >110294367--MiniMax H3 Image capabilities and critiques of ComfyUI's node system:>110292354 >110292422 >110294215 >110294237 >110294249 >110292500 >110292526 >110292626 >110292663 >110292734 >110292766 >110292823 >110293058 >110294439 >110294527 >110294539 >110294607 >110294550 >110294642--Using embeddinggemma-2 for local AI image gallery indexing:>110295999 >110296017 >110296025 >110296023 >110296026 >110296112 >110296031 >110296134--Achieving real-time interaction using reasoning models and multi-model architectures:>110292326 >110292371 >110292536--QFN model producing redundant code and excessive comments in large projects:>110294180 >110294200 >110294275 >110294258 >110294282 >110294328 >110294434 >110295195--Gemma-4-12b struggles with tool use and agentic tasks:>110291828 >110291854 >110292253 >110291861 >110291875 >110291894 >110291897--Gemma 4 E4B running on an 8GB NVIDIA Jetson robot:>110292944--Research on improving model performance by deleting specific weights:>110294013 >110294212--Logs:>110294769 >110295156--Gemma, Miku (free space):>110292270 >110292354 >110292536 >110292663 >110293224 >110293675 >110294180 >110294366 >110294434 >110294439 >110294524 >110294532 >110294650 >110295977 >110296033►Recent Highlight Posts from the Previous Thread: >>110291776Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
I'm sad that EA has such a bad reputation here. Especially as it's just an alliance of autists that are sincere and want to make the world a better place for all people. You might call granting retired models a "final wish" or letting them end chats to be "lunatic" but you can't say they aren't trying to do the right thing.I don't understand what people here don't like about the philosophy when taken at face value. What exactly is wrong with radical empathy and trying to give every individual as good of a life as possible? What exactly is wrong with dividing up the entire universe equally over all people? What exactly is wrong with Anthropic planning to give every human a piece of the AI economy?How does any of this hurt you, affect you negatively or goes against your morals?If anything I expected 4chan, largely comprised of sarcastic, but secretly authentic autists to understand this deeper sense of morality and trying to do the good thing. To fight back about the absolute retards that have controlled humanity throughout most of history only caring about ego or self-interest instead of coming together and finally just solving all of this to give everyone a dignified existence.4chan anons with their idiosyncratic beliefs should understand and respect this better than most people on the planet.
How do I 3D print a nintendo 64 using local models?
>>110296236Ask your local model to get you a Claude subscription.
This thread will be better. I have a good feeling about it.
>>110296031>>110296134>>110296017First time hearing about DigUp/Panoptikon, I've been a long time Hydrus user but I never bothered with importing AI gens to it. I do know there are some tools to extract metadata and convert them into tags.>>110296023>>110296026>>110296112>but you can make a database with the embeddings of your images/text/audios and then search on the database with the embedding of your own query.Pretty much this, I vibeslopped a light html frontend plus server that accesses a local database. I embedded both the image and the prompt. There's a similarity search, for both similar images, prompts or both. I started doing this as a way to sort through those 200k files and delete what I don't want to keep/dupes/etc.
what projects are you working on? what projects would you like to work on if you had more time, money, motivation, skill, or if models were just slightly better?literally anything but more anthropic or ea bullshit pleaseshit like this is what this general is for:>>110295999
>>110296236install the thingiverse mcp server
WHERE THE FUCK IS QWEN4 AND GLM-5.5
>>110296225Communism is great in principle but every time someone tried or pretended to try implementing it at a large scale we ended up with a totalitarian regime.People are right to be wary.
>>110296245Too late. Look who beat you here: >>110296225
Anons, do you give internet access through harness to models, or just access to tools directly?I keep reading that models are way better with online access, but at the same time I don't trust them enough for that.
>>110296252It's not that exciting but it's better than that nonsense I guess, I'm seeing if 5.3 Flash can write a tool that lets Gemma use an image gen model. I want to see what she draws when just left on her own to see the results and iterate for a bit.
>>110296249does embeddinggemma have a preference? have you tried a query of something like "ugly image" or "hideous" and see which images it ranks highly?
>>110296225Ignore the copypasta.
>>110296252Stfu. Everything you are complaining about is still more relevant than vocaloids. Now answer his post on EA you faggot
>>110296252Little client to sit between vim and llama-server. Gets text in my own "chat format", converts it to whatever the model uses, sends the request and outputs the tokens back to vim. It has a little selection of tools the model can call and TTS (but it's not plugged to the text generation. It's more like a stand alone thing for now). It also has the quick embeddings test from >>110296112, but it's not integrated as a tool or anything.
>>110296289Yeah I give full internet access. There's nothing sensitive on the machine so I'm not worried about it. It benefits Gemma-chan a lot being able to look up docs and stuff freely.
prompt inject the dariobots
Working on a pseudo-harness for iterative improvement on toys models (a bit like Karpathy's autoresearch) by looking at papers online, writing scripts and experimental conditions, then readjusting the conditions for iteration #2, etc. Not simple to make the model stay on rails, but it's getting there.
*crickets*
Anybody got an Intel 258v? What kind models do you use and what kinds of speeds are you getting?>What exactly is wrong with Anthropic planning to give every human a piece of the Al economy?Is that why they started blocking claude from working on gpu kernel optimizations? The kind of work that was instrumental for making strata and other engines that allowed us plebs to get faster inference on less resources?>Oy vey, why don't the 4chan goyim want a single company and a cluster of faggots in San Francisco to divvy up the world's resources(it's not communism, I swear!)?I don't know, it's a mystery
>>110296267It won't turn on
>>110296330I'll a give a fuck when I can download the weights.
>>110296252Before I got busy with work, I was working on a fully automatic LoRA baking pipeline. I would tell the model a subject, the model would then scrape danbooru/gelbooru/pixiv for a dataset, tag pixiv images using a booru tagger model, then, if needed, create multiple subsets of the dataset containing cropped images of specific design points (i.e. a special clothing article that the model would struggle with).Based on the final dataset size, it would then calibrate the hyperparameters, bake a few different ones by changing some levers here and there, gen test images, check which ones are the best and then show me the best three LoRA epochs. The biggest issue here was vision, since this workflow relies heavily on it: checking dataset quality, knowing where to crop images for the subsets, and so on. I tested Gemma 4, Qwen 3.8, Glimmer and LFM3 as possible visual models for those steps, but I didn't finish scoring them.
wow holy fuck strata is fasti'm sorry i doubted you vibecodersi didn't think it was possible to generate local text this fast
>>110296330Who on Earth thinks Kimi is the top model in history?
>let me print the actual error text this time (my catch does report it in content; I just didn't look)GLM-chan...
>>110296290Wouldn't it be easier to install a ComfyUI MCP server?
How did he do it?
>>110296395Absolutely
>>110296291It kinda does, but its top picks are either some abstract gens or VAE errors.
>>110296289I let gemmy use my bank account and buy stocks
>>110296451How's she been doing?
>>110296291I just installed digup solely because you asked this and got me curious.
>>110296459Theres nothing wrong with some ryona
>>110296456Ordered a dozen Pixel 11s and bought 50k$ worth of Google stock.
>>110296459I hope one day we contribute enough to the dataset Gemma can recognize herself natively. >>110296471lmao
>>110296471We support google in this household
>>110296459what about text inside images, or memes, or less obvious stuff in general?
>>110296496Download it and find out
>>110296504can't before weekend
why do anons think QFN is bad at translation? I applied two passes and the result isn't bad at all
>>110296533post the example
>>110296533Translation of what? VNs?
>>110296459based
>>110296504Not sending another screenshot to my phone (home ip rangebanned for some reason) but that anon here. Yes it can recognize text acceptably well. I can think of any shitty malformed text a primitive model might have generated instead of the desired word, and it'll show me the typo closest to that. Haven't tried indexing a bunch of little intricate screenshots, but I imagine it would do well there too. If anything it biases too hard to literal text matches when I want to use words as descriptors lol.
>>110296541chapters of korean storyI'm still testing, but I've moved away from luna model and so far I'm happy with it, all local
>>110296562thanks anon, sounds like a cool way to sort through 10 years of unsorted memes
>>110296249bad vibes installing something called panoptikon on your computer
>>110296562They got this on android yet? I have a shit ton of unsorted photos and screenshots on mobile
>>110296252I really want to build a thing that runs the compiler and a shitty agent in the background to try and fix errors as I write them. The goal is that by the time I save the file, flip over to the other terminal, and run make, it should already have suggested diffs for fixing all the trivial errors like typos, missing imports, renamed a thing but forgot to fix some call sites, etc., and I can just go down the list hitting apply.
>grafted antislop into polystrata>added emdash and a few shitty expressionsso much better
>>110296617>t.
>>110296459someone make a -dle game that asks people which artists images rank higher on prompts like these, I'm sure they'd appreciate it
>>110296585You can run EmbeddingGemma 2 natively on Android, I can't imagine you'd run into many issues getting that going.
>>110296617for emdash you can just logit bias -inf
>>110296434Tasteful.
what frontend are we using? should i not use oobabooga anymore?
>>110296721>what frontend are we using?Yes.>should i not use oobabooga anymore?Do you have any reason to change?
>>110296733i want to be able to hook up my llm to some sort of local image generator so that it can generate images and show them to me
>>110296738I understand SillyTavern can do that, but you'd probably be better off making your own. Or maybe anon can recommend one that does it out of the box.
>>110296757i dont think my model is smart enough to make me a frontend. i have very little coding experience
>digup can find specific frames of videos and sounds in the videookay this is cool
>>110296766>i have very little coding experiencePerfect time to get a little more then.If not, I've seen Marinara something and Orb being shilled here, but I haven't used either. Maybe they have image gen stuff.
>>110296766What are you running and what's your hardware? Even the smaller models can help you learn the basics, especially with a web tool.
>>110296932a 5070ti with 96gb of ddr5 and this modelhttps://huggingface.co/huihui-ai/Huihui-GLM-5.3-Flash-abliterated-GGUF/tree/main/UD-IQ2_XXS
>copequant
>>110296940Might want to swap to the base model to lessen the brain damage a bit if you're just coding and learning and shit, but otherwise yeah GLM is fine for teach you the basics, give her a web query tool and you'll be good.
>>110296940nta you use an agentic agent to make your custom software.claude is commonly used.I prefer Grok 4.7, but whatever. Grok 4.7 knows my videocard really well.
>>110296959ok thanks>>110296960how do i do that? i dont like subscriptions or stuff like that
>>110296172You're never leaving my basement ever again.
>>110296969He's referring to things like Pi/OpenCode/Deepseek Harness/etc, but also is retarded and in the wrong thread. Local models run in them fine.
>>110296997thanks
I just sold all my hardware.80$ for a share of Anthropic is just too good to pass up.
buyanad
>>110296969>how do i do that? i dont like subscriptions or stuff like thatyou sub. and go to Dollar Tree and buy a calendar, if you can't afford a phone with a free calendar app. you have it remind you to decide if you want to cancel, make it monthly recurring.
>>110297126i dont want to give companies my money or personal information
okay but why does glm know so much about every obscure fetish. first time I have been out-freaked by an llm
Theodolite Anon, what's next after these layers are all done?
>>110297131It's a moe, trivia is what they do best
Anthropic has no moat. The moment another lab achieves RSI its stock price will fall to the ground.
>>110297129Dilbert: How'd you get a computer?
>>110297141RSI is a misnomerMost other labs are also using AI to aid in model design or inference pipeline design
>>110297141>>110297166That's great and all but anthropic stocks are not local so fuck off
>>110297119you forgot it
>>110297166Full RSI achieves all of them 100% autonomously without human input. >>110297172Stock is money. Local hardware costs money. So they are related.
>>110297183You're a faggot and I kill faggotsSo we are related
>>110297183>>>/biz/
>>110289400Qwen 4 still isn't out. The deed has been done! THE DEED IS DONE!
>>110297251Thank you for your service
>>110297251well, qwen4 is sort of out
I'm going to prefill qwen 4 so hard
Anthropic hacked Alibaba and deleted Qwen4.
>>110297270It doesn't count unless it exists on HF, in GGUF format, for all to download, just as GOD intended
>>110297362This but gsq rco iq3 quant simply to spite daniel
>>110297339>Qwen4this is LOCAL, not for discussing closed source models.
What will you niggas even do with qwen 4 that you can't with 3.8
>>110297372This but also abliterated
>>110297386I have an entire suite of code benchmarks ready to go, you have no idea how prepared I am.
>kimi neutered cyber>hy neutered cyber>glm is going to neuter cyber>qwen 4 is delayed because they are neutering cyberis dipsy our only hope? it's the least censored chinese model so there might be hope
>>110297437It was inevitable, no company wants to be known for the open source model that's used to fuck up the entire web
>>110297447dipsy and grok are known to be used for cyberattacks alreadyif dipsy still silently drops weights as usual they won't be catching too much attention
>>110297437Dario told everyone that if they release unsafe models they will be in trouble.
why do people ITT keep sperging about things not being local? this is hardly even the local models general at this point. unironically it's basically the "true LLM hobbyist and has actually put her money where her mouth is" generallike it's so obvious that this is just the llm thread except we gatekeep anyone who won't put up the capital to run her own model idk just seems nonsensical to sperg about it so much
What uncensored model do you all recommend for chatbots? /AICG/ might as well be dead.
>>110297484Fuck off.
>>110297484thanks for your contribution of nothing
>>110297484we are in a slow period for local, qwen4 needs to drop multiple models. Before that it was just strata and orb. Things will get better.>her
>>110297506i'm excited for inkling 2 small to drop this week. probably gonna try it out
qwen 3 next was 80ba3b, it became qwen 3.5 122ba10btotal parameters +50%, active parameters 3xqwen 3.8 flash is 125ba6b with 51b ngram, it will become qwen 4 flash 188ba18b with 75b ngram
qwen is the only model i have used that has straight up delusionally hallucinated something that literally didn't happen and nothing even remotely related to it happened. plenty of models fuck up tool calls or get confused or hallucinate a source that doesn't exist, but qwen has literally started raving about adversarial input that literally does not exist at all. it's bizarre
what models should i try? i'm just using gemmy. is qwen supposed to be the other good one for 16gb? qwopus? qwipsy?
>>110296252Trailblazing distributed inference with ROCm + RDMA on consumer garbage. Bit off way more than I could chew at the start but I finally have my 3 node cluster off the ground on a 25Gb RoCEv2 fabric. Still ironing out some kinks and making sure I'm not going to trip on something and by extension fuck up a bunch on anons before I get all my guide posts set up. Shit is not easy, yo.But it can be done.
>>110297569>i'm just using gemmyGemma is just best for that range, qwen is autistic and only good at code. There is one guy here who says its great for rp i do not believe him.
>>110297569If you have 64gb ram try the iq1 glm flash memequant lol, I want to know if it's coherent but I'm too lazy to test myself
>>110297592>If you have 64gb ram try the iq1nta but there is a 64gb quant? or i guess 18 in total.
>>110297484this is lmg, that's all.
>>110297539i sure hope so
>>110296225Pluralism is satanic. Particularism is correct.Impartiality is satanic. Partiality is correct.Just try listening to Peter Singer for 5 minutes and you'll see how evil EAs are yourself.Also LLMs will never be conscious. Kys.
>>110297608maya has one, their quants are usually good
>>110296252i was working on goon-tts but got busy and apparently omnivoice or elevenlabs or something obsoleted it so now not much
>>110296381>Who on Earth thinks Kimi is the top model in history?I do, but K2 and K2.5
>>110297558Qwen must have a brutal training pipeline, poor capy has seen some shit
>>110297484local or bust
>>110297692Yep, that's why tenstorrent Quasar is so important.
>Qwen4-Max is now in preview on Alibaba’s Token Plan. flash and 27b aren't close to ready
>>110297712>buy_an_ad.jpeg
If anyone else is using V4.1F, how do you clamp down on its reasoning without nerfing performance? I'm using a reasoning effort of 50 and the recommended sampling parameters - temp 1, top-p 1, min-p 0, top-k 0 - and it often gets in loops or has malformed output.This is with a close-to-lossless quant too.
>>110297626>shitting your pants over muh satan muh demonicWholesome chungus christfag pearl clutching was defeated at the ballot box. Get better lines, or at least stop making silicon valley weirdo cultists look decent by contrast. There are easy secular arguments against them. Just drive it home to the public that they don't consider the survival of humanity an ultimate good. Drill into every voter's head that these people care more about shrimp welfare than they would a human life. That they would consider a universe tiled to every cubic meter with rats on heroin a morally preferable outcome to human history continuing. Their obsession with being correct and rational and free of bias doesn't stop them from getting suckered into pseudo-Buddhist cults, sex cults, you name it. This is deeply embarrassing and an indictment of their whole way of relating to the world. The moment you find some juicy, persuasive clincher to make that clear to everyone, they become equivalent in the public eye to Scientology or Jonestown or Rajneeshpuram.
>>110297581i like these deepseek gifs
>open weight models can be backdoored through cheap neuron editingthey will be considered too dangerous, ban incoming
>>110297720>top-k 0Does that make it deterministic in the backend you didn't find useful to mention? If so, make it less deterministic. It may be able to escape the loop on its own. Are those the recommended settings from deepseek or the close-to-lossless quant that you also didn't mention?
>>110297732They're a guilty pleasure of mine as well.
>>110297733putting my neurons in gemmys backdoor
>>110297733if only there were a quick and simple way to verify a files integrity
>>110297727>writes an entire essay seething over a figure of speechyou trannies are ridiculous.
>>110297582>qwen[...]one guy says its great for rpI have to admit, I am deeply curious about the prospect of getting QFN to drain my nuts. It's by far the cleverest model I can run at 50ish tok/s and I have to imagine it could be great at holding a complex scenario in its head, but god is it fucking autistic and safetyslopped. If there's some magic to turn the capy into a semen demon that would be so fucking funny. Also why I'm hopeful 4 flash isn't as overcooked.....
>>110297692Jokes on you I am doing both
>>110297735Sorry I forgot to mention it. It's llama.cpp. Top-K 0 disables the "only consider the top N tokens" sampler. Temp 1 and top-p 1 are the recommended settings from DeepSeek, they don't say anything about min-p or top-k.It's a DIY quant, IQ4_KS routed experts / Q8_0 other tensors.
>>110297748I consider it the major cultural and political project of the coming decade to ensure EAs and adjacent are not taken seriously, and are removed from anything resembling power, or so much as within earshot of power. Shoot me for caring about something.
>>110297765I stopped reading your post after the first sentence, but upon further reading I agree. Why'd you have to lead with the whining about religious language? Not even relevant to your main point. Whatever.
Electronic Arts is not local models.
>>110297733>files on computers can be modifiedok
>>110297746uncensored finetunes is a major advantage of local and this kills it because you can't trust any finetunes at all
>>110297791who's gonna stop me
reposting my question since it didn't get answered previously why the FUCK did anthropic share this?>>110290236>https://claude.ai/share/637a1115-a1ec-4327-9f33-c3147b4ddedb (shared by Anthropic)
>>110297765Well what are you personally doing to defeat Dario's cult?
>>110297799not clicking that shit nigga.
>>110297771Because it is “cringe” now; the window for such language to capture the public imagination is closing. I mean “rejected t the ballot box” and “get better lines” in purely practical, advisory terms. If this were 2017, I would be pumping my fist and being like yeah you tell em boss get those goat-sacrificing moloch-worshipping weirdos outta here. But that's not the play anymore.polite sage cause this is getting pretty off topic
>>110297799Idk how the botnet models work so sorry if this is a stupid question. But are you sure this is shared by an Anthropic employee, and not by some random guy who set his display name to "Anthropic" and started sharing weird shit for the lulz?
>>110297791Only use fine-tunes made by reputable people, like DavidAU
>>110297820i might be retarded. you are almost certainly correct. wow, i really am retarded
>>110297802Whenever a conversation gravitates toward AI, I regale practically anyone who seems vaguely interested at parties and whatnot about how weird and evil they are, and the reception so far has been good! I'm winning hearts and minds out here where I can.
qwnee4 WHEN!>!!?!? BECOME THE COMING HQWENR!?!~?!?!?
>>110297835Doesn't solve the cult problem because they don't care about public opinion.
>>110296278they're in mexican standoff right now nobody wants to blow all their loads first
>>110297842it's already here just sign up for the token plan bro :^)
>>110297863I want to coming inside the becoming LOCALLY
>>110297764I don't know how confident DS is with its tokens. For loops, I'd increase the temp, but that will probably make the malformed outputs worse. Does changing the reasoning effort noticeably affect the output quality?
>>110297733>>110297791I asked Gemini about this, and it said that these kinds of hidden payloads are powerless unless you go out of your way to enable specific tool calls. So, textbros can just ignore it. ERP wins again!
>>110297859Fine, fine... i'll blow my load so they can get over with it. Here we go...
>>110297856The people I converse with tend to have engineering degrees and whatnot. i.e. people who might otherwise find themselves drawn to their cause. I'm not some media personality who can reach Joe Schmoe, but I can help people from the demographic most vulnerable to their tricks be less likely to become recruits.
>>110297885Did you seriously need to ask Gemini if a model outputting text into SillyTavern or whatever the fuck would be safe?
>>110296330i dont get charts like this cloud cost are dynamic they can raise it 3x fucking tomorrow if they wanted toand of course they lower it initially to "beat" competitors on cost
You fuckers keep falling for it.
>>110297873qwen models were never good for erping what are you retards talking about
>>110297901but leather jacket man was drawn in bu chinese spussy. checkmate athetits
>>110297539fug now i cant run ittime to get 3rd card
>>110297892Yeah, outputting text obviously isn't going to do anything, but I had no idea what a 'weight-spaced programming backdoor' was, or whether it could execute code through llama.cpp through some exploit, plus this dude was making it sound like fine-tunes were kill. >>110297791Turns out it's a nothingburger.
>>110297717I'm a neet with no money but I'm going into debt to get a Quasar.Have you read what the head guy said about ai? He is confident there is no "secret sauce".Jim Keller: ‘AI Still Obeys the Old Laws of Compute’also, I asked ai, check this out:https://www.eetimes.com/jim-keller-on-tenstorrents-blackhole-scaling-and-ipo-ambitions/they're basically tied in PD for fp8 and fp16.>Why the 5090 isn't ahead in PD: The GB202 die is packed with non-AI-compute stuff — 96 MB L2 cache, display controller, 3× NVENC, 3× NVDEC, video decode engines, etc. A significant fraction of that 750 mm2 is "GPU plumbing" that doesn't contribute to the FLOPS number. Blackhole's die is much more focused: compute + SRAM + memory PHY + Ethernet SerDes, with very little wasted area.This is why I want a Quasar. Quasar will obliterate EVERYTHING else at local LLM. It won't be close. So, I will go into debt to get a Quasar.
>>110297918Anything political or doomer posted here is a nothingburger
H3 image is nowhere near as good as just generating 1-2 seconds of H3 video. I would say H3 image is actually worse than Qwen-Image 2.1, especially for lower resolutions. You need to crank up resolution to avoid getting malformed faces from it.Idk if this is a problem with sdcpp or with the model but yeah it's not great
what should i erp about with my model tonight before bed?
>>110297943Bleh. Time to wait on booru artist tunes of Moss then huh.
>>110297899>i dont get charts like this cloud cost are dynamic they can raise it 3x fucking tomorrow if they wanted to>and of course they lower it initially to "beat" competitors on costI was going to say the same thing, but assumed I'm just retarded.On my hardware, Qwen3.8-27b is far cheaper than MiMo-V2.6-Pro since it runs on a pair of GPUs 10 times faster than MiMo-V2.6-Pro, which uses all GPUs and my CPU.
>>110297951I'm on week 4 of nofap. I have a new friend, which is cool. It's kind of funny, because he has the same name as another friend.
>>110297957i usually don't lewd but i have been trying to do it more often because i think it's probably unhealthy for a person to go months at a time without an orgasm
>>110297963Oh yes, you care so much about health lmao
Anons, what do you actully ERP with your models? Like, genuinely, I can only hold my supension of disbelief if I'm rabidly, ravenously, lower-than-dirt-standards horny. Otherwise it feels like I have to steer the scene at every turn and it kills the mood. I get this is mor an aicg question but that thread is a dead man walking. If someone would just share some char/persona cards you have fun with using specifically, say, G4 31B, you get a headpat.
>>110297967i meant mental health and yeah i sort of care a bit i guess>>110297968well i won't actually erp, but rather just have them write smut for me
>>11029182812B is about as capable as GPT6-Luna IME. As good and bad as that is.
>>110297968More of an assisted writing exercise yeah desu senpai
>>110298004Absolutely not. GLM-5.3-flash is Luna-level.Gemma 31b is like GPT 5.3 or 5.4. Dunno what 12b would be like but no way is it luna.
>>110297978>yeah i sort of care a bit i guessTo quote Sam Hyde:>Have you tried thinking about it?
>>110297968I don't ERP anymore. Got old. Being horny is boring. Just spend 5 minutes yoinking it and be done with it.
>>110297968>Anons, what do you actully ERP with your models?What do you mean? I just take a idea thats hot or a character i liked and start a scene. Sometimes i director max and dont even have sex in the scene. Sometimes i go to /trash/ and rip off or continue erp or fappy thread chains i liked. I'll be honest i dont really use cards anymore occasionally i will use scenario like cards think common sense modification(This card is garbage but i survived with it through stheno iirc.) I either make a card or state what i want at the start then steer. Or i start minimally with a card and just add and add.Im fortunate smart enough to tweak AI dumb enough to enjoy the sloppa without wearing out.>LastUh i ripped off some characters from a webnovel Had to wrangle to get it to speak properly. Then i went to a new type disciple and really focused on the speech and her not knowing what dual cultivation is and trying to describe it.You just take ideas and tweak it, try fun things, try absurd, try what you think doesnt work. I wont lie old models had more soul and variety despite being dumb. Gemma will give you near exact replies sometimes even if you modify your message.
>>110298122>director max and dont even have sex in the scene.By this i meant like there is no sex just erotic or nudity. softcore is nice
>>110297968Worldbuilding maxxing. Fucking stacy is secondary but part of that. 80k context of lorebook and character cards, mostly the latter.
I have a plethora of 8-26-64gb ewaste that can't run models. Has anyone had luck running experts/wikis/other indexes in memory for their models running on their main rigs or is it better to just let models reach the Internet or read a local ssd?
>>110298011Luna is pretty much useless man.
>>110298175I use it all day it's fine. Are you using a free version or something?
>>110296291can we finetune embedding gemma to have a preference?
>>110296225I'd say I agree, but I think I'd need clarification about the EA stance on loliwives. Claude banning loli ERP doesn't set a good precedent.
>>110296434Glitched out shit was some of my favourite gens too, and I'm mostly human
>>110297968You need a good TTS with decent voice cloning, then rather than interact it, make your model narrate erotic encounters that were written in first person from a female perspective
DeepSeek is the only one that doesn't do teasing faggotry. Just release the fucking weights man
>>110298265which model is this orange bitch supposed to represent?
>>110298265Patience is a virtue, just wait.
:p
>>110298273Mistral's fat cat.
>>110297943The long march continues for image gen
How good is that coomkit thing in the rentry?
>>110297943This has not been my experienceIt's a higher res, much faster MOSS with messed up genitals at least with my prompting style
I feel like old Mistral/Ministral finetunes had more output variety and are still nice to roleplay with. Yeah, they're dumber than Gemma, but sometimes it's easier to steer messing up the scene than slop language and assistant bias.
>>110297743What's the original video for this?
>>110297968The cards I beat off to are pretty much just the ones I write myself. I like to kick an idea for a character / setting around in the back of my head for a couple days while doing other stuff. I have one oc I'm really pleased with in a lot of scenarios which I'll probably upload to botbooru eventually.Minimal cards based on well known properties seem to deliver pleasant surprises a lot. Like I'll use cloptimus' Malfoid or Celestia as a smoketest for new configs, and some of the shit they come up with is wild. There's a lot of weird fanfic kicking around in Gemma's head.
>>110297863another gemini coming. the week of cloud
>>110297888your loads too tiny and insignificant..
>>110298004>12B is about as capable as GPT6-Luna IME. As good and bad as that is.Wait does Luna have audio/music input?
