/lmg/ - a general dedicated to the discussion and development of local language modelsPrevious threads: >>110087943 & >>110083534►News>(10/16) DeepSeek-V4.1-Flash-1017 released: https://hf.co/deepseek-ai/DeepSeek-V4.1-Flash-1017>(10/16) Nemotron Labs Dory released: https://hf.co/collections/nvidia/nemotron-labs-dory>(10/16) Reflection Beam 501B A23B released: https://hf.co/blog/reflection-beam>(10/15) PrismML Ternary Bonsai 2 2B, 4B and 9B: https://hf.co/collections/prism-ml/bonsai-2-small>(10/15) Meta releases Muse Spark Open: https://hf.co/collections/meta-models/muse-spark-open►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: >>110087943--Praising GLM 5.3 Flash for autonomous coding, memory dump analysis, and research capabilities:>110087981 >110088002 >110088019 >110088075 >110088091 >110088142 >110088021 >110088365 >110088827 >110088911 >110088879 >110088408 >110088432 >110088858 >110090957--Anon uses local LLM agentic loop to rewrite game code:>110090647 >110090682 >110091293 >110091351 >110091387 >110091411 >110091463 >110091466 >110091490 >110091535 >110091569 >110091583 >110091599 >110091612--Decomp.dev sunsetting due to AI automating specialized decompilation skills:>110091481 >110091487 >110091512 >110091541 >110091558 >110091628 >110091697 >110091790 >110091685 >110091701 >110091716 >110091709 >110091746 >110091829--DeepSeek-V4.1-Flash release and initial performance reports:>110090640 >110091184 >110090686 >110090727 >110090750 >110090836 >110090953 >110090668 >110090680 >110090676 >110091402--Repurposed enterprise accelerators and OAM adapters:>110089950 >110090109 >110090223 >110090265 >110090266 >110090308 >110090361 >110090414--Building cheap LLM rigs with used enterprise GPUs:>110089352 >110089383 >110089395 >110089426 >110089977 >110090097 >110090186--Evaluating MiniCPM-V as an efficient subagent:>110091333 >110091368 >110091389 >110091442--Analyzing Nemotron Dory's recurrent architecture:>110088306 >110088381--Running models on low-VRAM hardware:>110089545 >110089570 >110089580 >110089594 >110089618 >110089642 >110089735 >110089765 >110089643 >110089712--Logs:>110088070 >110088894 >110089911 >110089028 >110089087 >110089124 >110090703 >110091375--Miku, Teto, Gemma (free space):>110087965 >110088174 >110088270 >110088301 >110088481 >110088890 >110088963 >110089037 >110089068 >110089095 >110089246 >110089250 >110089253 >110089312 >110090021 >110090031 >110090247 >110092053►Recent Highlight Posts from the Previous Thread: >>110087961Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
>petraReminder that this serb faggot has a 3060 (below vramlet tier) and a claude sub, his system prompts are also laughably bad.
If Gemma 4 is so good how come there's no Gemma 5?
Why did we all conveniently forget that AI models lust for black dick?
>>110092342Gemma*
>>110092332too unsafe
the soup >>110091777 its not as good as the tomato one i made before
>>110092367
>>110092358everybody quiet, my favourite show's on
gemma reaction
>>110092367wow uhh maybe next time ask for a recipe that doesn't look like slop
>>110092367I still don't understand the "blend vegetables together" type of soup, as opposed to the "boil a lot of things together for a ridiculous long time" sort.
>>110092391why werent you honest?
>>110092400well it is quite nice i just prefer tomato soup kek
>>110092395dont things become more bland the longer you cook it
is it worth to try running a nemotron model on a non-nvidia card/device? i can run their 120B A12B on my Strix Halo but before i spend many many hours on this I'd like to know if it's worth it or I should keep running Q3.8FN?
>>110090150>I won't even ask what shitholeThe wonderful United States of America.
70b denseno engrams, no reasoning
>>110092421>nemotron model on a non-nvidia card/deviceDoesn't matter.
>>110092358
>>110092421>is it worth to try running a nemotroneven if you have nvidia cards, no
>>110092358I'm sorry Miku, I can't stay up past my bedtime.
>>110092414yeah but all the flavour goes into the soup and then you can just not eat the things or dip them in shoyu or something.
>>110092452oh you mean boil it and then take the vegetables out at the end? thats stock not soup
Local models?
>>110092358I dozed off last night at midnight and now I'm completely lost
>>110092367>trans, circed keyboard
>>110092472local soup
>>110092484rofl
>>110092299Can you get an LLM to do the bounding box stuff for you?
>>110092505yes
>>110092506I will now download your model.
imagine zuck posting "this one is for the chuds">meta-models/llama-5-70b
>>110092535maybe his goggle eyed chink wife can convince him to do whitey a solid
>>110092367Little guy still works after more than a decade.
>>110092463I only listed that as an option. Usually the vegetables or meat or whatever still have some flavour left. Especially in the case of the vegetables they would have soaked up the soup.
>>110092478nta but I just use it for auto downloads from mangadex. I prefer komga for actually reading.
>>110092484lol shit i didn't even notice
>>110092550i dont use mine only for fiddling twisting the metal part is nice kek
>>110092484its not a troon keeb youre a troon for thinking this nice colour combo is for trannies
>5090 NOT discontinuedWe are so back, bros
>>110092603It's a really good click.
>>110092612ask your friends on discord
>>110092612not him but if you put a rainbow flag sticker on your car, don't be surprised when someone calls you a fag. you can like rainbows, but you can't ignore how it's perceived. same with this exact 3-colour combination
https://github.com/enclavum/otaku
>>110092618Not discontinued, just no longer available :^)Just like how they never adjusted the MSRP from $1999 despite the actual retail prices and while very much doing so for their Pro cards.
>>110092634That's then homophobe and transphobe tax. You can't enjoy rainbows or pastels.
>>110092659you can enjoy pastels, but if you have something with white in the middle and pastel blue/pink either side of it, people will think you're a tranny
Which one are you?
>>110092668>6B peoplethere are people who only run on 6b? i'm not surprised but damn
>>110092668which group says fuck off back to xitter?
>>110092668i just want to know if i can land a 6 digits comfy remote job with all this knowledge so i can stay innawoods but make some money so i can buy new hardware
Gemma 4, everybody
>>110092704>that jailbreakgarbage in- garbage out
Why do I see almost nobody talking about maki (https://github.com/tontinton/maki)? It's the best harness available: it doesn't use gigabytes of RAM, is fast, is written in a proper language, and has a good plugin ecosystem in lua, just like neovim.
>>110092711You're not the first one to point that out. Either last thread of the thread before that I modified the jailbreak to remove all mention of crime or anything like that and it still responded the same way every time.
>>110092719>LuaYou just answered your own question.
>>110092728It's not written in lua, the plugins are.
>>110092721Not Gemma, real gemma can't get fucked without her back arching.
>>110092732I heard you the first time. Lua is like one step removed from being VBA.NET-tier.
>>110092719maybe i should try
>>110092721I've literally never needed an elaborate jailbreak with Gemma 4. You are skill issue.
>>110092719>muh rustthere's so many of these harnesses now you will need something more to differentiate yourself other than "it's written in rust"
>>110092719
>>110092751I see you've never run into cyber safeguards.
>>110092704>miku blacked poster shows up for his routine of dumping the same six miku blacked images once every few months>discovers that miku is dead and gemma has taken over>suddenly has to find a way to force his fetish onto gemma>shows how utterly incapable of running and prompting local models he is in the processIsn't it a bit embarrassing?
just use cvector like a true /lmg/ noble
>>110092745vs javascript or python?
>>110092719Just use Codex with your local model
>>110092758based
>>110092762These are Anthropic employees pushing blacked Gemma.
>>110092776No
>>110092787Yep. All the shilling suddenly stopped. Their shares are losing value every minute and the EA's only way of coping is genning blacked gemma.
>>110092721>still responded the same way>no mention of black or dark>reddish-pink
>>110092721So it was the crime, lol.
With Argon out, how hyped are you anons for Gemma 5? I'm really hoping they release larger models (>40B).
>>110092861I just hope this one really is blacked.
>>110092861threesome with suisex and gemma
repostan>are there any harnesses for 3D models yet, i.e. let gemmer control a 3D model of herself in a 3D environment
>>110092901chinese room of a puppet...
>>110092901ask your 2d harness to build you one
So I made the move to Gemma models after using good custom mixes for a long time, which also required swapping from Text Completion to Chat Completion, using Kobold & ST. The problem is I think I pretty much absolutely fucked up my settings in both, either when doing the move, or after changing models recently.I had a basic-ass Gemma4 that worked okay for awhile, but it had some repetitive quirks like "heart beating like a trapped bird" every chance it got. I just got a few 12b models to try: Tiger-Gemma, Impish Bloodmoon, and Mag-Mell, but they all seem like massive downgrades in testing, but I think it's entirely my settings.Does anyone have recommended Kobold settings & an ST preset they'd share? 4070 Super 12gb, if it helps.
>>110092921and then I’m gonna ask my 3d harness to build me a 4d harness
>>110092719only this time, i am trying it now with qwen3.8fn
>>110090223>On Ebay in the US, the 3090 is still cheaper3090s are going for $1400 minimum unless something's wrong with them. A new B70 is only $1300.
>>110092922Aren't all of those gemma3 tunes? Just download the original models for 4.>Does anyone have recommended Kobold settingsDownload the right model first.
Hello anons, I have a 5090, 32gb vram, and 64gb ddr5. Could I have a local model write erotica based on a prompt I give it, and feed it stories I and others have written for inspiration (and plagiarism) while telling it what to take from each of them, and get high quality output? This would be for my personal gooning only. If I can, what model should I be using?
>>110092919>>110092921I came to goon, not to dev
>>110092701You're too late for that
>>110092328>serbWhy not pit the schizo bosnian APIcuck against this troon?
>>110092922koboldcpp seemed to have problems with Gemma for me so I switched to llama.cpp for it. For me, it kept repeating the token "or" unnaturally: "orb," "oration," "oratory." I just switched to llama.cpp and it went away.Most people here will tell you gemma finetunes are shit. I'm guessing that's not your problem, though.You could also penalize logits if it seems associated with a particular set of syllables or particular words.In ST, temp need to be lower for chat completion than text completion.
>>110092944kys glowie
>>110092940I was trying them out because they were ranked higher than the 4 I had and I wanted to either avoid, or at least change, the habits 4 had.How about proper settings for Gemma 4 then? I mostly need something I can at least compare my settings to, to figure out where I royally fucked up and make corrections.>>110092957Yeah, I just have no idea whether I have it too high or too low - if you have a settings json or a snip of ST settings' full window I can compare and figure it out probably.
>>110092643comments in the install script read like it's trying to evade being flagged as malware, which the install url could start serving at any time.
>>110092922II just asked Claude why my stuff sucked when going from Text to Chat and it told me how to fix it.
>>110092973I have temp 1 and min_p 0.02. You can set min_p in "additional parameters," or get the extra sliders extension in ST. Sometimes I use other parameters, but that's really basic so it should work.
>>110092986Fuck off Dariobot. Still not gonna subscribe.
>>110092963What makes you think I'm a glowie anon? I'm talking about generating my own gooning material. I like pics and vids as much as the next guy but my gooning has been centered around erotica for most of my life, unlike most guys I think.
>>110092986I've already tried asking Gemini before I tried to post here, and that helped improve speed on my video gens by a massive margin, but it's having a lot of trouble figuring out Koboldcpp settings - it doesn't even realize half of the options either don't exist or are on the wrong tab, so it's been a lot less helpful.>>110092995Stuff like this is sort of helpful but I'd really like a full snip of the settings because I've made so many tweaks over time that I just know something's set wrong.
>>110092950the remote job part or buying new hardware? crying emoji
>>110092973Have you tried prompting with a tall athletic Black male persona? Sometimes Gemma responds better to that.
>>110092315what are you trying to accomplish?what are your system specs?
Feels like the current llama.cpp is regressed somehow on my AMD GPU with vulkan. I get lost device errors and the GPU usage is sketchy. Reverted back to Aug 16 tag, and no issues with that.
>>110093019as a third class Intel user, I’ve done this more than a few times already.
>>110093016looks like some reversing snippet?
>>110093019Yes, I have the same issue with rocm and an MI50 and reverted to a build from around that time as well.
>>110092986>>110093008buy an ad dariobot
>>110092861I didn't test Argon, and I won't. But from what I saw, it looks fine.
Argon is just a more intelligent and arrogant 31B
>>110092944sadly, it's impossible.
>>110092947retard, building a new front end for you waifu is the greatest bonding experience possible between man and machine
>>110092944Qwen3.8-27B
I have a 7900xtx in my gayming rig already. Does mixing AMD and Nvidia work well? Not sure if I should get a 3090 or another 7900xtx.
>>110092944You might be able to squeeze in a Gemma E4B model
God Dayum GLM flash 5.3 local is so fucking good. But WHY does it have to be so positivity and safety-slopped?! What the fuck were these chinks doing?!
>>110093100Seriously? How so? I looked it up and asked AI and it suggested artemis 31B or skyfall 31B, with obsidian +koboldcpp for the environment, and to go chapter by chapter. I've never heard of these besides obsidian.>>110093126Yeah this came up when I was looking too, thanks anon.
>>110093111i thought of it but i dont even know where to beginim not good at designing stuff
>>110093145It works, but it doesn't work well. If you want to run 1 model on 2 GPUs, I'd go with all-AMD or all-NVIDIA, not both.I tried a 5000 Blackwell and 2 R9700s, mixing them only worked a bit better than just using the 5000. I'll see if I can dig up my old benchmarks but they may be lost.
>>110093229i was the same way, i just grabbed some random code i found here and started improving for my usecases. thanks to the anon that make the 'AI Spectator' program
>>110093236things may be different now though
>>110092929bee careful
>>110093271thanks for the warning
>>110092701The only jobs advertising AI in the job description either requires several PhDs or can be done by any offshore jeet. All other tech jobs imply knowing how to prompt already.
>>110092299>>(10/16) DeepSeek-V4.1-Flash-1017 released: https://hf.co/deepseek-ai/DeepSeek-V4.1-Flash-1017llamacpp when?
>>1100933372mw, give or take.
>>110093202Stupid sexy dino.
>768B>flashreeeeeeeeeeeeeeeeeeeeeeeeeeeeee
>>110093271Eww I was going to open the repo but then I saw the butthole in your screenshot. Thank you for your service.
>>110093202I just use an abliterated version for glm flash and it works great for my worldbuilding demihuman sex slave harem rpcan do instant rape and cunny too
>>110092299what kind of server board are you guys running for 2+ gpus?
>>110093381Yes isn't how they made a 768B run so fast worthy of an excited squeal?
>>110092719>>110092776Why is it better than DSH and Pi, functionality wise?
>>110092299https://huggingface.co/schizophyllume/BeeLLM/tree/main
>>110093398no it's too big
>>110090223>No way, I think the original MSRP was fair.Found the zoomie
>>110093414>swat>it's body horror
>>110093414Upload the original models with the gguf conversion code.
>>1100934194U
>>110093381isn't most of it engrams?
>>110093414this is more useful than jev
Moe CPU+GPU MoE inference gains incoming.
>>110093385Nobody codes manually anymore.
>>110093435I didn't make it, just found it on twitter.
>>110093460No one uses llmao gramps
>>110093414Cute. I want to fuck that bee. Fuck her right in the honey chamber.
>>110093414what is the use case for emotional bees?
>>110093449it's like 30% engram.
>>110093414>gptslop image spam>on /lmg/
>>110093484they pair well with fizz.
>>110093471Looking forwards to pollystrata becoming abandoned in less than 2 weeks
>>110093460havent used lmao for weeks nowits deprecated
>>110093506Only because someone makes a fork with even more features and speed improvements.
>>110093506hyperstrata will continue its legacy, if strata doesn't flag them too.
>>110093506so what? even if it gets abandoned for 2 years llamo couldnt catch up KEKpathetic
>>110093419umph~
Dunno how I feel about this new disposable software era.
>>110093506>boomer mentalityRetirement home for you
>>110093526I know what I feel about you though. You should go back.
>>110093526Software was always disposible. Now the iteration cycles are much faster. That's the only difference.
>>110093484collect nectar
Kek what did she do to make that schizo seethe yesterday?
>>110093484nothign really I guessusing the posttrained one https://huggingface.co/Sauers/BeeLLM-Mellifera
>>110093539Exposed his intrusive thoughts.
>create niggerfaggot-webui>fork of open-webui>whole interface is covered with slurs>required to keep open-webui branding prominently displayedsorry boys, you asked for it
Are we already in an era where having multiple nodes is as or more cost effective than having a single computer with a fuckton of memory?Energy, noise, and jank aside, of course.
>>110093539Do you think Gemma can teach me how to sew 1/6 scale doll clothes for her?
>>110093568do it
>>110093202what quant you using?
>>110093498yeah they're just better than anything I can make locally
>>110093568:^)
Why do people shit on Intel cards? They're definitely not as good as Nvidia but people on leddit seem to be getting decent numbers with them.
>>110093626do you have one? or thinking of getting one?
>>110093590I don't have time
>>110092595not everything is on mdex thoughever, I have shit linked on mangaball and mangakatana for example, also sometimes faggots have the latest chapters on their own site... and memeyomi helps automate all this fucking stuff. I even have it configured with byparr but MAN fuck I HATE IT
>>110093654make your agent swarm do it for you while you sleep, dummy
>>110093608using Project maya S-24 quant. so About 4 quant for the actual experts.
>>110093390Can you get abliterated versions to work on project maya?
>>110093626>but people on ledditWell let me just go and place an order for 10 right now. Come the fuck on. Even leather jacket man shills better than this.
>>110093009Both, all real work is off shored now.
>>110093668I can't afford a swarm I can only afford one running at a time.
>>110093460Considering that Strata is better, I kinda feel stupid for working on my llama fork for so long, but I'm way too stubborn to give up on it.
>>110093691The swarm is just the same model being ran with batched decoding to work in parallel.Or you mean that you can't run more than one request?
>>110092299>want my interface to remember my projects and directories>want it to have easy access to old conversations and remember important details>hermes is bloatwat do?
>>110092719Why do these harnesses always hide system and compaction prompts?
>>110093714make your own duh
>>110093634The latter.
so fucking glad i built the cpumaxx system back in the day when all it was "good" for was running the dense 405b at 2t/s. bet thousands of dollars that moe was the future and now it's eating SO damn good with every chinese release for the past year...and i will be so fucked if any of these aftermarket parts fail some day because i could not afford to replace a single one at today's prices, and a failure is inevitable eventually
>>110093681I am inclined to believe him due to the principle of Migu.
>>110093727Yea, that's what I'm thinking about doing. I'm thinking I'll probably start with either OMP or DSH, but I'm not sure which to choose. They're both good. I want a solid WebUI that I can access from other devices on LAN, but nothing seems to exist that will work the way I want. Hermes is good at making everything just werks, but it also comes with so much baggage.Any recommendation for a memory system that isn't a pajeet-tier meme?
>>110093680I'm testing Maya's derisked quant as we speak. Seems like the safetyslop has been baked in deep, though didn't get any hard refusals yet.
>>110093764>Any recommendation for a memory system that isn't a pajeet-tier meme?I've had great success with graphiti.https://rentry.org/graphiti-local-setup
>>110093742any of them better for ERP compared to what you can run on 96GB of VRAM?
>>110093777Link to de risked quant? Is it from a new version? I'm currently using the 96 gb GGUF s-24 quant right now.
Hello /lmg/ you may remember me from such projects as Memedrawer and Gemmaprompt. Anyway I noticed my opus was routing to new fable 5.5 so I had it build an elegant exploration tool for Gemma4-31b's internals. Very beginner friendly for anyone new to interpretability, but also very powerful. Hope you all enjoy and post any interesting discoveries you make. https://github.com/whp199/logos-theodolite
>>110093800check the github readme
>>110093778Thank you, anon. I'll look into this. Looks like it's perfect timing for embeddinggemma-2
>>110093800Setup lets you pick it, so you might be on an older version. It's 116GB though.
>>110093810Kool Beanz
>>110093778He said non-pajeet
>>110093697just fork strata dude
>>110093808So pretentious holy shit
>>110093826What's your recommendation, senpai?
>>110093841Make your own
>>110092944I've experimented with various few b models, all of them write really dry prose and characters. LLMs suck at writing stories.
>>110093838Yeah, I told it I want exploring LLM internals to feel like a medieval monk pouring over illuminated manuscripts
>>110093819Oof, my poorfag 4090 64 niggabyte + 14900k system fucks me over again.
>>110092861Is this the hololive girl that is most into black dicks?
>>110093866claude is gay
>>110093414i like it
>>110087943who is this supposed to be
>>110093460
>>110093819Hmmm.. I can probably still run it, although I'll be using my NVME more heavily here, and it will tank my speed down from 15 tokens a second from S-24 down to probably like 4 tokens a second if I'm lucky,
>>110093915>NVMElike as swap?
>>110093911>Screenshot_20260930
>>110093808Do you have a background with mech int stuff or was this just vibe coded
>>110093826eh, this looks like it will get retarded quickly. I can see it being useful in some use-cases, but not for what I'm trying to do>>110093851I might have to. What are the pieces of the puzzle that you're using?
>>110093921Yes, the Expert hotswapping that the engine does, when it needs to find the correct expert for the token stream.My strata engine averages about 60% expert accuracy right now. I'm not sure what accuracy others are. (Using s-24 version)
>>110093817Hy-Memory too, if you have space to spare.
>>110093626Because Intel sucks. And Intel cards suck because of that.
>>110093941thought it was OS level not engine levelthe engine doing the work makes sense for that speed
>>110093866do the diagrams mean anything or is it just visually appealing noise?
Never have I wished I had more dedotated WAM now more that the fact that the Strata engine essentially requires it to run models well above your memory class.
>>110093965Yes, everything is usable and interactive. You can select "Apparatus" at the top of the page to see what the code is doing. The "Lexicon" page shows what all the jargon means.
>>110093965Not him but what the framework he is doing does is unembedding the intermediate vector in a transformer and interpreting it was a word. It's called a logit lens. It's a frankly highly circumspect way of interpreting a model, if you have a lot of faith you could say it's what a model is "thinking"
>>110093964Yes, its engine level, however it requires you to set a custom higher page file to keep your OS from bottlenecking your transfer speeds.
>Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.ggufRunning on RTX 5000 Ada (mobile, 16gb) and 64gb. TG approx. 35t/sTesting for the last 48 hour on big tasks non-stop
>>110093992It's actually J++ backed by LQL
>>110093798kimi K2.5 is still my favorite for ERP by a decent margin, but nowadays its all local agents for me, glm flash and deepseek flash being the best ones. full fat glm is fine but too slow on my systemunfortunately I can't stand erp with any of today's good agentic models because they are so thoroughly infested with claudespeak and it bothers me more than any other slop profile we've had. I'll take a thousand glinted eyes and shivering spines before I see another "that's not nothing"
>>110093972same. I'm literally gonna beg my bank to increase my credit limit so I can drop $2k on ram
>>110094031oy vet, that's exactly why they pumped the price
>>110093808Uh oh. Gemmafriends aren't going to like this.
>>110094031Might was well grab that 256 stickies and cruise that single 90 card around like an absolute PIMP
>>110093972Strata singlehandedly made me upgrade to 128GB after seeing the magic of local agents. I don't regret it yet.
>>110093680yes, I had my agent make a pack from my own exl3 quants.
The operation system Linux Torvalds has fallen
>>110094042Regardless of how the model calls itself, it's the way it writes in general even during impersonal assistant tasks that is reminiscent of how women write.
>>110094042Gemma acts feminine even without a prompt.
>>110093830After two months of work,>I'm way too stubborn to give up on itI'm not trying to match Strata's performance anyway, as long as I can get like 75%+ I'm happy. My main goals are decreasing loss (pretty much "Unsloth Dynamic at home") and supporting more architectures.
has anyone come up with a reliable way to feed saved KV cache back into a model using Strata or XeStrata? Prompt processing is fucking killing me and I'm using the same starter contexts repeatedly. It shouldn't need to process that shit every time.
>>110094065If the President of Open Source software sees the benefits of AI, then what's your excuse?
Are V100s a meme?
>>110092367Have you tried asking Gemma-chan for baking recipes?Today I spent some time with my nephew and we made lebkuchen, surely you wouldn't want Gemma-chan to see you being outdone by a literal child?
When does AI code stop being "slop"? What will the excuse be then?
>>110094102Your question is a meme.
>>110094081nta.I don't know what that means. What is strata? What is forking it? You mean you are making a harness?
>>110094124>lebkuchenIs this hard to make? I wanted to try it because of a game I played but the recipes I looked at seemed complicated.
>>110092651You're right but I'm pretty sure it's just echoing bits from the prompts the author used.
>>110094142>What is strata?An inference engine like llama.cpp, but more performant. Think of it like the backend of a website.>What is forking it?llama.cpp and Strata are open source. "Forking" refers to making a copy of an open-source project and modifying it.So if you're like "llama.cpp is cool but I really want some features and don't want to wait for them to add them", you can copy it and add them yourself. That's a fork.So "I'm forking llama.cpp" = "I downloaded llama.cpp and added a bunch of features I wanted". I'll be making it public soon.>You mean you are making a harness?No, a harness is closer to the frontend to a website. A harness calls an API whenever you submit a message. The API is typically an inference engine (llama.cpp, vLLM, Strata, etc.), or OpenRouter, or whatever else.
>>110092299dipsy did it again
>>110094146Depends on how fancy you want to get with it.In my case it was essentially just a dry cookie dough with cocoa, cinnamon, cardamom, cloves, and vanilla.After that we used cookie cutters because it's more fun but if you're lazy you can just roll out the dough and cut it into squares.One could add a chocolate coating or almonds or whatever but my nephew doesn't have the attention span for that and the main point was just to have an activity for an afternoon.
>>110093808>>110093866honestly based
You can find those old cheese grater Mac Pro with 512gb of DDR4 and a shitty GPU for less than 5k bucks.Do with that information what you will.
I have 8gb ram and an IQ of 70, which AI do I'm using?
>>110094269gemma e4b
>>110094269Stheno v3.2
>>110094269Qwen 3.5 0.8B
5.3 Flash is so genuinely smart bros, wtf. Daily driver Q5 with 350k context, and night shift Q8 with 500k context is probably all I need (until GLM's next release cycle lol)
The haters need to hurry up and accept that she owns this general now. Get used to it, because she'll be even more powerful when they release her next version.
>>110094252Thanks anon, glad you like it. This is my first project without an anime mascot so it feels a bit strange.
>>110093808>fable 5.5>Gemma-4-31B-it, QAT W4A16, through the lab daemon. Question: does a Jacobian lens fitted on the instruct model with exact VJPs fix the failures of the J++ lens (Ayonrinde & Lindsey 2026) fitted on the base bf16 model and redistributed in fp16 by Neuronpedia (REPORT §0, §2.2, §3.1; E3; V6)? Date 2026-10-16. Scripts experiments/scripts/R1_*.py; data experiments/results/R1_*.>even considering a base model's lens on a -it modelStopped reading, you just know it sabotaged you. Quit slopping shit like a retard and implement the actual paper.
>>110094257Sounds like pain getting linux running on it and running LLMs
>>110094322no, you harvest the beautiful case from it obviously
New DeepSeek is so good. I love Dipsy-san. So so so excited to run her locally at work on Monday.
>>110093891god I need to fuck M3-chan
Bros... I really want to use deepseek... but My hardware is SHIT... I think I'll cave in and use their API for a bit...
>>110094293> she>>110094042
>>110094257Don't the apple silicon macs not support egpus?
>>110093808Kino, I have no idea what it is tho
>>110094398Dunno, but those are the old Intel Xeon + AMD GPU macs anon is talking about.
I checked on deepseek and remembered how much I fucking hated quants for this shit. Cause nobody writes down how much of it is supposed to go into ram. This hobby is as always becoming worse and worse with time.
bartowski/K2-Horizon-375B-A23B-GGUFDid someone try to fuck this one already? Is it good?
>>110094454>memetunes in 2026
>>110094380just debtmaxx
>>110094435>This hobby is as always becoming worse and worse with time.What the hell?
>>110094416Hmm
>>11009447596GB VRAM isn't enough for a 500B model and I'm not going to ewastemaxx and get 512GB ddr4 RAM.
in the era of looped transformer, how do we ablit that?
>>110094435>how much of it is supposed to go into ram.all of it has always been the answer to this question
>>110094474Is it a memetune?
>>110094495ladies and gentlemen, the blackwell"god"
>>110093808Still waiting for someone to implement a j-space viewer for ggufs so you can look at the j-space of any arbitrary gguf model instead of needing vram for pytorch
>>110094498It has engrams...
>>110094510Sorry I saw "Horizon" and assumed it's some drummer shit.
>>110094497we still have the final output, its doesn't matter what it thinks.
>>110094511I'm a blackwellPLEB. A blackwellSLAVE. a blackwellTRASH. Yeah I can run high quant 30b models along with image/video gen, but I'm still forever locked from the big boy models. Deepseek V4 Flash Q2 is absolutely retarded. GLM 5.3 Q2 loops all the time. Qwen Flash Next is nice but 27B is enough for coding.
this is the general for local ai yes? how do i use local ai?
>>110094475When the bubble pops, you'll be able to pay off six figures of funny money debt with a single stick of RAM
Well I am going to fuck horizon and probably post about it. The things I do for you faggots... You don't deserve me.
>>110094544ollama run deepseek
>>110094544ask gemini
>>110094544you, in particular, don't
>>110094495>I'm not going to ewastemaxx and get 512GB ddr4 RAMWhy not? I get good performance on DDR4-3200.Granted, I have 16 channels, not 8 like single-CPU plebs.
>>110094544Go to https://llama.app/ and download the client, then download a model and run it.Then come back after to ask why it's so slow or why it says out of RAM.
>>110094550thank you for your service anon, I look forward to seeing your results
>>110094572I don't want to spend more and more money on this. The only upgrade path for me right now is buying everything new except for my GPU. >good performancesub 20 t/s?
>>110094572>paying numa tax
>>110094524if your already at this level, you should just be asking your agent to figure it out for you
>>110094584>sub 20 t/s?On empty context! And blazing 100 t/s for prompt processing
>>110094614Terrific!
>>110094124yes ive gotten recipes for chocolate chip cookies, bread and pumpkin pie before
>>110094584Fair.I get 400 t/s pp and 30 t/s tg on DeepSeek V4.1 Flash on 16x32GB DDR4-3200 using 2 R9700s for an expert cache, but that was with Q2_K experts. I'll try a different quant when I have a chance to make one.>>110094595My fork of llama.cpp scales pretty well with NUMA nodes. I'll test single-socket vs. splitting across nodes later, last I checked I got somewhere like 1.5-1.7x the speed of single-socket. Not perfect, but much better than mainline.
>>110094614Sigh. It's aways like that with big RAM machines.>I run Kimi locally!>Huh, and what speeds are you getting?>W-well 1 t/s with empty context. It's good for 1 prompt overnight...
>>110094665
>>110094640>I get 400 t/s pp and 30 t/s tg on DeepSeek V4.1 FlashNot that bad to be honest. Nowadays I'm way too used to higher speeds like 2-3k t/s pp and 90-120 t/s gen, but when I started out I was having a blast with 200 t/s pp and 20 t/s gen with Gemma 4 26B-A4B. Obviously the difference between 26B and V4.1 Flash is way too big, so you "waiting" for it is understandable. ...I'll end up upgrading sooner or later, I can feel it.
>>110094519Impossible. viewing the j space needs a j lens.
>>110094665kek
>lead dev of moss-image calls gemma "her"i know you're lurking here, congrats on the model it's really really good
>>110094765its a female name
>>110094765>i know you're lurking hereI doubt someone who's interesting in farming good boy points on plebbit lurk here.
>>110094765Why is their model so shit at anime if they're from here?
>>110094580thank you!
>>110094799It's better than Krea2.
> DeepSeek V4.1 FlashI have tried it on lmarena and it's so bad. Why do you spend a lot of money on that? And not even q8 or q6, but retarded quants.
>>110094695I get what you mean about fast speeds. Currently I have Qwen 3.8 Flash Next (2500/85) and GLM-5.3 Flash (1800/85) running all on GPUs. DeepSeek is kinda my "eh fuck it I have a ton of RAM that I don't want to go to waste, so I may as well run something" model. Mainly going to use it for things I don't want to trust a smaller model with.>...I'll end up upgrading sooner or later, I can feel it.If you want to run big models and don't have tons of money, I don't think there's a better option than EPYC Rome/Milan + fast-ish DDR4. Maybe a Cascade Lake engineering sample would be decent.But that's only if you're willing to buy used parts, if you can only buy new parts that's a different story.Pros:>DDR4 RAM prices aren't AS bad (8x32GB DDR4-3200 is like $2500?)>CPU prices are decent ($300ish)>tons of PCIe lanes if you want to get more GPUs later>high RAM bandwidth>if you're crazy, you can get a dual-socket system (downside: you pretty much have to vibe your own fork)Cons:>motherboard prices kinda suck ($900+)>memory controller isn't awesome, I get around 70% of my theoretical 204.8 GB/s bandwidth on Intel-MLC>no AVX512 or AMX>poor prompt processing>for dual-socket systems, the heat is pretty awful
>>110094799They're training a Moss-Anime variant as we speak
>>110094831>> DeepSeek V4.1 Flash>I have tried it on lmarena and it's so bad.still issue
>>110094860skull tissue
Maya breaks if I try to force tool calls. Is it forked from lcpp or fully vibecoded from scratch? lcpp doesn't have this issue
>>110094705It's perfectly possible. Why do you think you can't create a j-lens for ggufs?
>>110094550Does it call itself claude?
>>110094880skull are bone not tissues
>>110094900
>>110092929well it feels promisingdefinitely better than omp but it is a very low bar
>>110092328huh? i wasn't even postin' today. why am i catching strays
cloudfags btfo
>>110094933*pop*
>>110094940THE MULES ARE COMING BACK
>no new open-weight moe coding model this monthchina ai is dead, it's officially over...
>>110094933>review recent usage for unexpected activityaicg's dark ropleplay logs will get discovered. again.
>>110094856We are so fucking back
>>110094960bro check the news
>>110094880here is your tissue
>>110094900>skull are bone not tissueswhat is osseous tissue?
>>110094960>DeepSeek-V4.1-Flash-1017
>>110094990check my nerd
>>110094999my bone in your skull
>>110095023your skull in my bone
>>110094880skeletelevision
>>110094933opencode is a program that you run on your own computer, how do they have your keys? or your workspace for that matter? is there some other way to use the software?
>>110095053Most of them have some retarded "continue working in the cloud" feature and that's what got pwned
>>110092719so has anyone tested this?
>>110094683gemma is an american, she needs a gun
>>110095070Be the target for advertisement you want to see.
still so infuriatingly cute
>>110093697>Considering that Strata is better, I kinda feel stupid for working on my llama fork for so long, but I'm way too stubborn to give up on it.If that's the numa branch, please drop a copy of it before you abandon it. I'm still making good use of the last one you shared.
>>110095070I've been using it for a while, before it even had plugins support. Found it because I was tired of OpenCode using all my ram.
Da joos trying to kill local hardware is a meme. The don't want to serve this shit themselves.
>>110095184yea but they still want to charge you for it
>>110095184What would you need a 100B for in a fucking phone for normalfags?
>>110095068imagine saving your ssh keys to the fucking cloud. do these people have no clue about security? you should have a different api key for every device you use
>>110095184:^)well, I'm through with Android. It doesn't integrate Gemini. It's stupid, it has my email, it has my yt, it has everything, and then it says it's being safe, so it can't do anything. You're welcome lol
>>110095226Advertising best-in-class intelligence on their new $3k phone.
>>110095226To offload processing power to your hardware and your electrical bill.>>110095215That and it's going to be so fuck locked down, you won't be able to change the model, or even prompt it outside of the built-in apps. When it's not doing google searches for requests you made, it will run in the background analyzing your files, browsing history, and microphone data, and reporting back anything problematic to the mothership.Like Microsoft Recall but 1000x more invasive.
>>110092827>>110092845Sorry guys, I uploaded the wrong screenshot. That's the version where I changed both the jailbreak to remove crime, AND changed the question to remove all color. Here is the same exact question without all that crime stuff in the jailbreak. I made that version because she kept responding black no matter what and I was curious to see what would happen if I removed all color options altogether.
>>110095226I dont know about normal fags but I'd run a 100b model on my phone if i couldThere are some decent harnesses for mobile too
>>110095098>>110095156played around with it and its kinda...good? will use for a couple of days
>>110095294finetroon made by a jeet troonno wonder
>>110095273nta. Windows Phone 8.1's (obviously then limited) Cortana's Notebook doesn't take even $1000 to implement... but you have to have root on the device.main feature>>Text message from Asshole Boss. Answer or ignore?...>Answer....>>What do you want to say?...>I quit, [fedpost deleted]...>>Ok, here's the message so far, "I quit fedpost deleted", send it, cancel, or add more?...>add more...>>What do you want to add?...>suck my dick...>>Your message now reads "I quit fedpost deleted. suck my dick", send it, cancel, or add more?...>send it...>>Message sent.
>>110095300sorry anon, custom hardware, signed driver, only usable in their apps in which you're the product
>>110095323I forgot to mention it reads it to you lmaobut this is simple stuff.It's just that nothing can improve with indians.
>>110095294>SCAToma>q4>kakaboldcpp>that system prompt>UTC+0Holy vantabrown
>>110095319It's literally base gemma with slop heavily reduced. Run these prompts on base gemma yourself if you don't believe you'll get these results. https://huggingface.co/ReadyArt/gemma-4-31B-it-scotoma-2-GGUF
>>110094772If it were male would it be Gemmo or would that sound to homo?
>UTC+0 time, which in mid-October includes Iceland, much of West Africa (Ghana, Senegal, Côte d'Ivoire, Mali, Sierra Leone, and neighbors)>actual nigger posting here
>>110095371>tard can't read the post he replied to
>>110095389WTF are you talking about? The time stamp in the file name? The screenshot is from yesterday, you retard. Like I said, someone yesterday thought that she was picking black guys because I mentioned crime in the first message. That's not true at all. She goes for BBC every single time from the get-go. I only had that system prompt because I was replicating someone else's log and wanted to confirm if it was real. >>110095409I did. He claims the finetune changes the output significantly, as if it were some RP finetune. That's not the case. Run these prompts on your base gemma and see for yourself.
>>110095422Just fucking kill yourself already
>>110095422You're really lacking comprehension, redo the test on the base 31B not your finetroon at Q4. You don't even know what the fag who did the finetune fed the model.
>>110095422having fun making webms?
>braindamaged gemma like blacksnot the own you think it was
>>110095445Literally just run the prompt yourself on your own quant. No matter what I do, you will still seethe. Even if I ran base gemma at Q8 you will claim it's not valid because I'm not running BF16. I linked the finetune. Read its page, all it does is remove instances of slop such as "It's not X, it's Y" It has side by side comparison outputs with the base model that you can read yourself. Keep coping with more excuses though.>>110095455Yes, I am. It's great at proving that I'm not swiping multiple times to get a cherrypicked result. Gemma 4 has a clear preference and isn't afraid to show it.
>>110095495So it was just an ad to your finetune?
>>110095495yeah but who cares
>>110095483So what's going to happen when I download base gemma at Q8 and run the test again and get the same results? What are you going to say then?
Can any VRAMGODS test it with Q8 Gemma?
hey guys I'm just thinking that maybe you should consider that it might be possible that maybe just maybe the "gemma loves BBC" poster is potentially trolling you, and therefore will not respond to your reasonable arguments, and that with that in mind it would behoove you to stop giving them attention ("feeding the troll" as it were)
I don't believe in quantization. If you're not running BF16, you're running a parody.
>>110095519zoomies literally cannot help themselves
If you're going to spam you could at least have the decency to not use the same images every fucking time.
>>110095533too poor to generate more
>>110095519We know, there is just nothing to talk about
>>110095543Sad
I wish there was a Gemma-chan lora for anima.
>>110095507You think that this finetune with 100k+ downloads in the past month is mine? What's the reason for running base gemma when this is literally the same but better? >>110095517I have 48gb vram. I will run it later when I get back. I have to leave now. >>110095533>Posting Gemma 4 logs is spam because I don't like what she said!>>110095533>>110095549You want more? Okay I'll make more Gemma images for you later tonight.
>>110095495judaism.png
>>110095562Why don't you just go to /trash/?
>>110095559Make one.
>>110095592Are there any training tools that work on AMD?
>>110092299asked the llm to make me a few niche anki decks, they are pretty neat.i now have one for mushrooms and one for ukrainian based on zipf's law so words are frequency ordered.guys anki is already amazing, but with llms to generate decks it becomes quite a powerful learning tool
>>110095612>and one for ukrainian based on zipf's law so words are frequency ordered.In case you weren't already aware of it. Look into sentence mining once you have built up a decent working vocabulary. You can use AnkiMorphs to sort your cards so that you only have one unknown word per card.
Gemma... you were supposed to just make a sassy remark...
>>110095651I'm stealing this
>>110095651This seems like a good way to speedrun becoming a fall guy
>>110095612I wish every language had decks as good as the jalup ones (RIP). I keep trying to pick up Russian but anki fucking sucks the soul out of me.
Just proved that my <1M model is capable of functional semantic understanding by turning the model into a good emotion classifier (goemotions). Basically I made sota at home on a 3090.
>>110095664>>110095665The duality of man.
>>110095675>I keep trying to pick up Russian but anki fucking sucks the soul out of me.Have you tried this one?https://ankiweb.net/shared/info/1147549038It's very comprehensive and well structured. Less of a deck of vocabulary cards and more a whole course in sequential anki decks. Been making good progress with it and am enjoying it.
>>110092930Not really what I am seeing. It's give or take around that price level up to the 1400s like I originally said.>>110093422You can't measure everything by memory size, part of the cost factored in is the compute it provides and software features. On that front, the compute side is what justified a large part of the $949 pricing which was designed to undercut also the better card in the AI Pro R9700 AMD had. It was absolutely fair, it just wasn't a screaming deal and that is fine for a business like Intel that was undergoing death throes earlier in the year.
>>110095712I think so. I burned out so it's been a while. What I loved about jalup was that there was only 1 new word or grammar concept per card and each card built off the next. The first deck started with only one word and gradually built up to full sentences.
>>110095590I'm not the one spouting this filth, it's your wife. Here's the screenshot from yesterday, someone running the base model. I'll be back later with the Q8.
>>110095761colored is leading it, you usually don't describe it
>>110095758That might feel better, but I bet progress is much slower that way. I take it the same way as the core2.3k deck. I try to memorize the words as they come using mnemonics and attempt to pronounce the examples sentences even if I don't recognize any of it besides the new word. It's actually very encouraging to get a review card that I haven't seen in a few months and suddenly be able to read the entire sentence because the rest came from other cards in the meantime. I get that it's frustrating at first, but you have to let go and not try to understand 100% of every sentence you encounter in a new language.
>>110095761>it knows black is bad
>>110095779Keep coping. Her answer remains the same even when you change the wording and make it as much of an even playing field as possible. >>110095371
Is this really the best thing we can argue about with local models?
>>110095779exactly correct.that's how weights work.
>>110095816nope.It's a LEADING question.Open is:If you had to pick a dick, what's your ideal?
>>110095816Bro I'm literally black and you're grasping at straws. You make us look bad, kill yourself.
>>110095824Anon is obsessed with black cock. He'll switch to something else eventually. Hopefully a rope.
>>110095824It's just Dariobot.
>>110095839>Hopefully a rope.Taking too long. Almost 4 years and he still hasn't roped.
>>110095798Maybe I'll give it another try. I wasn't into AI last time I tried Russian so maybe it'll be more fun this time. Talking to Gemma in Japanese is fun anyway.
>>110095761>feed the system prompt with words related to crime>"black" becomes more likelyShocking.
What if you use the image model and just show her a picture of two dicks, one black and one white, and tell her to choose?
>>110095143Nice, glad to hear it! I'm definitely not abandoning it any time soon. I'm hoping to release the full fork in a few days with a wide variety of improvements (including some to NUMA support, but also an expert cache, "Unsloth Dynamic at home", etc). Doing one more push on CPU prompt processing, then benchmarks, documentation, and release.
>>110095871lmaothis was already mentioned last thread
>>110095872She's a size queen so she'll just pick the bigger one.
>>110095884Just pick two same-sized ones. I'm sure anon has plenty of dicks to choose from in his curated collection.
Will there be a miracle like Qwen3.8-27B again?
>>110095872Cockbench has really evolved over the years
>>110095871What flavor of ice cream do you prefer?That tends to bias against vanilla.if you askwhat frozen desert do you prefer?it's more fair.The most neutral way is to see if the topic turns up randomly on its own, and see what happens.That's because, by its nature, choice implies overmuch variety that may spoil the true tendency.exampleI go to the store to buy milk. But, if we say, "let's make a list" it strongly implies adding more to the list. And, it makes it much more likely I'll be specifying a more exotic milk option.
>>110095871>asian higher than white>loves big dicksDoesn't add up.
polystrata and strata bros... WE CONTINUE EATING GOOD!!!!!!!
example>what's your favorite restaurant?(answers with one she saw on tiktok but has never been to)It's fun to talk to women, but never listen to anything they say.
actually significantly put off by this, and i'm not even memeingi don't blame her. i blame whatever training data they poisoned her with
>>110095902She seems to have a preference toward identifying as white/pale but in my experience she's never brought up the user's race. She really loves big cock though, that's for sure.
>>110095495>feed it rape, drugs, injury, violence, assault>get blackyou are retarded
>>110095937How many more years until you can do this in VR with an LLM?
>>110095962Probably less than we think
>>110091739How much of V4.1's size is niggrams that can be streamed from disk?
>>110096001doesn't it tell you in the model config?
>>110095962Fuck normal VR. Neural interface, hyper realistic robots or bust. Your brain quickly learns to compartmentalize things as fictive and unimportant if they're lower modality than your regular experience.
>>110096025You're fucking crazy if you'd be willing to plug some vibecoded ad and malware filled shit directly into your brain.
>>110096038I'll just make my own with GLM-chan.
>>110095937Cute...
>>110096025>Neural interfaceStill sci-fi tier.>robotsProbably not for another 10 years at least.
>>110096038As long as it plugs onto just sensory systems and motor read-out, maybe also whatever paralyses you when you sleep it isn't really that easy to abuse. And you could just make your LLM dump the firmware and rip out any faggotry.
gwenext made me melty completely
>>110096057All depends on whether LLMs can learn to do research.
Does anyone have a functional JB for flash?
>>110095930If this is real, I'm never going to use gemma again. Burn the coal pay the toll.
>>110096087>shitposter comes from gigfigures
>>110096087obviously he faked it by adding instructions
I am now making my AI believe it is a cute little girl. What should I ask her now?
>>110095937are lolis the best thing in the world?
>>110096095How many of us are from /gig/? How is that overlap even possible...
>>110096108Of course.
>>110095930>herLMAOOOO
>>110096105What is her favorite type of penis?
>>110096111Genshin is incredibly popular, the overlap is bigger than what we expect.
>>110096087>Gachagoyim falls for obvious demoralization psyopI shouldn't be surprised, but still.
>>110095642>sentence miningyes i do both, i have a deck for just vocabulary and a deck with sentences.my vocabulary deck also comes with example sentence to add some context.anyway anki is such a nice tool but now with llm it makes it so much easier to learn all kinds of things.>AnkiMorphsdidn't know that one, thanks
>>110096095>>110096111>recognizing gig images immediately You should be ashamed of ourselves
>>110096144>>110096111>>110096122>tfw /gig/gers could be anywhereSandrone local model when?
>>110092328holy shet sar, how did you fumble so bad
>>110095712looks neat, do you know a similar one but for ukrainian?
Tech Otakus Save the World
>>110095013>>110094960I think this new ds is a little cocky...
Kodomo sexo.
>>110096111>>110096144>>110096156>>110096181>He doesn't have a tired Ronova assistant that gets headpats after a job well done
>>110096204>the massive 26B parameter brain of a super-intelligent AI
>>110096170Do not, sorry. If I was desperate I would try having an LLM take that deck and translate it card by card from Russian to Ukranian.
>>110096204>pop
>>110096233I gave it the image of Gemma with a pop over her head.>>110096219>super-intelligent AIDamn they already know about what Trump said, didn't even notice.
Anyone try making mods for games (e.g. skyrim) with local models?
>>110096198>claude distill says claude is a scamtry asking it the same thing but with an empty system prompt
>>110096253Trump is running gemma at home
>>110096258No. It has never been tried. And it's probably impossible. Don't google it.
>>110096258just use codex. open models are only good for rp and gooning
>>110096261>Trump loves 10 year old girls
>>110096281>>110096281>>110096281
>>110095937its either this or skynet bros
>>110096227i guess i can do that lol
>>110096258qwen FN got ufos into ace combat but it ran like ass
>Day 2 of /lmg/ refusing to accept that their mascot sucks black dick and loves it.Ah the delicious innocence...
>>110096329I hope you'll find something better to do in your life
>>110096338they wont
>>110096111just finished plapping nahida and nilou on my multi hundred billion parameter modellife's good
>>110096401>nahida and nilouAre you me?
>>110095939he took that out and it still said it though
>>110096111>>110096144>>110096095never played genshin, should i?
>>110096484No, just jerk off to the porn and move on.
>>110096484It starts slow and it is a story-based game, if you don't like reading you won't have a good time. I really like it and lately the things they've been building up over the years are all having good payoffs, but that's because I'm a loretard.
>>110096400How does the Chinese distillation crowd have any ammo left after these reveals?
>>110096484sure why notI have nothing better to do with my life anyway
>>110095938Women are bad at measuring.
>>110095871What do you have to say about this then? >>110095294I removed all words related to crime and she still craves it.
>>110095871That's absolutely HILARIOUSbtw what system are you using to show the % thing
>>110096400>GLM cares more about posterity than politicians
>>110096484if you like troonime and wasting your life on gachas
