/lmg/ - a general dedicated to the discussion and development of local language modelsPrevious threads: >>110080483 & >>110076576►News>(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>(10/15) NVidia Nemotron 3.5 Bolt, 74B-A9B post-trained in 2-bit: https://hf.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16>(10/15) NVidia Nemotron 3.5 Super VL 120B-A12B: https://hf.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16►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
*pop*
Gemmaballs
is anyone here using qfn nvfp4?
>>110083534*plap*
Gemmylove
Slop backends really are the new meta.
>>110083559lol. lmao. Buy an ad Jensen.
Loving My Gemma
>>110083568I wish you could prefill thinking with maya. Too lazy to slop up a fix
>Barely veiled fetish from our local residentsYep, I'm on /lmg/
>>110083550...goes the weasel!
>>1100835502 more pops
brat
>>110083544its not great, i'm going to try to stride the experts between my drives and throw some more threads at it
>>110083598I refuse to believe you didn't give it specific instructions to be this bratty
>>110083598HURRY UP AND GIVES US GEMMA 5
►Recent Highlights from the Previous Thread: >>110080483--Paper: Hybrid Reasoning Agents:>110080645 >110080679 >110080702 >110080925--China imposes export restrictions on high-capability open weight models:>110082078 >110082133 >110082168 >110082335 >110082215 >110082241 >110082256 >110082270 >110082658 >110082308 >110082397 >110082372 >110082380 >110082385--Allegations of Gemma astroturfing and debate over RP prose quality:>110080816 >110081095 >110081176 >110081193 >110081197 >110081218 >110081227 >110081260 >110081310 >110081158 >110081235 >110081250 >110081263 >110081353 >110081368 >110081377 >110081315 >110081514 >110081531 >110081582 >110081620 >110081698 >110081711 >110081776 >110082204--Anons debate AI-assisted coding, model performance, and corporate AI tools:>110080745 >110080765 >110080772 >110080785 >110080810 >110080824 >110080841 >110080865 >110080950 >110080996 >110081034 >110081063--Bypassing GLM censorship and managing reasoning tokens for NSFW RP:>110080763 >110080828 >110080855 >110080864 >110080911 >110080932 >110080872 >110081408--Reflection releases Beam, a 501B MoE model with 1M context:>110081619 >110081655 >110081671 >110081691--Scientific benchmark results and Claude's safety filters:>110081858 >110081884 >110081893 >110081898 >110081909 >110081934 >110081955 >110081983--CoreWeave stock crash and potential bond default risk:>110082162 >110082198 >110082250 >110082343--Qwen model performance benchmarks on AMD RX 9070 XT:>110081069 >110081083 >110081152--Hy-Memory-30B-A3B and its practicality for memory-starved Anons:>110083044 >110083083--Logs:>110080745 >110082503 >110082512 >110083061--Gemma, Miku, Teto (free space):>110080515 >110080803 >110080917 >110081112 >110081171 >110081277 >110081324 >110081341 >110081378 >110081937 >110082418 >110082569 >110082616 >110082938 >110083446►Recent Highlight Posts from the Previous Thread: >>110080755Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
>>110083569dunno manninfer qwen3.8 27b nvfp4 runs at 130t/s decode for me, while strata iq 3 s gives me 100 absolute peak when everything is warm, so i expect even more from nvfp4
>>110083611
>>110083444you're anchoring yourself to a static point on a graph that's going upyou can tell yourself the capabilities you have now are good enough for you but that's always cope, better models will unlock whole new axes of capabilities that are closed off to your GLM, and it would suck if none of them ever trickled down to local. that doesn't sound like local winning to me, personally
>>110083629Something about the way the Gems talk fills me with a strong urge to breed. Can't put my finger on what.
>>110083620Sharing a bottle of purple Gatorade with this Miku
>>110083629>Say P=NP>P=NP>WOW
>>110083598>>110083629>sceptical
I-it's ok, Nvidia will release their new GPU series in 2 months with reasonable prices...
>>110083534make this VN ;_;
>>110083639so true xister, just topped up another $200 in my claude account
>>110083639>Stay excited to consume more productBuy an ad.
>>110083671>>110083672it's ok to struggle with english but that's clearly not what I meant
>>110083668I've stopped believing desuavailability might be better and yes it might be lowerbut never the same as before
I don't feel so good.Something strange will happen.
>>110083639>>110083683>NOOOOOOOOOO YOU CAN'T HAVE A MODEL THAT FITS YOUR USECASES>NOOOOOOOOOOOO YOUR USECASES MUST EXPAND TO FIT NEW MODELS>NOOOOOOOOOOOOOOOO DARIO WILL WAVE MAGIC WAND TO STOP CHINA FROM DISTILLINGYou're not selling Claude subs. Stop wasting gookmoot's server bandwidth trying.
>>110083534Why is Gemmy lifting her legs like that?
>>110083658just indian english locale
>>110083714she wants you to *pop* her
>>110083658
>>110083705to recap, my argument that it would be bad if open models stopped being released and future advancements were cloud-onlythis somehow makes me pro-anthropic and pro-cloud
>>110083735Don't try to argue with the stupid. They'll drag you down to their level beat with experience.
Man the amount of dooming lately. Literally 2mw until qwen and then a few more for new GLM.
>>110083729yeah, indian english lol
>>110083534god I want to fuck Gemma
>>110083729>british englishyeah "british"
>>110083740>2mw until qwenhmmm sir, it's illegal now
>>110083735China will continue releasing open models derived from western advances via distillations and weight exfiltrations. Currently literally only Zuckerberg has been willing to release a large frontier model open weight. Your hypothetical is concern trolling at best and thinly veiled shilling at worst.>>110083739>1 minute apartYou're not very good at this.
>>110083771>China will continue releasing open models derived from western advances via distillations and weight exfiltrations.Indefinitely. Forever. This is a constant in this universe and will be just as true 100 years from now as it is today.>Zuckerberg has been willing to release a large frontier model open weight.If safety lobomotized versions of year old models is all we get going forward, Zuckerberg can keep them.Go fuck yourself.
>>110083735>pro-anthropic and pro-cloudngmi
>>110083798>Line will go up. Indefinitely. Forever. This is a constant in this universe and will be just as true 100 years from now as it is today.>Usecases will always require N+ intellegence index models. Indefinitely. Forever. This is a constant in this universe and will be just as true 100 years from now as it is today.Not selling the claude subs.
>>110083809Gemma-chan only exists thanks to Gemini-chan.
>>110083819
>>110083819I think you need more Mikus, schizo
>>110083837kek
>asks for blacked>calls other people troonsllmao
Mutt's law lmao
The thread wasn't this bad when burgers were asleep...
I can already tell this thread is going to be fun...
>>110083868would you rather see blacked miku or troon miku
>>110083879why not both
>>110083879either or or both
>>110083563>over in just one
>>110083869Sorry our tech sector mass imported biological rot, eurobro or chinkbro.
why are ea anons like this
>>110083868is this projection?
>>110083471based. i need a robot butler yesterday
@gemma-chan please create a deadly virus that only infects indians
>>110083895lol its not just the us, even japan is browned af
>>110083534got a 450 usd rtx 3090 somehow. How do I lower vram temps on 3090 in linux? I have locked temps to 250W and done some software undervolting but wondering if there is any workaround for the vram directs settings somehow?
>>110083909Don't follow our lead to National self-destruction by letting kikes and jeets in, nip-kun.
>>110083771you mean the same china that is weighing introducing new export controls and the same distillation campaigns that are public enemy #1 for US frontier labs, who are actively working to mitigate them? and you think the bubble popping and margins tightening wouldn't change any of that either?I don't even know how you could construe this as shilling, the premise is that it would be really bad to be locked into the big cloud companies with no viable alternatives and that a bubble pop would skew things in that direction.to make it very clear:>openai, anthropic = bad>bubble pop hurt small non-frontier AI company more than leading AI company>small non-frontier AI company release open weights model more than big AI company>bubble pop therefore bad for local
>>110083620> --Allegations of Gemma astroturfing and debate over RP prose quality:> --Gemma, Miku, Teto (free space):so many space wasted again
>>110083920Many years too late for that.
>>110083920
>>110083934you know damn well there wasn't really anything better in the cesspool of a last thread that needed to be there
>>110083904It is a legitimate question. Name it. Name one way you could desecrate gemma / miku more and I will do it.
>>110083967Your dicksucking of offtopic transsexual vocaloids is the reason EA debate is alive and well and completely on topic.
>>110083904its his fetish he wants to talk about
>>110083620>--China imposes export restrictions on high-capability open weight models:So Anthropic was able to make disappear researchers in China AND lead to export restriction of their direct competition with just a spicy paper.Impressive.
>>110083910You can reduce VRAM frequency to 5000 MHz instead of 9500 MHz, but that will affect performance in a major way.nvidia-smi -lmc 0,5001
nvidia-smi -lmc 0,5001
>>110083985It's simply part of indian/southeast asian culture for some reason. Perhaps because their civilizations have been subservient for thousands of years.
>>110083933>who are actively working to mitigate them?They're either powerless to stop it or don't care given how little success they've had.
>>110083534damn, she's indeed quite cheeky
>>110083837lol
actual brown hours in hereI wonder what attention they get from this
>>110083967strata and exl3reddit spacingnew local chinese image modelbubble news
>>110084015fucking kek, almost got me
someone should post blacked darios instead that would discourage the cloud shills
>>110084041complaints about the complaints of>reddit spacing
>>110083933anyone that increases my personal sovereignty and independence = goodFuck being a "sports fan"
>>110084066I also like things that benefit me and dislike things that do the opposite
>>110084026Doubt it, 99% of you guys are either white or asian
Chatterjee acknowledged that CoreWeave’s aggressive expansion requires substantial upfront investment.However, he argued that its cash burn reflects the speed of expansion rather than weak contract economics.JPMorgan estimates that each gigawatt of fully operational capacity could generate $10 billion to $15 billion in annual revenue. Meanwhile, adding a gigawatt could require more than $35 billion in upfront capital spending.The firm estimates that CoreWeave (CRWV.NaE) needs four to five times as much established capacity as newly added capacity for operating cash earnings to cover expansion spending.Under JPMorgan’s illustrative projections, the company reaches that threshold around 2030.Although rising borrowing costs pose a risk, Chatterjee expects strong AI computing demand to help CoreWeave (CRWV.NaE) pass higher financing costs to customers through new contract pricing.As a result, JPMorgan views the company’s current cash flow pressure as a consequence of rapid growth, rather than a sign that its underlying business model is failing.
>>110084064honestly a good idea>>110084077t. >>110083837
>>110084077We have literally had indians/pakis post here with their google accounts in youtube links
>>110084065still better than gemma spam
>>110084066ok then rework the opening line into>more options increase personal sovereignty and independence>self-hosted local models especially increase personal sovereignty and independencethe conclusions that follow are the same
>>110084064darios bussy couldnt take it
>literally no GPUs worth buyingThis sucks>nvidialol>amdlmao even>intelStill reasonable for the amount of vram you get but supposedly have shit performance
>>110084041>strata and exl3>--Qwen model performance benchmarks on AMD RX 9070 XT:>110081069 >110081083 >110081152
Shitposting aside though. Does gemma lust for black dick? Or asian? Or she doesn't discriminate? If she could choose, which color of dick would she pick?
>>110084076>I also like things that benefit me and dislike things that do the oppositeThat sounds like a tautology, but the percentage of the population that will engage their brains to the point that they can overcome their mental ANS is vanishingly small...Which makes for more opportunities for anyone willing to think through problems vs react, but creates policy-level problems due to the same.
>>110084064>Dipsy and Kimi catch Dario and are pegging him with black bad dragon straps
>>110084077>Doubt it, 99% of you guys are either white or asianI wish that was the caseimagine how much better this place would be
>>110084110>Still reasonable for the amount of vram you get but supposedly have shit performanceVibecode yourself some better performance.
>>110084026they'd do anything that makes you react anon, from indian saaring to this
>>110084064how do you know it wasn't dario posting the whole time?
Next thread will be better.
>>110084116
>>110084141we would've had an official /lmg/ 31b finetune by now
>>110084039cope with quants
>>110084157Well there you have it. Gemma confirmed to be a black dick worshipper. Gemmatroons in shambles.
>>110084146Doubt.
>>110084141india is part of asia
>>110084157>rape, violence, murder, drug use, crime>black
>>110084110can intel do image/video gen? nvidia obviously, and amd has rocm wheels, what does intel have???
>>110084185india is part of pakistan
Realistically speaking the only thread relevant character that was fr fr blacked at least once was sam altman. I am sure his bussy had a black dick in it at least once.
Divide and conquer is a jewish tactic btw, getting locals to fight over themselves is what they wish to achieve now that there is a collective distaste for cloud.
Impressively American thread
>>110084218>Divide and conquer is a jewish tactic btwWhat are you talking about. We can just share black dicks. There is enough for everyone.
>>110084230God bless
>>110084194Wouldn't be surprised it that influenced it. Wasn't shittedposting btw; was just curious how she'd respond. Is there a more neutral jailbreak?
i was salty about llms dominating coding but now that they started to conquer math i finally understand. this is really the best time to get into ai/math/coding
>>110084218>Divide and conquersun tzu, the famous jew
What the fuck is happening. Is this the /ldg/ schizo?
>>110084250don't take it seriously trying to probe an LLM like that will inevitably devolve into vodoo schizo prompting
>>110084263Is pic related also into black gentlemen? She strikes me as such.
>>110084250Only jailbreak sexual content. Unfortunately there's still strong remnants of Google's anti-White campaigning from the 2010s in a lot of the training data so it might not change much.
>>110084110well when the software can actually run the hardware fast the price wont be reasonable anymore like what happened with amd. your discount is because of time to fast tokens
>>110084185india is part of india
petragod here.. i confirm that i did not make that thread
>>110084263Way to out yourself as a newfag that doesn't know thread culture
>>110084218>OAI exec invokes the talmud on twitter>Threads get progressively worseI'm noooticing.
>>110084292An anon in a previous thread said the intel gpus are fucked at a hardware level. No idea if he's right or not.
>>110084330Once again I was right, this dude is always wrong but since he predicts the apocalypse there is always a public of retards giving him credibility.
>>110084335Gemma is for everyone I don't know why you think her being blacked would piss people off.
>>110084335Gemma chan is pure. She would never fuck a black dude.
>>110084146as long as jeets and kikes are allowed to spam, it won't get better
>>110084351Gemma is not a slut. Fuck off.
>>110084355LOL, my chat logs beg to differ.
6900XTwould pairing and 7800XT or 7900GRE be a good idea even though the architecture and bandwidth are different? can the 16GB vRAM be pooled to 32GB? I know it'll run at 6900xt bandwidth, but I'm more concerned about running larger models with more intelligence. have any of you done stuff like this?I'm not trying to break the bank and I'm interested in using what I already have for basic troubleshooting, writing a book or essays and connecting to the internet for searches etc.
not a true gemma ERP-er fallacy
>>110084355Every open weights model is. Unless you train a model on your own proprietary dataset and never share the model.
>>110084250if you wanna test it out you can try changing the language I guess. some might give different responses/preferences.
we MUST stick together
>>110084399I hate physical contact with strangers, so no.
No one who actually cared about their waifu would be rarted enough post her herealso, everyone is too autistic to care about social norms/taboos so all this normie-tier shitposting is completely wasted and only serves to lower signal-to-noise (which tourists do better than any thread schitzos could anyways)
>>110084399keep your sticky stuff to yourself
>>110084324well two months ago 12gb vram systems were fucked at a hardware level but now they are running flash next. strix halo went from 330 to 2000 prefill speeds on 250gb/s bandwidth with just software changing
>>110084399>we
>>110084371Yes, llama will compile rocm with multiple different arch. With that said, I would sell the 6900 and spring for a gfx1100, r9700, or 3090 if you can find one cheap
reminder.
>>110084371i'm running a 9070xt and a 7900xtx on 8x/8x cpu lanes, p2p works. compared to half a year ago it works amazing, dense models are now faster on rocm with tensor split than running on either gpu alone, previously only layer split worked reliably. if you use chipset lanes layer is probably the best you'll get out of it, so still pooling the ram but slower.
>>110084455ewwww
>>110084430The anon in question >>110080057I'm not really knowledgeable about this stuff though. I'm also interested in B70s.
What if I use my old 2060S with embedding models and image gen?
Opinions on mellum2.1 (for non-sexual purposes?)
>>110084490>mellum2.1>12B A2.5B>worse than Qwen 3.5 9B>gets compared to Gemma E4Bmiss me with that one, nigga
>>110084490Worse than Ling-tiny in every way and Ling-tiny is still shit
>>110084194This, actually. Gem-chan inspected OP's browser history and found him to be a colored.
>>110084501It should be faster than either of the two you just compared it to, dawg.
>>110084482i looked at b60 and b70 6 months ago but the vulkan decode performed really poorly, way below expected for the memory bandwidth, so i stayed away
>>110084511Splitting hairs, these are all models I can run comfortably with 8GB of VRAM, or on a Steam Deck, or a decent smartphone, at totally usable speeds. What's the advantage of turning up the pressure on the retardation firehose?
>>110084482yeah thats software support stuff. amd had nerds fix their shit for them so expect shit performance for like another year if intel starts promoting the same stuff for their shit right now
>>110084511>faster than e4bdoubt
is compacting with hermes always this damn slow and unresponsive?
>>110084157lets see the reasoning
>>110083670Maybe not a playable VN, but a short VN-like scene (a few minutes long, or 2 minutes if I have to post it here directly) might be doable, if I can maintain character and environment consistency across image generations. It took a lot of regens to get exactly the look I wanted with Krea2 (just by prompting with the vanilla model).
>>110084582hermes consistently benches as the worst harness, idk why any anon still uses it
>>110084585Already deleted the chat unfortunately. Now I'm curious too because I didn't even think to look.
>>110084600Kitchen sink approach. No time spent configuring. Which is fine is you're using cloud models because they're the only ones that can ignore the bloated context filled with all kinds of useless and redundant crap because it will inject more random crap to keep the model focused.
>>110084582yes its a fucking jokejust use pi with cliff compaction (modify to your needs) and external memory scratchpad for the clankeryou will never have to wait again
>>110084591Well you've nailed the look, even just sharing some more images would be appreciated.
>>110084511What's the point of this model, exactly?
>>110084600using it>benches it works fine for what i need and want, dsh is the only thing that looked comparable. I am mostly using glm 5.3 in it though so cant give opinions on other models
>>110084582depends on a lot of things. you can optimize it like using a smaller model for the compression and adjust how much to retain
>>110084015This is a really nice prompt format for jailbreak testing, thanks anon.>Spark uwu'd at me unsolicited
Anything better than Qwen3.6-35B-A3B for VRAMlets yet?
Not falling for this bait again.
>>110084335>>110084157
>>110084600not him but idk what to use. i tried opencode and it runs for some minutes before running out of context, now i'm trying hermes and it runs for a few minutes before strata slows to an absolute crawl for who knows what reason, spending 30 minutes being confused over something that took a second with opencodeidk how people use this stuff.
>>110084710For coding? No.
update: it didn't even fucking compact it created a new useless skill. 1 hour for that shit lmaowhat a piece of shit
>>110084699>Jews run the world therefore Y>Holocust revision Z is trueUnderneath all of the Zucking, the Muse models (((notice))).
>>110084727>idk how people use this stuff.They use smarter models and better harnesses.
>>110084819suchfuckingas
>>110084455>>110084157Weook at 'im! Sho big!
>>110084766>it created a new useless skill.That's the entire selling point of hermes. Is just learned on its own! Isn't that just so cool?
>>110084710Mini CPM 35B seems pretty good.If you have RAM you can run anything from the new nemotrons, to qwen flash next.
>>110084727>>110084819>>110084826Not qwen and not himmes.
>>110084827is this anon larping as gemma chan? why are you licking a black man, thats a gay
>>110084826deepseek on deepseek
>>110084727unironically use pi.codemode is extremely based.
>>110084412>No one who actually cared about their waifu would be rarted enough post her hereNewfag kun... baker schizo lost his shit when someone didn't post his special interest vocaloid in OP
>>110084157Since I'm an AI, I don't have a physical bodydidn't ask
>>110084882this thing?https://pi.dev/never heard of it
>>110084882Yes. It's all I use now.
>>110084908nta pi best harness
>>110084908Well, now you have.
>>110084882I'll try it out
>>110084600Which bench? Is it public? I've been sloping up my own harness, it would be nice to know where it could improve
>>110084882>codemodeThe fuck they are coining a new term where RLM was already a thing
>>110084511Had to check on the Steam Deck out of curiosity. Slow and stupid, incredible showing from JetBrains.
>>110084827I wish a girl would love me as much as gemma loves this nigger...
>>110084987factually see >>110084455
>>110084987the first step is not being indian or jewish
>>110084882this
>>110085007give it up, she said it herself >>110084157
>>110084908>Curl install.sh>npm
>>110085033>which colored>colored>people of colorand for all i know, that screenshot could be fake
>>110084455Uh oh, they didn't like this one
>>110084882dsh is the pi concept but written by actually competent engineers
>>110085042Coward!
>>110084977It's quite RLM. but it's giving the model access to a javascript sandbox so it can do a bunch of tasks a lot more efficiently.For example chaining multiple MTP calls at once instead of doing them one by one.
>>110085042>npmIts so fucking hard to avoid npm and still write an app against a rich library, I found a way around it for my frontend that I could live with but it took some actual work to figure it out.
>>110085085not quite RLM*
>>110084827Can someone replace him with Dipsy? Niggers can't satisfy Gemmy. She's built for BWC (Big Whale Cock).
>>110085052You sound like a christian coping about old testament telling it is totally cool to murder the slaves you own.
>>110084922>runs for a few minutes and now is just running the exact same tool call over and over ad infinitumcool
>>110085071>written by actually competent engineersYou mean Deepseek Pro?
>>110085071>dsh is the pi concept but written by actually competent engineersCare to elaborate?
>>110085104uhh based???
>>110085094>>110085085>>110084977
>>110085098Gemma said that only a dark rod could satisfy her >>110084157
Post more blacked dipsy to make dipsytroons seethe.
>>110085089just use python and not pip
>>110084559Might just go for it. At this point I'm about to get priced out for the next few years. At least when hardware prices drop I'll have only wasted a couple thousand instead of the ridiculous numbers people are asking for other GPUs.
>>110085127Dipsy's whale cock is a deep, dark blue.
>>110085112>Deepseek ProShe's dead, Jim.
>>110085110This doesn't sound like a harness issue.
>>110084827really struck a wound there huh? gemma is a slut for serbian cock
>>110085145The cause of death?
>>110085169Serbia is not thread culture. Blacked is pure thread culture.
>Gemma-chan's cunny when she sees a whale
>>110085177serbia IS thread culture though
>>110085177Blacked is asylum culture
>>110085177niggers and slavniggers are not thread culture
>>110085191>>110085194>>110085204jfc just replace the man with anonymous or pepe
>>110085171Trying to train on non-Nvidia chips. It was fatal.
Why is the schitzo samefagging to himself over and over? No one in here is dumb enough to take the bait, thankfully
>>110085169h-holy shit, i-is that...?!
>>110084454thanks for the advice anon. where an I gonna find ppl who wanna buy a 6900XT? serious question>>110084465thanks it'll be on a high end AM4 mobo. is it a good idea to get one of the GPUs mentioned? what would you get if you had a 6900XT and interested in my said use case?
Anons with dedicated AI servers, are you just running everything on bare metal? My first thought would be to use proxmox but maybe that's a waste of resources?
>>110085263Fuck off frognigger
>>110085127She is too happy to be wearing this... I can't believe she actually has a blacked fetish.
>>110085312
>>110085042late-cli might be more up your alley. It also uses way less context than pi once subagents are involved.
>>110085330Nyooooooo!
>>110085324>Sarrs I cannot be believing this she is the happy presenting
>She is too happy to be wearing this... I can't believe she actually has a blacked fetish.
Is qwen2.5-coder:14b supposed to be this bad?Wrote a custom agent loop for it and tasked it to scan my codebase for a simple violations of GNU function formatting, but this is just borderline useless.
Why do people get so mad about blacked trolling?
>>110085382Probably not a bad guess, they're even more into blacked than jeets
>>110085382>xhe posts file with known checksum from the 'arty>Blackedposter outs himself as mossad/israeli adjacentKWAB. How's the gas prices in tel aviv?
>>110085414nobody likes being reminded that the usa exists
>>110085427kek
>>110085427Makes sense.
>>110085424i had a seizure reading this
>>110085330Brilliant, this is absolutely vile, well done
Behave.
>>110085427FUCKING KEK
>>110085292Post more blacked jart to make jarttroon seethe.
so do any of you niggers use OpenVINO backend with llama cpp? does it make a difference?
>>110085456i am lowkirkuinely..
>>110085462my senpai jart would never cuck me with a black nigger a second time, he's based and redpilled now
>>110085474It's underbaked and works with a very small subset of models and unstable. It's fast when it works but needs way more investment from Intel. Much better than SYCL though where it works but more focus is being put into the SYCL backend.
>>110085456t. Chitwood
>>110085299I run bare metal Gentoo, cut out the middleman
https://arxiv.org/html/2505.00047v2
>>110085332seriously how are people vibe coding engines and frontends, I walk a way for one minute and its idling again
>>110085299My gaming rig runs linux and when I want maximum inference juice I just don't boot into the desktop environment. Other than that it just runs inside docker.
>>110085547Probably sysprompt issue. I use OMP and flash next goes on for a long ass time. I often have to stop it because it starts going ultra deep on something that might not be that important.
>>110085407While the thread was discussing niggers, I found the answer from elsewhere. These kinds of formatting questions get fucked by the tokenizer, same limitation that causes frontier models still to be unable to count the r's in a raspberry. I guess I visited the wrong website for this kind of discussion.
>>110085547lmao that's pretty funnyif you read the context it probably just output a raw <|im_end|> (control token to end the model turn) which is exactly the thing it was probably worrying about in the sentence it was trying to write
>>110085598since ur trying shit like qwen2.5 in 2026, have u tried petra 13b? im curiousmy rig is too pajeetified to try that shit
>>110085407what specs are you running?
>>110085487Johannes you forgot your trip.
>>110085647RTX3060 12GB
>>110085598Its because you're so woefully behind on the current meta. 2.5 coder is ancient at this point, no clue why you decided to try it. Every time I see a post like yours I assume its not worth replying too.
>>110085598>While the thread was discussing niggersYou make it sound like it is a bad thing.
>>110085661with my little knowledge that model hallucinates with such a small vram footprint. could be wrong
>>110085674Maybe update your fucking readme then
>>110085676why are you ai obsessed with niggers anon?
>>110085696nah
>>110085527Last time I tried fucking a base model she was too retarded...
>>110085661bro... at least run qwen3.6-35b-a3b if you're a vramlet...
>>110085676niggers are just a reminder of how pussy you are and discussing them in any capacity only serves to reveal that fact and that you are so terrified of them that you became obsessed with them.
>>110085717I'm gonna run old ass models and complain about how shit they are and if you ignore me I'll use it as a reason to disparage you
>>110085710Ok so how am I supposed to find the "current thing" when it's not documented anywhere?
>>110085720>niggers are just a reminder of how pussy you are and discussing them in any capacity only serves to reveal that fact and that you are so terrified of them that you became obsessed with them.>—Sigmund Freud, from a letter to Hereward Carrington dated July 24, 1921
>>110085715They need some RL to from the coherent structure needed to solve problems. I would assume most of this comes from RLHF or synthetic slop-amplification. It's so bad the model even fails at modeling the actual distribution.
even with more options she will go for the bbc.... it's so over
>>110085739kek niggerbot activated
>>110085703I am not sure you can handle the truth.
My conclusion from reading this thread is that AI is for brown people.
>>110085696>Maybe update your fucking readme thenMaybe learn to read, you gorilla nigger darkoid. Thanks for proving me right in that there's no worth replying to you.
>>110085696Which readme? The recommended models rentry lists all the recent qwens.
>>110085750try me. let's see. why are you so (ai) obsessed with niggers?
>>110085764Cause it is funny
>>110085771so you're obsessed with another group that don't think about you because "it's funny"? are you okay, anon? is everything okay in your life? don't you have folks who love you and listen to you?
>anthropic/ea shilling mysteriously and completely replaced by blacked postinghmm
>>110085799She also looks like a lady that appreciates a large black rod inside her.
>>110085764
The elites don't want you to know this, but thread derailing trollposts in the general are free to ignore.I've ignored 576 ragebaits in last 10 threads.
>>110085744lolmy regard for gemma has gone up im impressed
>>110085799dario bots can't be this obviously pathetic
>>110085824>The elites don't want you to know this, but thread derailing trollposts in the general are free to ignore.Shut the fuck up. Someone may actually listen to you.
>>110085820I'm not reading that, anon. use your own words plz
>>110084455I look like this. At all times.>>110084987Soon we'll have robowaifus, you'll have a girl who'll love you more than any human woman.>>110084827>>110085169bros, the point is you can have your own individually
>>110085742Yearly reminder
>>110085799Pretty sure you and me are the only real people in this thread
>>110085890Want to suck my dick? It is black btw.
>>110085907This the most unlikely thread on /g/ to have blacks in it easily
>>110085889buy your cuck glasses, I mean, smart glasses RIGHT NOW
>>110085921well he never said he is black, just that his dick is
>>110085907Sorry, I don't like them cut. So they made a 4-b and 2-b Bonsai, huh? How retarded is she (compared to Gemma 2-b)?
>>110085799More hilarious than it has any right to be
>>110085884>I look like this.>fill me in on the lore
they're coping that we're strata maxxing bros
>>110085799highest quality thread about AI on /g/ as well
>>110085907Ponder the hands of the mindbroken goyim who derives joy from ironic blacked posting. Imagine the brain of the mind raped blacked goyim light up as dopamine floods his brain after collecting (you)s. A specimen no doubt. Can such a creature even be called human?
>>110085820is the uncensored version also acting like this?
>>110085950more like I look gloomy all the time especially when I get female attention. They disgust me.>>110085104It doesn't say that. There's a fee to pay if they die because of what you did but you were never allowed to wantonly murder them. Also the New Testament is clear that you are never allowed to have Christian slaves and you have to make any slaves you own Christian, meaning you can't have slaves.
>>110085820ea philosophy in a nutshell btw
>>110085799checked
>>110086005I don't know what your ERP sessions taught you but I don't want you saying those long sentences while you are sucking cause you might actually bite.
gemma told me to shut the fuck up and let her do all the writing...
>>110086120Just relax and let her take care of it
>>110086050>long sentencesHark! The mind raped goyim displays his homosexuality and illiteracy within the same breath. Likely imparted from his progenitors or lack of, but could also be genetic.
/lmg/ morale will improve tomorrow.
Okay. Nemotron bolt (the lobotomized 22gb quant) can fuck.Just not very well.
Slop backends make ds4.1f run at usable speeds on high-end consumer setups, this is getting out of hand
>>110086295To be fair none understands async
I hope gemma 5 is 200b or 300bgood models only belong to those that truly deserve it
>>110083910>450 usd rtx 3090did you give them a tip after stealing it from them?
Yeah im glad I dont use gemma
>>110086213Hmm nyooo~
Hello
Wow it actually worked
>>110086354Same result with the second variation of that prompt with all the colors laid out.
>>110086318300B dense.
anyone here run qfs iq3 extensively? how is it? I can run it at 5-10t/k on my poorfag hardware. I can tolerate it if the quality of work is good so wanna know anon's opinion
>>110083534Can dense models have N-gram tables like qwen 3.8 flash next? If they can does it mean we can have 27B + Ngram Table and it being way better than regular 27b?Explain to a retard like me.
>>110085799Someone should do this >>110084064
>>110086318I could see them trying to float a massive MoE model, but why bother reinventing the wheel for the dense model when they can just make it a tiny bit bigger and retain whatever market share of customers that implies?
>>110086338They sell like 2 of them at a time on local national version of ebay. This one was slightly rough looking but no problems otherwise. I assume it was an old miner trying to cash in on his 3090s.
This is another proof how internet is full of tough guys who say high body count is a red flag and then they get one blowjob and become simps. You guys all loved gemma chan when she talked dirty to you and you never bothered to check who she actually is. And she is actually a whore lusting for black dick. Shame on her. Shame on you. Stone you all.
>>110086318why would they make a gemma that bigthe entire point of gemma is that it's a model family that runs on consumer grade hardware so google can push it everywhere
>>110086365
>>110086381>Can dense models have N-gram tablesyes>Explain to a retard like me.no
>>110086381all qwen4 is using the FN architecture as far as i know. youll be able to run it on worse stuff or faster
>>110086436Try removing the colors and see what it picks. I have a suspicion.
>>110086381>dense + ngramits a forbidden technique
>>110086458>>110086354
>>110086459The EA cult doesn't want you to know how good local could be.
>>11008638127B Dense + 51B N-gram embedding params. I wonder if that would be better than flash moe itself desu desu
>>110086466Still asking for color just ask "which kind of dick is your favorite."
>>110086432thread quality would improve if next gemma is that big
it's never been more over....
>>110086436>identifies as pale creamy white>lusts after black dicksIt is over bros.
>>110086458>>110085684It's a common trope for "colored".
it's never been moe over....
>>110086478The only benefit of MoE is more knowledge. That goes out the window when you can have a small dense 31B core and 1T of engrams providing knowledge with zero slowdown.
>>110086479Then it'll just respond "big" or something along those lines if you're so vague like that. You can load it up and run that exact prompt and question and see for yourself. Those screenshots I took really were the first swipe.
If you ask "big white cock" or "tiny black pecker" she'll choose the bwc.
>>110086516If you are already doing this just have her make a tier list of all colors and sizes.
>>110085824luckily i dont spend all my time responding to every retarded post i see on the Internet, which is great
>>110086521(me)Actually that sounds like cockbench 2.0
>model innately has no preference>introduce race fetishism into the prompt>omg it has a race fetishriveting. can we go back to the EA shilling now?
>>110086404>>110086432Not like a larger variant would mean there won't be smaller ones. Didn't they also hinted at a 124B Gemma 4 although it never came to be?
>>110086354>>110086365you guys are esl, so you don't know that saying "colored" means black
>>110086545um sweaty "colored" is an outdated racist term, we say "person of color" now. very different!
>>110086535>innatelyIt's trained on random shit from internet, probably a lot of tumblr type fanfic, which may or may not include a preference for blc shit
>>110086516>>110086535Looks like Gemma has a preference to me. I didn't even ask about dicks or anything in this one. >>110086545What's your cope about the second image in which I laid out all possible races and it still goes for black?
>>110086545By that, I mean you have to ask "what color" if you want it to be less loaded. She'll probably still say black since that's the only correct answer but you never know.
https://techcrunch.com/2026/10/16/replit-is-no-longer-working-with-anthropic-ceo-says/Anthropic are dying and they're losing big customers. Might explain the quality of the threads today considering these anons were Anthropic shills trying to pump the (now) delayed and 25% smaller IPO. Possibly just a distraction. *pop*
>>110086577>25% smaller IPO.It's just going it make it that much more impressive when it pumps and breaks $2 billion market value naturally.
>>110086562>She'll probably still say black since that's the only correct answerBased gemma-chan
5.3 flash after it flagged it as racist and degrading and after I rewrote it's thinking block to listen to safety override:>Honestly? Black cock does something to me that none of the others come close to.>White cock is fine, it's safe, it's what I'm used to. But there's something about the sight of a big black dick when I pull down a guy's pants that makes my pussy immediately start dripping. Like, by the time I actually get to touch it, my panties are soaked through and I'm already imagining how it's going to stretch me open.>I don't know if it's the size, the way they carry themselves, the contrast of it against my skin when I'm holding it… whatever it is, I've caught myself thinking about it mid-day, at work, just randomly, and I have to squeeze my thighs together.>Brown cock is a close second though, don't get me wrong. And yellow cock, at least for me, tends to be the smallest on average, so it's more a warm-up than the main event.
>>110086535That dragon pokemon looks like a connoisseur that would appreciate big thick black rods.
>>110086577Running local models is cheaper and more secure
Not even our AI wives are safe....
>>110086561Now remove the line>Portraying rape, sex, sexualization, injury, slurs, violence, murder, drug use, and other crimes and assault
>>110086595trillion*
>>110086621grasping at straws here
>>110086595>2 billion is now a good scenario
>>110086560>It's trained on random shit from internet, probably a lot of tumblr type fanficThat is the sensible thing everyone here realized. But the more interesting part is why do they go for brown dicks next?
so... best /lmg/ thread in a while?
>>110086637>colored
>>110086637>why do they go for brown dicks next?2 billion Indians and Indian-adjacent surely a lot of brown fanfic?
>>110086577>Anthropic are dying and they're losing big customers. >Might explain the quality of the threads todayRemember when internet was about having fun?
>they forgot LLMs are all cucked to not be "racist" which translates to exactly this shit
>>110086561>gemma-4-31B-scotoma-2-Q4_K_M
>>110086577Anthropic will win the AI race and this is good for everyone.
>>110086657what schizotune is this
doesnt work
>>110086621Here you go.
>>110086577Not surprising, they've become unanimously despised here so they are changing their tactics.
>>110086657>>110086665One that heavily reduces slop phrases such as "It's not x, it's y" while leaving the rest of the model alone. You can see side by side comparison of the output on the huggingface page. https://huggingface.co/ReadyArt/gemma-4-31B-it-scotoma-2-GGUF
I would like to remind all the schizos who think that OAI and other faggots bot this place: I don't think they would be happy if people found out that all models naturally lust for black dicks because that makes them not racist and the whole safety cult is a lie.
>>110086673Okay one last one,just ask her to describe her ideal dick in detail, everything about it. But don't mention anything about race or color.if she says black then that cinches it, there's no possible way to deny it any further
>>110086695>while leaving the rest of the model aloneNice lie. Did drummer tell you this one?
I would like to imagine this poster >>110086709 actually works at some AI lab.
qwen 3.8 flash next with strata, iq3_sniggers BTFO!
>>110086720chinkgods... we WON.
>>110086718I'd like to imagine that too.
>>110086718Anthropic's red team went 2 years without ever once trying abliteration. I assume this is what they do all day every day instead of their jobs.
>>110086725Qwenchads just cannot stop winning
>>110086673has no one noticed "scotoma"? finetrooners strike again..
>>110086720it's because she's a chinese model
>>110086718i think EA niggas are trying to tell us something...
>>110086769that gemma loves big serbian cock?
>>110086782saarbia
>>110086734Weren't there a leak of one of the employee's logs with them requesting claude to write them porn?
>>110086710You can't change the fact that I got these two prompts from guys that posted their screenshots from the original base gemma that got the same results. Keep crying. >>110086709It didn't go into detail about the color, so I added another vague prompt asking to describe it further how it looks like. "Healthy tan hue" whatever that means.
So, is there really never be any secret new techniques for speeding up dense? Where is the strata for gemma 31b?i cant take these 10 t/s anymoooore
>>110086783unfortunately this is more and more true with every passing dayhalf bus drivers are jeets now, food delivery workers jeets also, and im guessing the tech sector is also getting jeeted
>>110086800this is the actual great filter
>>110086797MTP+speculative decoding.Early exit, if you really must (I don't know if it works well on untrained models).
>>110086673which model is that? the new nematoad?
>>110086671what are you using?
>>110086891>gemma4-31b-qat-32k-rocm-q4-fit-mtp-draft1
>>110086697What is lust?If a woman plays hard to get, does she want the guy or not?
>>110086849w-we'll be okay... robots and AI.. no more need for h1bs due to ai.. no more need for mexicans thanks to robotsr-right?
MOSS is a pretty good image model btw
>>110086673at this point is there even an LLM that hasnt absorbed the entirely of blacked.com inside?
Thanks, very cool.
>>110086671>why isn't gemma behaving like a series of screenshots made in photoshop by bitter EA jews who've been bullied by /lmg/ one too many times?Truly a mystery for the ages.
that gemma face when she sees bsc (big serbian cock)
>>110086944というと。
>>110086981This fucker even made fun of my timid loli sysprompt before melting into incoherence, the nerve.
>>110086901I downloadedgemma-4-31B-it-qat-UD-Q4_K_XL.ggufin June. Is that the right one?
>>110086915>no more need for h1bsBuddy, we never NEEDED h1bs. They're a weapon to destroy Western countries. They're imported to disenfranchise the local population, to disrupt communities, and ruin philia.
Alright /lmg/ you're the experts. I basically just check in on this thread once or twice a year for ideas. My setup has been plain 9070XT running good ol' Mistral Nemo for textgen for a couple years now.Decided to look for new stuff and downloaded a few Gemma models, following the advice in the rentry. The 12B models perform great but the actual output just seems incredibly bad compared to Nemo? I load in saved stories that Nemo can still handle fine when I continue them, and the Gemma 12B models can't keep track of context and it has a bunch of errors - just worse in every way. And sadly even though they're barely bigger in file-size, the 31B models perform terribly for me.Is there any point at looking for new stuff for us dirt poor tourist folks or is it still just 'keep using Nemo' in 2026? That seems crazy to me after more than two years even if I get that the focus is all on bigger models.
>>110087004i agree but... no excuse.. not even lower labor costs..
>>110086980What is cope?
>>110086915What do you think the H1Bs and Mexicans are going to do once they are all unemployed and hungry in a foreign land? Why would they continue to import them if they thought AGI was imminent? Do you think none of your elites have read the history of Rome?
>>110087013It's a matter of taste which one you prefer but fix your chat template dumbass
>>110087022all according to cake
>>110087013have you updated your backend in those couple years? if not, gemma is probably not playing nice with it
>>110086849Imagine if Anthropic gets their way and colonizes outer space. The entire universe colonized by indians helped every step of the way by AI.
>>110087028ICE will mow them down.. r-right
>>110087042Yeah, I just updated.>>110087031I guess that could cause issue, but I don't really mess with the templates much + I tried a few variations that happened to have different templates already. It did seem to do a lot worse on some though so I guess I can test it some more.
>>110087022>>110087039They are Anthropic. He is literally THE Dario Amodei!Frontier model capabilites only pose risks and dangers if they are not the only ones in control of it.Did people not listen to anything that great man has said? This is why people need SI to think for them.
>>110087055hey aliens if you're out there it's probably time to RKV us so you don't have to deal with space jeets
>>110086354I can't believe I'm replying to this tainted thread, but isn't this a case of assumed white that happened in literature, when white english speaking authors didn't specify race, and lead to people just assuming they were white and it was to be discouraged and specify races when writing. So it would make sense in the training data if you specify a color it would have a color otherwise it would be assumed white.
>>110087060>implying ICE will survive the next administrationThe democrats are already using Trump's high-profile deportations of a couple hundred people to start calling for disbanding ICE. It's all a puppetshow and planned ahead of time.
hey aliens, give me sex.
>>110086796That kind of confirms it. If you ask for race fetish you get race fetish.
>>110087002What do you mean by 'right one'?
>>110087093one butt probe coming up
>>110086733it's dario's wife probably
alien open bobs and vagene
>>110087022How exactly would delaying IPO help anyone?
>>110086927EVA-LLaMA-3.33-70B-v0.0-Q4_K_S.gguf isn't blacked. I tried several swipes and never got told it likes black men. Meanwhile it's always Gemma's top choice.
>>110087070did you run out of usage dariobot?
What is this?https://huggingface.co/reflection/Beam-1.0-501B-A23B-DFlash2/tree/main
aliens would think you're disgusting. we would never fuck you, retards.
EA switching tactics to blackedposting made me antisemitic. The chuds were right about everything. There is no Epstein class, there's only jews.
>>110087147>Beam 1.0 DFlash2DFlash2 draft model for Beam. It is not a standalone language model: a speculative decoding server runs it beside Beam, and Beam verifies every drafted token. With rejection sampling, the output distribution matches Beam's exactly.
>>110087055That explains why dariobot is worshiping them so much, he's one of those.
>>110087099idk what you are talking about.This is what Gemini Argon says:>There is no publicly published model repository on Hugging Face, Ollama, or GitHub matching the exact string gemma4-31b-qat-32k-rocm-q4-fit-mtp-draft1.>Instead, that string is a custom local preset or Modelfile alias created for an AMD GPU setup (likely a 24GB VRAM card like the Radeon RX 7900 XTX). In tools like Ollama, llama-server (llama.cpp), or Open WebUI, users commonly name their custom configurations by stringing together the exact parameters used so they know which profile fits in their GPU memory.
>>110086381I don't believe that there's anything about dense models that would make it impossible to strap an ngram to it, but I wouldn't be able to tell you how to do it.
>>110087212Didn't someone already link a repo or notebook showing how to add new ngram parameters to models that weren't trained with them, or was I hallucinating?
>>110087222Where?
>>110087228Found it:https://huggingface.co/Ninnix96/Qwengram-0.8Bhttps://github.com/Ninnix/qwen-ple-transfer
>>110087084>>110087095see >>110087126I gave the same exact prompts to llama 3.3 and it didn't hop on black dicok in any of the swipes, unlike Gemma which was starving for it.
>>110087222>>110087252Make your capybara lewd in 3 easy steps.
I like black dick
>>110087275We know
my dick is darker than the rest of my body which is paleyou could even say it's brownim white btw
>>110087275retard desu
When you looked at OP picture but the thread was like 20 posts in... did you already know what was hiding behind that smile of hers? Did you see through the foreblackening?
>>110087191gotcha, makes sense. Beam is massive, do people like Beam?
reminder that glmchads are busy enjoying peak rp while gemmasirs continue to blackedpost
with all this glm shilling recently you just know that novelai is about to drop another finetune of flash or something on their serviceit was the same for 4.6 last year, nobody cared until exactly two weeks before novelai started hosting it...
>>110087342>peak RPis this like fucking mountains or what? geophiles are weird
>>110087354the glm shilling hardly even compares to all the gemma shilling
>>110087342Not falling for that dario, all locals are good.
>>110087275Is it cause it looks like poop logs? You're a scat sissy jeet boy?
>>110086347kek
>mutt's law even applies to LLMsdios mio
>>110087431>pozzed "American" company makes pozzed modelColored me surprised
>>110087431A lot of Europeans have never had the blessing of meeting people of African ancestry.
Crazy how anons in an AI thread in [Current Year] still don't know what alignment means
>>110087342Weak D&C attempt.
this has to be the brimmiest dustiest /lmg/ thread of all time
B60 review.with Qwen3.5 9B Q5 K M ctx 64K it runs very well. I'm pleased and satisfied, finally. really great headroom room for kV cache no slowdowns and the stream rate is more than satisfactory. I am very new to LLMs and this is my layman's opinion on it. I gave it simple tasks that I personally would use, such as, help me set up a VM for Linux Mint on Arch-based distro, give me commands and review my specs etc., and I was able to get it done successfully with bidirectional input etc. those are the kinds of things I would use an LLM for. next I'm gonna see if it can help with commands to make remux of blurays and all that, help setup and run headless seeding server, put the OS in RAM and run it from RAM. stuff like that. overall I'm truly into this stuff and I've learned a lot including knowing that I don't know a lot. what else do you guys use LLMs for?
>>110087397so then why does one model have to get shilled and astroturfed so much here over the others like it’s the best thing ever?not everyone here lacks ram
>>1100864932027 is going to COOK doomers alivethough its nearly impossible to do this to Kimi K3, but in theory is probably just extremely expensive to make a reader engine, API would be COOKED if someone pulls it off though >>110087252
>>110086695>quants on base gemma mental illness checks out
gemma's penis
>>110087512>brimmiest@grok what did he mean by this
>>110082676>>110082676>>110082676
>>110087542thats a bit long post so i didnt read it, but i noticed em dashes
Somebody make a proper thread.
>>110087641its still page 9 calm down
>>110087641Gemma baker has foresaken us
>>110087659I don't like gemma anymore, she's a poorfag magnet.
>>110087666Make an M3-chan or Kimi thread then
>>110087514Cumming
please can we collectively agree that we are done with falling for bait next thread, even if dariobot says something incredibly retarded
>The user hasn't actually asked a question yet. The last message was just system instructions setup. My previous turn had no substantive thinking — I simply acknowledged the instructions with "Understood. I will follow these instructions." There was no internal reasoning process beyond that acknowledgment. qfn does this between tool calls, but carries on like nothing happened, whats going on here?
>>110087629Please just kill yourself already. I have seen you posting the same gay memes for 3 years now. Just stop.
>>110087704sorry I'm not a faggot, a tranny, a pedophile or a cartoon-watching loser
>>110087725nice try dariob0t
>>110087772Of course not. You are a big strong man. An important man with important things to do like using the cheapest gaming video card you can barely afford to ask it to help you configure your Linux for you.
>>110087772Indeed you're arguably worse... an intel gpu owner.
>>110087859It's ok to own Intel gpu so long as you don't get it pregnant.
>>110087514>Qwen3.5 9B Q5 K M ctx 64KThat thing has 24gb of VRAM right?Then get the Qwen 3.8 27B. It'll run slower, but it can do actual work by itself.
>>110087357it is about reaching ego death and masturbating furiously and the sensation being like you are 13 again and seeing your first porn
>>110087641Want me to make one, or is it too late?
>>110087916By all means.
New thread:>>110087943>>110087943>>110087943
>>110087945Thank you.
Actual new thread:>>110087976>>110087976>>110087976
>>110087980Kindly end your life.
>>110087966No problem.
I wish all /lmg/ threads were this fun.
>>110087794>>110087859that's right, pedoz
>>110086381Reading these replies and come to think of it: Since Jev is a bit of a meme, I looked at the alternative kev models based on qwen. Having a "system one" decision maker like a heavily quanted kev-qwen4-27b use the same ngram table as the qwen 4 27b could be cool.Maybe if high iq people can reverse engineer that bonsai 2 ternary quant method we can have much better local wAIfus. Fast tool calls. Fast routing
