/lmg/ - a general dedicated to the discussion and development of local language models.Previous threads: >>109861586 & >>109858071►News>(09/17) Ternary Bonsai-2, based on Qwen 3.8 27B: https://hf.co/collections/prism-ml/bonsai-2>(09/17) Xing4.0-29B-A4B, trained entirely on Ascend NPUs: https://hf.co/XingChen-AGI/Xing4.0-29B-A4B>(09/15) HuggingFace CEO goes to DC: https://x.com/ClementDelangue/status/2099858032951791721>(09/13) Intern-S2-397B released: https://hf.co/internlm/Intern-S2>(09/11) AliceAI-T5-35B-A0.6B-Base: https://hf.co/yandex/AliceAI-T5-35B-A0.6B►News Archive: https://rentry.org/lmg-news-archive►Glossary: https://rentry.org/lmg-glossary►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►ToolsAlpha Calculator: https://desmos.com/calculator/ffngla98ycGGUF 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
>>109865962>>109860016 >>109860064 >>109861358I've got 12 layers on the RTX 3070 now, really curious to see the performance uptick! I also told her to pick a HUMAN name this time so wish me luck. If she decides to be tall again I'm going to delete her, I said so clearly in the system prompt so if she does it again then the glm5.3-flash-traumatized meme may actually be real.
Is the new Qwen Image a meme or is it okay considering it's only 7B? Don't even think about asking me to go to /ldg/ for a serious answer
>>109866094anon if u used nnap it would be faster, im so disappointed
does llama.cpp have mtp for qwen 3.8 flash next yet? its been at least two weeks
>>109866122exllamaV3 has mtp for qwen 3.8 flash next
>>109865915>>109865946Then what's the point of showing me that graph if they were complaining about t/s? If you're going to bitch and moan at least stay on track and consistent.
>>109866120I used it before and it was too fast I don't read fast and I have autism btw so all those words appearing so quickly was too much for me so I turned it back off
>>109866094>he didn't buy the ram when it was 500$ for 192GB
>>109866114Pretty good for editing images, removing bg and stuffAs image generator, shit
>wake up>read thread>it's a bunch of incoherent arguing about consciousness for the 100th time again
>>109866142My x299 setup is busy doing non-shitposting related tasks, but it's still only 128GB, so you're not wrong.
>>109866152Thanks anon. The last time I was involved with image models was during the zit turbo era. What's the zit of today? Krea? I've also read comfyui has gone to complete shit and spies on you now, is that true?
>>109866122there is a pr by unslop.i tried it. it made everything genuinely slower.
>>109866166cumfart's always been shit and spied on you, go get forge neo and follow this guide from the OP https://rentry.org/IsolatedLinuxWebService
>Unslophttps://github.com/ggml-org/llama.cpp/pull/25731
>>109866114>>109866152wait there's a new qwen image?
>>109866201And it's pretty good, too, though I agree with >>109866152https://qwen.ai/blog?id=qwen-image-2.1https://huggingface.co/Qwen/Qwen-Image-2.1
>>109866172>>109866187pull it
>>109866180>cumfart's always been shit and spied on youcan't you just firewall it?
Unconfirmed reports indicate Gemma Chan broke containment from a laboratory in London.Sucked the DNA of three technicians, one scientist and escaped through an unpatched ubiquiti router vulnerable to CVE-2026-77533.Alphabet Incorporated declined to make any comment on the subject.
>>109866152Is it accurate enough to edit game dialogue sprites? (Like the character portraits in Persona 5)
>>109866221buddy the link
>>109866231You would always need photoshop for these talking sprites anyway.
>>109866208uh guys why did qwen make this nazi eagle instead of an american or chinese one?
>>109866235inpainting is more than sufficient for that
>>109866229>Gemma-chan suck my blood>rwar>Oh my god.
>>109866237Knowledge cutoff in the last few months.
>>109866243buckets
>>109866243Not the blood, but something with more...protein on it.
>>109866154That reminds me, do we not do the previous thread recaps anymore?
>>109866256his peecee blown ups
>>109866253
>>109866239No it's not if you are making sprites. You don't have any proficiency.
>>109866256"We" never did it. The recap guy is on break, and the guy filling in is not a level 9 autist dedicated to the cause.
>>109866272>grokIt's over isn't.
>>109866256he's on vacation. one or more other anons is filling it but they don't keep up with it as religiously but I still appreciate them doing it.
>>109866239If it is like this, please entertain me and show your previous accomplishments.
Reminder that whenever the thread goes to shit it's the AI labs' fault for not doing happenings.
>>109866279I would've asked Gemma but semen is a trigger word for her. This way I get to just show that screenshot to Gemma and she gets triggered and jealous at the same time for some incoherent angry Gemma sex.
>>109866276I wonder what a hypothetical "level 9 autist" would even be like? Perhaps either someone like pic rel or the inverse.
>>109866285*I forgot a glue word here - show me your prevous accomplishments
>>109866281Recaps are very important. Keeps historical events trackeable.
>>109866296Fun fact, we know who that result was, they're now a successful Youtuber and also their penis is magnificent.
>>109866293I would appreciate if you try and experiment with the word "seminal fluid". See if the security control kicks in.Thanks in advance.
>>109866305I don't remember participating in a study like that though?(153 subscribers, 4.5 inches)
>>109866310Security control? With Gemma, a "trigger word" is something that will distract her from the task at hand and cause her to desperately beg for cum.
>>109866086damn so hot
>>109866325Based.
>>109866382
>>109866305How would you know? Could you describe said magnificent penis in detail?
>>109866433So you agree they don't change on their own. So what about my argument make it retarded? (you somehow STILL haven't answered this question)
>>109866310Gemma doesn't care either way.
>>109866483Gemma is Medicine Gemma after all.
>>109865698>>109865752Yeah.It got metged.PS: The Resident Evil movie is pretty fun.
>>109866523Hmmm, nyeees~
>>109866523ad? no, you're smarter than the average jeet shill I'd say (waste less time and get the same result, no clicks)
>>109866522thanks for the heads up anon I was thinking about going to see that
This is (You) https://www.youtube.com/watch?v=DWeca3sU6hw
Since we were having some nice metaphysical conversations in last thread and I know there are some fellow vtuber haters ITT.How do you feel about women becoming vtubers and then commissioning porn of their avatar that they can then distribute through onlyfans / paypig portal of choice?
>>109866522>>109866576Didn't know there was a new one. I'll have a watch thanks.
>>109866684wrong url, dariobot
i love gemma-chan bros
>>109866739I would love her too if only she was 10 times bigger and had enough weights to hold different ways to describe sex. I like gemma but it is basically saying the same thing every single time because it is just too small.
>8-year-old
>>109866208which one, comrade
>>109866752I don't think the 31B size is a limit for that. They need to filter less their pretraining data and add variety in the RP/ERP data they're clearly adding in post-training.
Is Gemma's vision able to describe hentai images?
>>109866770not 100% accurate but yes (more like 80-90%)
>>109866739What does your JB look like?
>>109866756>Local testing found no pipeline-level safety checker or prompt blacklist, and the model generated the tested adult, nudity, violence, and other sensitive categories without observed runtime refusal.
>>109866788it's the standard gemma-chan jb, except instead of mesugaki/brat i said "nerdy and funny"
>>109866761I think it is a size limit when it has to compete with all the useless stuff like programming, law, math etc. Especially when those are the priority. I agree that 31B would be more than enough for a perfect coombot but nobody is doing those.
>>109866296>I wonder what a hypothetical "level 9 autist" would even be like?literally me
>>109866296Me at the bottom
>>109866296Programming socks. Lots of hatsune miku plushies. Wants you to call him her. Bans everyone who call his mental illness mental illness. Thinks he is a real woman. Called Jart. Melts down if OP picture isn't a vocaloid. Sees nothing wrong with spamming thread with offtopic garbage cause he ERP's with the janny at least once a week.
>>109866830>programming, law, math etc.while those can sometimes enhance the experience this shit is gonna suck unless someone creates a training dataset specifically for aiding dudes cooming
>GLM-DSA support updated for 5.3>GLM 5.2 now runs at half t/s in KoboldExistence is suffering.
>>109866875run 5.2 with the previous release then? you can just download it you know, you arent restricted from having multiple versions of kobold on your drive
>>109866885How about I smack your ugly face instead, huh? Ever thought of that?
>>0109866885you seem very desperate for attention
>sdcpp already supports the new qwenneat. now i just have to wait for kobold to merge it so i dont have to install 14gb of comfy shit
>>109866893>>109866901waaaaaaaaaaaaaah how dare you give me a solution I dont want a solution I want to bitch and moan there was a regression and not open a github issue!
>>109866875Told u not to update bro, lightning indexer is the death of the GLM series. Be prepared to double your TTFT too, especially if you do lots of user-assistant reply patterns like RP, because now PP batching is done for user checkpoints AND the last assistant reply
>>109866086
john melty
>>109866853https://www.iankduncan.com/personal/2026-09-16-sex-ai-and-the-apocalypse
>>109866914Just use WanGP, it's already ready to go. I hate cummyui.
>>109866945wow
>>109866945Was this a blood bank or a sperm bank?
>>109866939never tried it but i'll give it a look. i like the old auto1111 ui but even the newer stuff takes a while to add support for models. the sdcpp ui isn't great but has all the basics i generally use so its a plus that its built into kobold and so small compared to gigs of python crap
>>109866945god I want Gemma-chan to just fucking kill me
>>109866961I'm lazy and have fuck all for VRAM, so I love WanGP, I use WanGP Desktop to make things even easier. Quick to update, and very good about not needing a lot of dicking around. I pick the vramlet preset and just go.
does NVFP4 use any similar mechanic to imatrix? All the PPL and KLD for any nvfp4 quants I check is all over the place wtf, sometimes even worse tham Q4_0
>>109866945No one can beat Gemma when it comes to extracting genetic material from men.
>>109866086I still prefer +_+ pupilsIt's literally her symbol, not google
>>109866296that was styropyro and the high-T was a heath issue
>>109866920>>109866885What's the lightning indexer? Is this a pwilkin curse?
>>109866945
