/lmg/ - a general dedicated to the discussion and development of local language models.Previous threads: >>107700909 & >>107686942►News>(12/29) WeDLM-8B-Instruct diffusion language model released: https://hf.co/tencent/WeDLM-8B-Instruct>(12/29) Llama-3.3-8B-Instruct weights leaked: https://hf.co/allura-forge/Llama-3.3-8B-Instruct>(12/26) MiniMax-M2.1 released: https://minimax.io/news/minimax-m21>(12/22) GLM-4.7: Advancing the Coding Capability: https://z.ai/blog/glm-4.7>(12/17) Introducing Meta Segment Anything Model Audio: https://ai.meta.com/samaudio►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://livecodebench.github.io/gso.htmlContext Length: https://github.com/adobe-research/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-sampling►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/vllm
►Recent Highlights from the Previous Thread: >>107700909--Modern setup strategies for real-time knowledge access beyond static model training:>107707804 >107707936 >107707959 >107707983 >107707985 >107707990 >107708011 >107708020 >107708035 >107708037 >107708708 >107709078 >107708058--Multi-character story challenges with Mistral 24B models:>107707485 >107707507 >107707600 >107707883 >107707948 >107707670 >107707718 >107707771--Quantization challenges for running GLM 4.6 on limited VRAM:>107705394 >107705411 >107705425 >107705450 >107705516--Evaluating 4.7 AI model's artistic adherence and natural dialogue vs 4.6:>107705364 >107706320 >107708117 >107708121--FunAudio-Chat Technical Report:>107708791 >107709016 >107709079--Resolving assistant response prefill incompatibility with enable_thinking:>107702566 >107702587 >107702629--Google's early 2000s chatbot experiment with knowledge reuse:>107705377 >107705409 >107705424--Updating software version fixed launch error for GLM-4.5-Air-UD-Q2_K_XL:>107702400 >107702426 >107702428 >107702530--Critique of model thinking processes and their impact on response quality:>107703015 >107703056 >107703071 >107703094 >107703119 >107703268--Exploring local voice cloning alternatives to SoVits:>107704130 >107704193 >107704277 >107704319 >107704453 >107704482 >107704829--Mixed performance and limitations with Minimax at IQ2_M quantization:>107702412 >107703627 >107703661 >107703732--Z AI's IPO implications for the AI-native LLM market and competing models:>107708784 >107709044--WeDLM-8B-Instruct release and comparison to Qwen3-8B:>107709163--Miku (free space):>107701017 >107701268 >107701433 >107701631 >107701715 >107704951 >107707361 >107708317 >107708548►Recent Highlight Posts from the Previous Thread: >>107700912Why?: >>102478518Enable Links: https://rentry.org/lmg-recap-script
>>107709264>>107709248kek
The joke went too far
>>107709282I think you are confused or very much clueless. These are just regex filters you mongoloid.
Is llama.cpp broken again? Gemma's more retarded than usual.
>>107707382thank you for bringing this to my attention
So.. when are we getting something?
Mistral my beloved
>>107709593https://huggingface.co/bartowski/allura-forge_Llama-3.3-8B-Instruct-GGUF
>>107709613The blandest and most mid LLMs on the market, only worth using because of their lack of strict guardrails.
>>107709653The Honda of LLMs. Nothing fancy but gets the job done.
>107709657can this faggot get out of my thread?
>>107709248I dont get this new meme
>>107709664Gone.
>>107709670are you a janny? if so, thank you.
>>107709672No. But we all can do things we're not supposed to advertise.
>>107709666A few threads back a few people tried generating miku using glm or whatever and almost every time she looked bald because she only had twintails or the hair was drawn too low so her head was poking out.
>>107709672
>>107709679right. i also may or may not have done that thing. never seen such a fast response time
>>107709683come on now. some of them work really hard. i even pay their salaries!
Well. That was quick.>>107709685Yeah. And the funny guy that just joined got donned.
teto my baldloved
>>107709692
how do i stop destroying my keyboard while waitin for my ai's responses
>>107709264>>107709248>>107709259wew ladthread theme: https://www.youtube.com/watch?v=423Nmfpo828
>>107709628It was released on the Meta API in April, but I bet it was trained about at the same time as Llama 3.3 70B; who knows why they didn't release a smaller model back then. So it's probably a year old, at this point.
>>107709691you havent been paying long enough, paypiggieGLM AIR WHENGEMMY 4 WHEN??
>schizoids goes rampantI blame bald migu
>>107709725Aim away from the keyboard.
llama 3.3 cockbench where?
>>107709781dunno about the samplers but i grabbed the cockbench paragraph from https://desuarchive.org/g/thread/105354556/#105354924q8
ahahahaha llama 3.3 7b? more like ollama deepseek-r1
>>107709259The recap missed the most interesting conversation from the last thread.
>>107709813>dunno about the samplersAlways greedy.
>>107709743I think they hate us. They released only large models that work on enthusiast systems in the last round. Meta are true prog believers but zuck is a fickle suckup to whoever is in power. Tuning a model that people liked on arena and then uploading cuckmaxxed weights is absolutely something else. Remember how they gimped their omni model despite there being way better image gen and text already out there. Who the fuck even does that?
>>107709880lmao it's MMLUmaxxed
https://huggingface.co/upstage/Solar-Open-100B 2mwh
>>107709902so its gonna be a shitty glm air clone basically?
>>107709894>mom is in another room entirely>We're caught.>they just had lunch>Dinner's ready!Where is the anon claiming that dense models have better understanding?
>>107709908its gonna be fimbulvetr sexo but air intelligencereminder they released solar-10.7b and that was the go-to SEX modelwe are going to be so back.>>107709913its 8b anon...
>>107709913ah yes a year old butchered 8b is representative of all dense models
>>107709919oh was that them? i was wondering what these random 10.7b moe abominations were.https://huggingface.co/tensorblock/SOLARC-MOE-10.7Bx6-GGUF
>>107709919>>107709922Llama 3.3 only exists as 70B.
>>107709934https://huggingface.co/allura-forge/Llama-3.3-8B-Instruct
>>107709934>reading comprehension of a moe...
Got tired of seeing her bald head
>>107709978Hair is stored in the ears, prove me wrong!
>>107709882>Who the fuck even does that?Management scared of collecting even more lawsuits. You can't release anything good if your primary concern is not to get sued because you are hated by pretty much everybody.I bet they just couldn't make Llama 4 both good and "safe" (according to their own internal parameters) at the same time, and so it got butchered before release with poor results. That, and not providing smaller versions for the local LLM community (even though ironically the current best Chinese MoE models are Llama 4-sized or larger) killed their reputation and in the end their open LLM efforts.
>>107709987
>>107710043No model creators or huggingface got sued yet. They were much more likely to get hit for copyright and yet ignored that. It had to be ideologues.
>>107709978>>107710051so much glaze
>>107710043What I don't understand is why the "safety" debate even exists. Nobody is suing home depot because some wannabe terrorist was able to buy ingredients for home made explosives there, so why do we have this whole fake debate about "security" for LLMs?
>>107710088refer to thine digits for a clue
>>107710088because yud unironic
>>107710088
>>107710088homedepot did remove potassium nitrate stump remover for one that doesn't work. so they do cuck all of our products when it comes down to it. also lawnmower blades sold unsharpened only.
>>107710088Just try going to a Home Depot and asking an employee to recommend the best products and methods for making a nice pipe bomb.
>>107710108He's literally worse than a faggot
>>107710088>our product is so powerful, it might even destroy the whole world if we're not careful (so we definitely can't let you download the weights, but we'll still sell you API access)
>>107709978can she grow hair long and sentient enough to twist and tie itself into twin drills?
>>107710051why does her hair change color when she moves her arms? what happens when she waves?
>>107710426>whats blushing
someone make a model only trained on the king james bible
>>107710463Already done
>>107710444hair is not supposed to blush
>>107710605meds
ai can be very educational. i learn so many useful things from ai
>>107710745Pretty funny!
