Autism Egregore

@lostboys @TeksEdge A dumb model going fast would be useless to me.

@AutismEgregore @TeksEdge ?

🎗️ "Medium-Sized" LLM Burners Coming Soon! 🔥 This Could Make Local HyperToken Generation a Reality. ⚡️ NVIDIA’s worst nightmare? 😱 ⚙️ Application-Specific Hardware Taalas new PCIe ASIC board would burn the entire medium-sized Qwen 3.5-27B LLM straight into silicon 🤯 (already doing it with small models) Taalos said medium models on ASIC would be available in their lab by Spring '26. 💭Imagine: 🚫 No more loading weights 🚀 ~10,000 Tokens Per Second locally (Llama 3.1 8B already @ 17,000 tps) 💻 Standard PC slot, ultra-low power (10x less) 🔋 🌍 100% offline with no cloud, no GPU farm 💰 Reddit unit cost rumor $300 to $400 🖥️ Imagine HyperToken generation on your desktop. 🤖 AI agents that think at light speed. ⚡️ Are you ready? 👀

@lostboys @TeksEdge No, it definitely would not. I can generate token faster than that with a python script, but it's just going to be the letter "A" forever.

@AutismEgregore @TeksEdge ye its not a criticism by any means, its awesome. id like to see the tradeoff between intelligence/speed for the dumber small models. would being able to process at 15kt/s outweigh the intelligence of running kimik2.5 at 15t/s?

@stevibe Any chance of testing the abliterated, heretic, and aggressive versions?

Qwen3.5-27B went 15/15 on our tool-calling benchmark. But which quant should you actually run? Tested Unsloth's Q2_K_XL all the way to Q8_K_XL TL;DR: Q8 — 15/15 ✅ Q6 — 15/15 ✅ Q5 — 14/15 Q4 — 14/15 Q3 — 14/15 Q2 — 13/15 Q6 is the sweet spot. Same perfect score as Q8, smaller footprint. Also, the results scale almost linearly, seems like ToolCall-15 is actually measuring something real.

@lostboys @TeksEdge Don't care. If they do qwen3.5-27b-abliterated, I would pay an arm and a leg for one of those cards.

@TeksEdge no more loading weights because you can only run one model forever. this is completely incompatible with MoE btw

@twlvone @victormustar Adding "the real unlock" to my blocked phrases.

@victormustar text to motion for robots is the real unlock here. programming movement by describing it in english

NVIDIA's Kimodo is the release of the week 🔥 Prompt the timeline whatever your want like: "a person walks forward" → "a person starts jumping", hit Generate, and watch a 3D character do it in seconds (700hrs of pro mocap training. Works on human + robot skeletons. Super fast + free to use on HF)

@wildmindai ITT: anti-AI cope

You never actually had an anonymous Reddit account. You just had a digital footprint that was too expensive for a human to piece together... - your fake username means nothing to an LLM - it reads years of casual comments in seconds - it extracts your city, your job, your minor complaints - it builds a unique psychological and demographic fingerprint - then... fingerprint + LinkedIn - what used to take a private investigator days now costs a few dollars no more illusion of online obscurity arxiv.org/abs/2602.16800

@TheZachMueller @UnslothAI For abliterated models, does the obliteration happen before or after the model is quantized?

PinchBench results for Qwen3.5 27B using @UnslothAI K_XL quants, best of 3, thinking disabled. TL;DR: Q3 KXL (14.5GB) easily In a shocking twist (thanks to randomness and averaging), Q3 was a top performer when doing non-thinking, which does mean good speed + memory! BUT 👇

@SlipperyGem tbf, those poses are blipping in and out of existence pretty sporadically

Look like LTX 2.3 it's still not strong enough in the consistency department.

@wildmindai Is TRELLIS2 still the top model for image to 3D? Can it accept multiple angles at once?

SegviGen. Repurposing 3D generative models for part segmentation. Looks cool. Easily select, isolate, or modify any part of a 3D model. -TRELLIS2 with SC-VAE. - tops P3-SAM by 40% fenghora.github.io/SegviGen-Page/

@zhengyiluo @br

288 hours of high-quality, text-annotated human motion data are now available! 140k motion sequences! Do you know that a large part of SONIC's training data is now open-sourced? Check out the dataset here 👇🏻 from our friends at Bones Studio! Full human + G1 retargeted motion! Stie🌐:bones.studio/datasets/seed Data💿:huggingface.co/datasets/bones… SONIC training code coming VERY VERY soon!

@william_weimeng @xiaohu With enough ram you should be able to offload layers

@xiaohu 5060ti 16G 能跑吗

Qwen3.5 去审查版来了 0拒绝 4090就能跑 有人把 Qwen3.5-35B-A3B 的安全拒绝机制给拆了，做了一个完全不拒绝的版本。 测了465个通常会被模型拒绝的提示词，拒绝次数：0 而且这还是个 Aggressive（激进版），意思是完全解锁，不留任何安全护栏。 支持文本、图片、视频多模态 原生 262K 上下文，可扩展到 100 万 支持 201 种语言

@SlipperyGem Gonna need more of you to start replying to Brie's requests for information on models. I stopped my NEET life and can no longer contribute and test things like I used to.

Uncensored version of MM-Audio for all your squishy noise generation needs. This same author also has a Hunyuan Foley uncen model. I've never ran Hunyuan Foley, but its there if that strikes your fancy. (Any of you run Hunyuan Foley? How is it?) huggingface.co/phazei/NSFW_MM…

@jtydhr88 Pulling frames from a WAN rotation LoRA seems like it would yield better (although less precise) results.

Tried to recreate PS’s image rotation feature inside ComfyUI - 2

@ltx_model Can it do cartoons?

LTX-2.3 is a clear upgrade from LTX-2. The improvements translate into more stable motion and better detail retention in complex scenes and cleaner audio output. See for yourself ⬇️

@SlipperyGem @realrebelai To a 4090. Last I remember you were at 12 or 16 GB vram.

@AutismEgregore @realrebelai To LTX 2.3? Yes, definitely.

After FURTHER further testing, i revoke my claim about the GGUFs for LTX-2.3 being better than the fp8. Too many artifacts, hallucinations, and weird glitch frames. Not worth the lower compute or gen time, stick with the fp8 🫡

Meet LTX-2.3-Workflows: a powerful image-to-video AI model that's buzzing in the community. It takes any static image and brings it to life with motion. Think of it as giving your pictures a soul. This is the next frontier in generative AI.

@SlipperyGem @realrebelai You finally upgraded?

@realrebelai Which one by Kijai do you recommend for 4090 users?

@KobertKirk @toyxyz3 I'll give it a shot

@AutismEgregore @toyxyz3 --chat-template-kwargs "{\"enable_thinking\":false}"

@Chris_Linux @toyxyz3 Tried that one and it will sometimes just "think" outside of whatever llama.cpp recognizes as the reasoning part.

@AutismEgregore @toyxyz3 and append --reasoning-budget 0