Local AI News You Missed - October 2026
extract
Here's what you probably missed in October 2026: **🧠 LLMs** 1. [**Ornith-1.5-122B**](https://huggingface.co/ornith-ai/Ornith-1.5-122B) - Handles million-token agent tasks on a single GPU. 2. [**Raon-1-mini-preview**](https://huggingface.co/KRAFTON/Raon-1-mini-preview) - Runs Korean agentic AI on a single RTX 5090. 3. [**Timestral-209M**](https://huggingface.co/mistralai/Timestral-209M) - Forecasts time series in one pass with 21 quantiles. 4. [**Carbon-X**](https://huggingface.co/HuggingFaceBio/Carbon-X) - Reads DNA in chat and predicts variant effects on one GPU. 5. [**Sage-A1.5**](https://huggingface.co/InternScience/Sage-A1.5) - Orchestrates scientific agents with self-review. 6. [**Spark-X2.6-4B**](https://huggingface.co/XHToken/Spark-X2.6-4B) - Advances on-device agents with controllable thinking. 7. [**Nanbeige4.3-33B**](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) - Teacher model leads agentic coding and cowork. 8. [**Flint**](https://huggingface.co/RecursiveSI/flint) - MoE coding agent investigates repos and fixes bugs on one GPU. 9. [**Hy-Memory-30B-A3B**](https://huggingface.co/tencent/Hy-Memory-30B-A3B) - Writes durable agent memory at 3B active parameters. 10. [**Underdog-Saluki-27B-1.0**](https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0) - Packs Qwen3.8 tool calling into under 8 GB. 11. [**D3**](https://huggingface.co/vllm-sr/d3) - Answers choice, yes-no, and score questions in one call. 12. [**LFM2.5-24B-A2B**](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B) - Packs fast on-device agent reasoning into 2B active. 13. [**Qwen3.5-2B-RCOL**](https://huggingface.co/trubisky/Qwen3.5-2B-RCOL) - Low-bit quants hold quality under 1 GB. 14. [**GLM-5.3-UNCENSORED-EXL3-3.0bpw**](https://huggingface.co/Infatoshi/GLM-5.3-UNCENSORED-EXL3-3.0bpw) - Fits a 753B MoE in 273 GB. 15. [**Humanizer**](https://huggingface.co/jialinyyzz/humanizer) - Rewrites AI drafts to sound human in English and Chinese. 16. [**Kolibri 1**](https://huggingface.co/Aleph-Alpha/Kolibri-1) - Brings 78B MoE reasoning to German and English. 17. [**GEV-26B-Decide**](https://huggingface.co/autotrust/GEV-26B-Decide) - Thinks only when needed on Gemma-4. 18. [**Victoria**](https://huggingface.co/rmonsurate/Victoria) - Trims Qwen3.8-Flash-Next experts for a single GPU. 19. [**Hearim Gateway**](https://github.com/ziozzang/hearim) - Routes Jev System One decisions over local LLMs. 20. [**Sharp-Spark-X2.5-4B-GGUF**](https://huggingface.co/peculiar-ragdoll/Sharp-Spark-X2.5-4B-GGUF) - Fits a long-context coder in 6GB. 21. [**LingBot-World 2.0 Realtime**](https://github.com/kaarelkaarelson/lingbot-world-v2-realtime) - Runs a world model at 16 FPS. 22. [**deepseek-v4.1-flash-gguf**](https://huggingface.co/antirez/deepseek-v4.1-flash-gguf) - Runs DeepSeek Flash on Apple Silicon. 23. [**ThreadShelf**](https://github.com/ChrystianSchutz/ThreadShelf) - Archives and searches your AI chat history locally. 24. [**Aurora1.0-150M**](https://huggingface.co/AuroraAI-Research/Aurora1.0-150M) - Packs an efficient tiny model you can run anywhere. 25. [**Intern-S2-397B**](https://huggingface.co/internlm/Intern-S2-397B) - Scales up scientific reasoning and agents. 26. [**NeoHorse-1-9B**](https://huggingface.co/TokenRhythm/NeoHorse-1-9B) - Agentic model steps toward self-improvement. 27. [**Atria-Dawn-Preview**](https://huggingface.co/internlm/Atria-Dawn-Preview) - Agentic model aims for verifiable research results. 28. [**Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF**](https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF) - Speeds local reasoning. **🔀 Multimodal** 1. [**MiniMax-M3.1-Flash**](https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash) - Runs multimodal agentic AI with a 1M token context. 2. [**Indra-OCR**](https://huggingface.co/datalab-to/indra-ocr) - Reads PDFs and handwriting on CPU at 44 pages per second. 3. [**Qwen3.8-Flash-Next-FP4-Local**](https://huggingface.co/nvidia/Qwen3.8-Flash-Next-FP4-Local) - Runs a quantized 125B model on Blackwell GPUs. 4. [**MiniCPM-V-4.7-1B**](https://huggingface.co/openbmb/MiniCPM-V-4.7-1B) - Packs efficient multimodal AI for phones and tablets. 5. [**Muse-Spark-Open-816B-A42B**](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) - Packs 816B parameters into a multimodal agent. 6. [**MiniCPM-V-4.7-35B-A3B**](https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B) - MoE runs multimodal AI smoothly on phones. 7. [**DeepSeek-V4.1-Flash-1017**](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017) - Sharpens agents and vision. 8. [**FrogNano-4B-2609-GGUF**](https://huggingface.co/bartowski/FrogNano-4B-2609-GGUF) - Adds image input and MTP speed to quants. 9. [**Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF**](https://huggingface.co/SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF) - Drops refusals without retraining. 10. [**D1-3B**](https://huggingface.co/LiquidAI/d1-3B) - Runs calibrated decisions on edge devices in one pass. 11. [**Clef-Flash**](https://huggingface.co/Cloudflare/clef-flash) - Brings low-latency decisions to a 9B Qwen model. 12. [**GLM5.3-Flash-E224-DGX-Spark**](https://huggingface.co/autotrust/GLM5.3-Flash-E224-DGX-Spark) - Packs NVFP4 MoE for two DGX Sparks. 13. [**Clef Model**](https://huggingface.co/Cloudflare/clef) - Turns state and questions into calibrated decisions. 14. [**EmbeddingGemma 2**](https://huggingface.co/google/embeddinggemma-2) - Maps text, image, video, and audio into one space. 15. [**JEV-27B-VL**](https://huggingface.co/autotrust/JEV-27B-VL) - Brings calibrated vision decisions to Qwen3.8. 16. [**HuatuoGPT-3-9B**](https://huggingface.co/FreedomIntelligence/HuatuoGPT-3-9B) - Brings one-stage medical reasoning to local models. 17. [**Realtime-Venus**](https://huggingface.co/inclusionAI/Realtime-Venus) - Holds full-duplex conversations with voice and video. 18. [**Agnes-3.0-Flash Preview**](https://huggingface.co/Agnes-AI/Agnes-3.0-Flash) - Packs long reasoning into 33B weights. 19. [**ZDTaichu5.0-9B**](https://huggingface.co/TaichuAI/ZDTaichu5.0-9B) - Brings spatial reasoning and agent skills to VLMs. **🖼️ Image** 1. [**Kromoss**](https://huggingface.co/lodestones/Kromoss) - Fine-tunes MOSS-Image for eight-step turbo generation. 2. [**Qwen-Image-2.1-Turbo-Uncensored-GGUF**](https://huggingface.co/AtomicChat/Qwen-Image-2.1-Turbo-Uncensored-GGUF) - Drops encoder refusals locally. 3. [**MOSS-Image-Base**](https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base) - Brings layout control to local image generation. 4. [**MOSS-Image-SFT**](https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT) - Sharpens text and layout control in image generation. 5. [**Qwen-Image-2.1-Turbo**](https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo) - Renders images in eight fast steps. 6. [**Qwen-Image-2.1-Uncensored-Quantized**](https://huggingface.co/abenzerps/Qwen-Image-2.1-Uncensored-Quantized) - Ships with filters off. 7. [**FLUX 3 Image Klein**](https://huggingface.co/black-forest-labs/flux-3-image-klein) - Blends up to 10 references into one image. 8. [**Iris-3B**](https://huggingface.co/speridlabs/iris-3b) - Paints pixels directly for images, depth, and restoration. 9. [**Qwen-Image-2.1 Fun ControlNet Union**](https://huggingface.co/alibaba-pai/Qwen-Image-2.1-Fun-Controlnet-Union) - Drives eight controls at once. 10. [**Agate Preview 003**](https://huggingface.co/Logolabs/agate-preview-003) - Renders 512px images from only 261M parameters. 11. [**Wulver v0.5**](https://huggingface.co/Vaelico/Wulver) - Fine-tunes Krea 2 for anime and furry art. 12. [**PrunaSuperPoint**](https://huggingface.co/PrunaAI/PrunaSuperPoint) - Cuts SuperPoint keypoints up to 2.1x faster on edge. 13. [**VisionHOPE**](https://huggingface.co/PSRben/VisionHOPE) - Backbone self-modifies to see and segment images. 14. [**Anima-Light-Lavender**](https://huggingface.co/Johnny-Z/Anima-Light-Lavender) - Boosts natural-language anime prompts. 15. [**Krea 2 Turbo SDA LoRA**](https://huggingface.co/F16/krea2-turbo-sda) - Restores sampling diversity. 16. [**SMACK LoRA**](https://huggingface.co/LeechTM/SMACK) - Adds punchy impact effects to MiniMax H3 video. 17. [**Krea 2 Lineart Edit**](https://huggingface.co/Cierpliwy/krea2_lineart_edit) - Turns sketches into rendered images. 18. [**Krea-2-merge-tool**](https://github.com/aoleg/krea-2-merge-tool) - Merges and analyzes Krea 2 models on Windows. 19. [**Krea-2-Pixel-Art**](https://huggingface.co/e-n-v-y/Krea-2-Pixel-Art) - Adds retro pixels at three sizes to Krea 2. 20. [**MageTrail**](https://huggingface.co/RicemanT/MageTrail) - Brings Danbooru-style tuning to Microsoft's MageFlow base. **🎬 Video** 1. [**DMAD**](https://huggingface.co/ZhengmingYu/DMAD) - Distills MiniMax H3 into a four-step audio-video generator. 2. [**LTX-2.5-22b-IC-LoRA-Layout-To-Render**](https://huggingface.co/Lightricks/LTX-2.5-22b-IC-LoRA-Layout-To-Render) - VFX tools add native 4K and 8K video post production. 3. [**DuoMatching**](https://github.com/JohnZhan2023/DuoMatching) - Distills Wan2.1 into a two-step video generator. 4. [**MiniMax-H3-X2-Detail-VAE**](https://huggingface.co/speach1sdef178/MiniMax-H3-X2-Detail-VAE) - Sharpens and upscales H3 video. 5. [**LTX-2.5 IC-LoRA Ingredients**](https://huggingface.co/Lightricks/LTX-2.5-22b-IC-LoRA-Ingredients) - Keeps characters and props consistent in video. 6. [**FastH3 V2**](https://github.com/hao-ai-lab/FastVideo) - Runs fast video models on consumer GPUs. 7. [**Meridian**](https://huggingface.co/Viggle/Meridian) - Re-creates camera moves around existing video. **🎧 Audio** 1. [**Fish Audio Drama 3**](https://huggingface.co/fishaudio/drama-3) - Speaks for 16 characters in one prompt. 2. [**Drifting-TTS**](https://github.com/kadirnar/drifting-tts) - Makes one-step Turkish speech on low latency. 3. [**Whistle**](https://huggingface.co/Cactus-Compute/whistle) - Runs tiny on-device speech recognition without a GPU. 4. [**EMA Lightning**](https://huggingface.co/canberkkkkkk/ema-lightning) - Makes a tiny, fast Turkish text-to-speech model. 5. [**GRVL LoRAs**](https://huggingface.co/becausereasons/yue2-grvl-raspy-rock-soul) - Give YuE2 raspy rock-soul female vocals. 6. [**YuE2 BLGR LoRAs**](https://huggingface.co/becausereasons/yue2-blgr-bulgarian-voices) - Add Bulgarian choir and Tuvan throat singing. 7. [**YuE2 Qawwali LoRAs**](https://huggingface.co/becausereasons/yue2-qwwl-qawwali-sufi-tabla) - Add Sufi singing to music generation. 8. [**Audio Slop Server**](https://github.com/stevelittlefish/AudioSlopServer) - Runs many audio models on one GPU. 9. [**audio.cpp YuE2**](https://huggingface.co/audio-cpp/Yue2-3B-GGUF) - Runs Suno-class music in 8-9 GB of VRAM. 10. [**YuE2 Real-Audio Tokenizer**](https://huggingface.co/Mothersuperior/yue2-mothersuperior-realaudio-tokenizer-v4) - Lets you tune music on your own tracks. 11. [**YuE2 Instrumental LoRA**](https://huggingface.co/Mothersuperior/YuE2-instrumental-cot-full-loras) - Writes score-led instrumental tracks. 12. [**YuE2-hum-to-song**](https://huggingface.co/Mothersuperior/YuE2-hum-to-song) - Turns a hummed melody into a full song. 13. [**Rumik-oss-1**](https://huggingface.co/rumik-ai/rumik-oss-1) - Speaks 22 Indian languages with expressive control. 14. [**Confucius4-R2T2**](https://huggingface.co/netease-youdao/Confucius4-R2T2) - Delivers low-latency streaming speech recognition. **🤖 Agents** 1. [**Tencent Octop**](https://github.com/TencentCloud/Octop) - Runs a self-hosted multi-agent assistant at home. **🎲 3D** 1. [**AssetFurnace**](https://github.com/Bingeljell/image-to-3dlab) - Turns a prompt into a rigged, animated 3D character. **⚡ LoRA** 1. [**MiniMax-H3-Person-Remover-LoRA**](https://huggingface.co/akatz-ai/MiniMax-H3-Person-Remover-LoRA) - Deletes people from video. 2. [**Pruna LoRAs**](https://huggingface.co/PrunaAI/Pruna-Qwen-Image-2.1) - Cut Qwen-Image-2.1 down to six fast steps. 3. [**Krea 2 Inpaint Edit LoRA**](https://huggingface.co/Cierpliwy/krea2-inpaint-edit) - Fills masked areas in ComfyUI. 4. [**YuE2 Deathmetal LoRA**](https://huggingface.co/smereces/YuE2_Deathmetalv1_lora) - Adds melodic death metal to AI songs. 5. [**TRBDR Folk Troubadour LoRAs**](https://huggingface.co/becausereasons/yue2-trbdr-folk-troubadour) - Give YuE2 a 1960s balladeer. 6. [**AnyAngle LoRA**](https://huggingface.co/lilylilith/QI_2.1_AnyAngle) - Retains style while changing camera angles. 7. [**QTSTRM Quiet Storm LoRAs**](https://huggingface.co/becausereasons/yue2-qtstrm-quiet-storm) - Bring 80s R&B vocals to YuE2. 8. [**LTX-2.5 Restore IC-LoRA**](https://huggingface.co/Lightricks/LTX-2.5-22b-IC-LoRA-Restore) - Revives old footage in color. 9. [**MiniMax-H3 Motion Continuity Repair V2**](https://huggingface.co/JOKER141/MiniMax-H3-General-Motion-Continuity-Repair) - Fixes broken action frames. 10. [**MiniMax-H3 CrossView LoRA**](https://huggingface.co/Cseti/MiniMax-H3_Ref2VA-LoRA-CrossView-Warp_v1) - Shoots a scene from new angles. 11. [**Krea2-Turbo-Distill-2step-LoRA**](https://huggingface.co/lvladikov/Krea2-Turbo-Distill-2step-LoRA) - Cuts image steps from eight to two. 12. [**TaoMate-H3-3Step-ComfyUI**](https://huggingface.co/Robert1212star/TaoMate-H3-3Step-ComfyUI) - Speeds up H3 runs in ComfyUI. 13. [**Minimax-H3-ComfyUI LoRAs**](https://huggingface.co/Alissonerdx/Minimax-H3-ComfyUI) - Add style and VFX editing to H3 video. 14. [**BFS-Best-Face-Swap LoRAs**](https://huggingface.co/Alissonerdx/BFS-Best-Face-Swap) - Swap faces across Qwen Image and Flux. **🏋️ Training** 1. [**LoRAlab TrainerStudio**](https://github.com/AcademiaSD/AcademiaSD_LoRAlab-TrainerStudio) - Trains latest models on low-VRAM GPUs. 2. [**ModelFiche**](https://github.com/latentwill/modelfiche) - Organizes image-model training for agents and reviewers. **📊 Datasets** 1. [**Danbooru 2026 Tag Cleaning Corrections Dataset**](https://huggingface.co/datasets/Grio43/Tag_cleaning) - Released with cleaned-up tag corrections for training. **☰ UI** 1. [**TinyLuma**](https://github.com/ThetaCursed/TinyLuma) - Is an 8MB desktop editor for finishing AI images. 2. [**Goated Prompter**](https://github.com/Jolanoff/goated-prompter) - Refines image and video prompts locally. 3. [**Fooocus-Qwen-Image-2.1**](https://github.com/ogoun/fooocus-qwen-image-2.1) - Brings Qwen-Image editing to small GPUs. 4. [**Slopus**](https://github.com/bitti-ai/slopus) - Turns MiniMax H3 video generation into a desktop app. 5. [**Image Studio**](https://github.com/tamimKTH/image-studio) - Brings local Qwen-Image 2.1 editing to Apple Silicon. 6. [**Pinhole**](https://github.com/DavidGudovic/PinholeAI) - Turns any GPU into a simple one-button local image generator. 7. [**H3 Prompt Studio**](https://github.com/BesianSherifaj-AI/h3-prompt-studio) - Directs MiniMax H3 films from a local app. 8. [**Qamba Studio**](https://github.com/qamba-labs/qamba-studio-oss) - Edits AI video end-to-end on your own machine. 9. [**PotionUI**](https://github.com/PotionUI/PotionUI) - Puts self-hosted diffusion studios in a browser. 10. [**H3 Studio**](https://github.com/janishar/h3c-studio) - Gives MiniMax-H3 video a local Mac web UI. 11. [**Livebound**](https://github.com/MoonBlack87/livebound) - Brings a local publishing workspace to CivitAI creators. **💻 Hardware** 1. [**AMD EPYC 9006 Venice CPUs**](https://videocardz.com/newz/amd-lists-zen6-epyc-9006-venice-cpus-from-8-cores-at-700-to-epyc-9996-256-cores-at-14904) - Listed from $700 to $14,904. 2. [**Inno3D AWS-511Z-R1 Tower Workstation**](https://www.techpowerup.com/353146/inno3d-launches-the-tower-workstation-aws-511z-r1) - Handles up to 4 GPUs. 3. [**GMKtec EVO-X5 Pro**](https://www.techpowerup.com/353151/gmktec-launches-evo-x5-pro-with-amd-ryzen-ai-max-pro-495-and-192-gb-unified-memory) - Runs up to 320B LLMs locally. 4. [**AceMagic F9A**](https://www.techpowerup.com/353163/acemagic-launches-f9a-usd-6-500-workstation-mini-pc-with-amd-ryzen-ai-max-pro-495) - Packs 131 TOPS AI into a 2L mini workstation. 5. [**Innodisk DDR5-8000 RDIMM**](https://www.techpowerup.com/353201/innodisk-launches-ddr5-8000-rdimm-for-edge-ai-and-high-performance-computing) - Launched for edge AI and servers. 6. [**Panasonic Toughbook 34 And G3**](https://www.techpowerup.com/353264/panasonic-connect-launches-the-toughbook-34-and-toughbook-g3-fully-rugged-2-in-1-mobile-computers) - Bring AI-ready rugged laptops. 7. [**Minisforum AtomMan G1 Pro**](https://www.techpowerup.com/353262/minisforum-atomman-g1-pro-89a6-mini-pc-launches-with-rx-9060) - Adds a 16 GB RX 9060 XT to its mini PCs. 8. [**Cervoz T449 And T459**](https://www.techpowerup.com/353299/cervoz-launches-mil-std-810h-m-2-nvme-ssds-for-mission-critical-embedded-systems) - Add rugged MIL-STD SSDs for edge AI. **🛠️ Other Tools** 1. [**Open-omnivoice-tts**](https://github.com/hypersniper05/open-omnivoice-tts) - Serves a 0.3 second voice clone API. 2. [**VRAM Manager**](https://github.com/Elhelali/vram-manager) - Flags idle AI servers holding GPU memory. 3. [**qwen-image-local**](https://github.com/Mats2208/qwen-image-local) - Runs Qwen-Image-2.1 on a 12 GB GPU. 4. [**Lumibelle Video Studio**](https://github.com/radiatingreverberations/lumibelle) - Turns a screenplay into a local film. 5. [**AutoRef**](https://github.com/KuOnoda/AutoRef) - Coaches FLUX.2 into composing multiple reference images. 6. [**MiniMax Code**](https://github.com/MiniMax-AI/minimax-code) - Turns terminal prompts into working projects. 7. [**Tool-prune**](https://github.com/hemanth/tool-prune) - Prunes tool schemas to cut agent prompt costs. 8. [**ZCode**](https://github.com/zai-org/ZCode) - Open sources an AI coding workbench in three forms. 9. [**YuE2-Radio**](https://github.com/PasiKoodaa/YuE2-Radio) - Creates an endless local AI music radio. 10. [**Inkvec**](https://github.com/logolabs/inkvec) - Traces images into clean vector SVGs in the browser. 11. [**wan2gp-h3-latent-continue**](https://github.com/g3n3rativ3/wan2gp-h3-latent-continue) - Lets Wan2GP continue H3 video from latents. 12. [**Jeff**](https://github.com/firelex/jeff) - Puts a fast 0.8B decision model in front of your 27B. ## ComfyUI Releases **Speed & Performance Boosts** 1. [**ComfyUI-VELA-H3**](https://github.com/Speach1sdef178/ComfyUI-VELA-H3) - Speeds up VDN H3 on MiniMax H3 in ComfyUI. 2. [**Veda-on-ComfyUI**](https://github.com/veda-sparse/Veda-on-ComfyUI) - Speeds up MiniMax H3 in ComfyUI. 3. [**ComfyUI H3 W4A4 VSA**](https://github.com/sepiablue-ai/ComfyUI-MiniMax-H3-W4A4-VSA) - Cuts Ref2VA video time down to 3 minutes. 4. [**ComfyUI-Ref2VA-VSA**](https://github.com/Kablex/ComfyUI-Ref2VA-VSA) - Speeds character video generation to 72 seconds. 5. [**ComfyUI-Spectrum-Qwen2.1**](https://github.com/awdqwdasdg/Comfyui-Spectrum-Qwen2.1) - Speeds up image generation by about 2x. 6. [**ComfyUI-Qwen-Image-2.1-PromptEnhancer-MTP**](https://github.com/mozophe/ComfyUI-Qwen-Image-2.1-PromptEnhancer-MTP) - Runs faster in ComfyUI with MTP support. 7. [**ComfyUI-HunyuanImage3**](https://github.com/PedroMarinhoDev/ComfyUI-HunyuanImage3) - Runs an 80B image model on a single GPU. 8. [**ComfyUI-AppleSilicon-FP8**](https://github.com/pawel-mazurkiewicz/ComfyUI-AppleSilicon-FP8) - Runs quantized models on Mac. 9. [**LynnReal-Omni**](https://github.com/LynnReal-AI/LynnReal-Omni) - Packs many video tasks into one fast model. **Video & Image Creation** 1. [**H3 Long Shot Studio**](https://github.com/r34vtraining/Longshot_Studio/tree/main) - Makes multi-shot films inside ComfyUI. 2. [**ComfyUI-Olm-SoL-Refiner**](https://github.com/o-l-l-i/ComfyUI-Olm-SoL-Refiner) - Brings video refinement to nodes. 3. [**Fizgig H3 Still**](https://github.com/shootthesound/ComfyUI-Fizgig-H3-Still) - Renders true MiniMax H3 frames in ComfyUI. 4. [**Fizgig H3 Tweaks**](https://github.com/shootthesound/ComfyUI-Fizgig-H3-Tweaks) - Gives ComfyUI fine control over MiniMax H3. 5. [**Scumble**](https://github.com/DenRakEiw/scumble) - A free layer-based AI inpainting editor. 6. [**ComfyUI-BubbleText**](https://github.com/arturor1990/ComfyUI-BubbleText) - Puts real text inside genera
retained text (truncated)
Here's what you probably missed in October 2026: **🧠 LLMs** 1. [**Ornith-1.5-122B**](https://huggingface.co/ornith-ai/Ornith-1.5-122B) - Handles million-token agent tasks on a single GPU. 2. [**Raon-1-mini-preview**](https://huggingface.co/KRAFTON/Raon-1-mini-preview) - Runs Korean agentic AI on a single RTX 5090. 3. [**Timestral-209M**](https://huggingface.co/mistralai/Timestral-209M) - Forecasts time series in one pass with 21 quantiles. 4. [**Carbon-X**](https://huggingface.co/HuggingFaceBio/Carbon-X) - Reads DNA in chat and predicts variant effects on one GPU. 5. [**Sage-A1.5**](https://huggingface.co/InternScience/Sage-A1.5) - Orchestrates scientific agents with self-review. 6. [**Spark-X2.6-4B**](https://huggingface.co/XHToken/Spark-X2.6-4B) - Advances on-device agents with controllable thinking. 7. [**Nanbeige4.3-33B**](https://huggingface.co/Nanbeige/Nanbeige4.3-33B) - Teacher model leads agentic coding and cowork. 8. [**Flint**](https://huggingface.co/RecursiveSI/flint) - MoE coding agent investigates repos and fixes bugs on one GPU. 9. [**Hy-Memory-30B-A3B**](https://huggingface.co/tencent/Hy-Memory-30B-A3B) - Writes durable agent memory at 3B active parameters. 10. [**Underdog-Saluki-27B-1.0**](https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0) - Packs Qwen3.8 tool calling into under 8 GB. 11. [**D3**](https://huggingface.co/vllm-sr/d3) - Answers choice, yes-no, and score questions in one call. 12. [**LFM2.5-24B-A2B**](https://huggingface.co/LiquidAI/LFM2.5-24B-A2B) - Packs fast on-device agent reasoning into 2B active. 13. [**Qwen3.5-2B-RCOL**](https://huggingface.co/trubisky/Qwen3.5-2B-RCOL) - Low-bit quants hold quality under 1 GB. 14. [**GLM-5.3-UNCENSORED-EXL3-3.0bpw**](https://huggingface.co/Infatoshi/GLM-5.3-UNCENSORED-EXL3-3.0bpw) - Fits a 753B MoE in 273 GB. 15. [**Humanizer**](https://huggingface.co/jialinyyzz/humanizer) - Rewrites AI drafts to sound human in English and Chinese. 16. [**Kolibri 1**](https://huggingface.co/Aleph-Alpha/Kolibri-1) - Brings 78B MoE reasoning to German and English. 17. [**GEV-26B-Decide**](https://huggingface.co/autotrust/GEV-26B-Decide) - Thinks only when needed on Gemma-4. 18. [**Victoria**](https://huggingface.co/rmonsurate/Victoria) - Trims Qwen3.8-Flash-Next experts for a single GPU. 19. [**Hearim Gateway**](https://github.com/ziozzang/hearim) - Routes Jev System One decisions over local LLMs. 20. [**Sharp-Spark-X2.5-4B-GGUF**](https://huggingface.co/peculiar-ragdoll/Sharp-Spark-X2.5-4B-GGUF) - Fits a long-context coder in 6GB. 21. [**LingBot-World 2.0 Realtime**](https://github.com/kaarelkaarelson/lingbot-world-v2-realtime) - Runs a world model at 16 FPS. 22. [**deepseek-v4.1-flash-gguf**](https://huggingface.co/antirez/deepseek-v4.1-flash-gguf) - Runs DeepSeek Flash on Apple Silicon. 23. [**ThreadShelf**](https://github.com/ChrystianSchutz/ThreadShelf) - Archives and searches your AI chat history locally. 24. [**Aurora1.0-150M**](https://huggingface.co/AuroraAI-Research/Aurora1.0-150M) - Packs an efficient tiny model you can run anywhere. 25. [**Intern-S2-397B**](https://huggingface.co/internlm/Intern-S2-397B) - Scales up scientific reasoning and agents. 26. [**NeoHorse-1-9B**](https://huggingface.co/TokenRhythm/NeoHorse-1-9B) - Agentic model steps toward self-improvement. 27. [**Atria-Dawn-Preview**](https://huggingface.co/internlm/Atria-Dawn-Preview) - Agentic model aims for verifiable research results. 28. [**Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF**](https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-27B-GSQ-RCO-GGUF) - Speeds local reasoning. **🔀 Multimodal** 1. [**MiniMax-M3.1-Flash**](https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash) - Runs multimodal agentic AI with a 1M token context. 2. [**Indra-OCR**](https://huggingface.co/datalab-to/indra-ocr) - Reads PDFs and handwriting on CPU at 44 pages per second. 3. [**Qwen3.8-Flash-Next-FP4-Local**](https://huggingface.co/nvidia/Qwen3.8-Flash-Next-FP4-Local) - Runs a quantized 125B model on Blackwell GPUs. 4. [**MiniCPM-V-4.7-1B**](https://huggingface.co/openbmb/MiniCPM-V-4.7-1B) - Packs efficient multimodal AI for phones and tablets. 5. [**Muse-Spark-Open-816B-A42B**](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) - Packs 816B parameters into a multimodal agent. 6. [**MiniCPM-V-4.7-35B-A3B**](https://huggingface.co/openbmb/MiniCPM-V-4.7-35B-A3B) - MoE runs multimodal AI smoothly on phones. 7. [**DeepSeek-V4.1-Flash-1017**](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017) - Sharpens agents and vision. 8. [**FrogNano-4B-2609-GGUF**](https://huggingface.co/bartowski/FrogNano-4B-2609-GGUF) - Adds image input and MTP speed to quants. 9. [**Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF**](https://huggingface.co/SC117/Qwen3.8-Flash-Next-GSQ-RCO-abliterated-GGUF) - Drops refusals without retraining. 10. [**D1-3B**](https://huggingface.co/LiquidAI/d1-3B) - Runs calibrated decisions on edge devices in one pass. 11. [**Clef-Flash**](https://huggingface.co/Cloudflare/clef-flash) - Brings low-latency decisions to a 9B Qwen model. 12. [**GLM5.3-Flash-E224-DGX-Spark**](https://huggingface.co/autotrust/GLM5.3-Flash-E224-DGX-Spark) - Packs NVFP4 MoE for two DGX Sparks. 13. [**Clef Model**](https://huggingface.co/Cloudflare/clef) - Turns state and questions into calibrated decisions. 14. [**EmbeddingGemma 2**](https://huggingface.co/google/embeddinggemma-2) - Maps text, image, video, and audio into one space. 15. [**JEV-27B-VL**](https://huggingface.co/autotrust/JEV-27B-VL) - Brings calibrated vision decisions to Qwen3.8. 16. [**HuatuoGPT-3-9B**](https://huggingface.co/FreedomIntelligence/HuatuoGPT-3-9B) - Brings one-stage medical reasoning to local models. 17. [**Realtime-Venus**](https://huggingface.co/inclusionAI/Realtime-Venus) - Holds full-duplex conversations with voice and video. 18. [**Agnes-3.0-Flash Preview**](https://huggingface.co/Agnes-AI/Agnes-3.0-Flash) - Packs long reasoning into 33B weights. 19. [**ZDTaichu5.0-9B**](https://huggingface.co/TaichuAI/ZDTaichu5.0-9B) - Brings spatial reasoning and agent skills to VLMs. **🖼️ Image** 1. [**Kromoss**](https://huggingface.co/lodestones/Kromoss) - Fine-tunes MOSS-Image for eight-step turbo generation. 2. [**Qwen-Image-2.1-Turbo-Uncensored-GGUF**](https://huggingface.co/AtomicChat/Qwen-Image-2.1-Turbo-Uncensored-GGUF) - Drops encoder refusals locally. 3. [**MOSS-Image-Base**](https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base) - Brings layout control to local image generation. 4. [**MOSS-Image-SFT**](https://huggingface.co/OpenMOSS-Team/MOSS-Image-SFT) - Sharpens text and layout control in image generation. 5. [**Qwen-Image-2.1-Turbo**](https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo) - Renders images in eight fast steps. 6. [**Qwen-Image-2.1-Uncensored-Quantized**](https://huggingface.co/abenzerps/Qwen-Image-2.1-Uncensored-Quantized) - Ships with filters off. 7. [**FLUX 3 Image Klein**](https://huggingface.co/black-forest-labs/flux-3-image-klein) - Blends up to 10 references into one image. 8. [**Iris-3B**](https://huggingface.co/speridlabs/iris-3b) - Paints pixels directly for images, depth, and restoration. 9. [**Qwen-Image-2.1 Fun ControlNet Union**](https://huggingface.co/alibaba-pai/Qwen-Image-2.1-Fun-Controlnet-Union) - Drives eight controls at once. 10. [**Agate Preview 003**](https://huggingface.co/Logolabs/agate-preview-003) - Renders 512px images from only 261M parameters. 11. [**Wulver v0.5**](https://huggingface.co/Vaelico/Wulver) - Fine-tunes Krea 2 for anime and furry art. 12. [**PrunaSuperPoint**](https://huggingface.co/PrunaAI/PrunaSuperPoint) - Cuts SuperPoint keypoints up to 2.1x faster on edge. 13. [**VisionHOPE**](https://huggingface.co/PSRben/VisionHOPE) - Backbone self-modifies to see and segment images. 14. [**Anima-Light-Lavender**](https://huggingface.co/Johnny-Z/Anima-Light-Lavender) - Boosts natural-language anime prompts. 15. [**Krea 2 Turbo SDA LoRA**](https://huggingface.co/F16/krea2-turbo-sda) - Restores sampling diversity. 16. [**SMACK LoRA**](https://huggingface.co/LeechTM/SMACK) - Adds punchy impact effects to MiniMax H3 video. 17. [**Krea 2 Lineart Edit**](https://huggingface.co/Cierpliwy/krea2_lineart_edit) - Turns sketches into rendered images. 18. [**Krea-2-merge-tool**](https://github.com/aoleg/krea-2-merge-tool) - Merges and analyzes Krea 2 models on Windows. 19. [**Krea-2-Pixel-Art**](https://huggingface.co/e-n-v-y/Krea-2-Pixel-Art) - Adds retro pixels at three sizes to Krea 2. 20. [**MageTrail**](https://huggingface.co/RicemanT/MageTrail) - Brings Danbooru-style tuning to Microsoft's MageFlow base. **🎬 Video** 1. [**DMAD**](https://huggingface.co/ZhengmingYu/DMAD) - Distills MiniMax H3 into a four-step audio-video generator. 2. [**LTX-2.5-22b-IC-LoRA-Layout-To-Render**](https://huggingface.co/Lightricks/LTX-2.5-22b-IC-LoRA-Layout-To-Render) - VFX tools add native 4K and 8K video post production. 3. [**DuoMatching**](https://github.com/JohnZhan2023/DuoMatching) - Distills Wan2.1 into a two-step video generator. 4. [**MiniMax-H3-X2-Detail-VAE**](https://huggingface.co/speach1sdef178/MiniMax-H3-X2-Detail-VAE) - Sharpens and upscales H3 video. 5. [**LTX-2.5 IC-LoRA Ingredients**](https://huggingface.co/Lightricks/LTX-2.5-22b-IC-LoRA-Ingredients) - Keeps characters and props consistent in video. 6. [**FastH3 V2**](https://github.com/hao-ai-lab/FastVideo) - Runs fast video models on consumer GPUs. 7. [**Meridian**](https://huggingface.co/Viggle/Meridian) - Re-creates camera moves around existing video. **🎧 Audio** 1. [**Fish Audio Drama 3**](https://huggingface.co/fishaudio/drama-3) - Speaks for 16 characters in one prompt. 2. [**Drifting-TTS**](https://github.com/kadirnar/drifting-tts) - Makes one-step Turkish speech on low latency. 3. [**Whistle**](https://huggingface.co/Cactus-Compute/whistle) - Runs tiny on-device speech recognition without a GPU. 4. [**EMA Lightning**](https://huggingface.co/canberkkkkkk/ema-lightning) - Makes a tiny, fast Turkish text-to-speech model. 5. [**GRVL LoRAs**](https://huggingface.co/becausereasons/yue2-grvl-raspy-rock-soul) - Give YuE2 raspy rock-soul female vocals. 6. [**YuE2 BLGR LoRAs**](https://huggingface.co/becausereasons/yue2-blgr-bulgarian-voices) - Add Bulgarian choir and Tuvan throat singing. 7. [**YuE2 Qawwali LoRAs**](https://huggingface.co/becausereasons/yue2-qwwl-qawwali-sufi-tabla) - Add Sufi singing to music generation. 8. [**Audio Slop Server**](https://github.com/stevelittlefish/AudioSlopServer) - Runs many audio models on one GPU. 9. [**audio.cpp YuE2**](https://huggingface.co/audio-cpp/Yue2-3B-GGUF) - Runs Suno-class music in 8-9 GB of VRAM. 10. [**YuE2 Real-Audio Tokenizer**](https://huggingface.co/Mothersuperior/yue2-mothersuperior-realaudio-tokenizer-v4) - Lets you tune music on your own tracks. 11. [**YuE2 Instrumental LoRA**](https://huggingface.co/Mothersuperior/YuE2-instrumental-cot-full-loras) - Writes score-led instrumental tracks. 12. [**YuE2-hum-to-song**](https://huggingface.co/Mothersuperior/YuE2-hum-to-song) - Turns a hummed melody into a full song. 13. [**Rumik-oss-1**](https://huggingface.co/rumik-ai/rumik-oss-1) - Speaks 22 Indian languages with expressive control. 14. [**Confucius4-R2T2**](https://huggingface.co/netease-youdao/Confucius4-R2T2) - Delivers low-latency streaming speech recognition. **🤖 Agents** 1. [**Tencent Octop**](https://github.com/TencentCloud/Octop) - Runs a self-hosted multi-agent assistant at home. **🎲 3D** 1. [**AssetFurnace**](https://github.com/Bingeljell/image-to-3dlab) - Turns a prompt into a rigged, animated 3D character. **⚡ LoRA** 1. [**MiniMax-H3-Person-Remover-LoRA**](https://huggingface.co/akatz-ai/MiniMax-H3-Person-Remover-LoRA) - Deletes people from video. 2. [**Pruna LoRAs**](https://huggingface.co/PrunaAI/Pruna-Qwen-Image-2.1) - Cut Qwen-Image-2.1 down to six fast steps. 3. [**Krea 2 Inpaint Edit LoRA**](https://huggingface.co/Cierpliwy/krea2-inpaint-edit) - Fills masked areas in ComfyUI. 4. [**YuE2 Deathmetal LoRA**](ht
document fetched 2026-11-03 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xlwezy/local_ai_news_you_missed_october_2026