The open‑source AI scene delivered an avalanche of new models, tools, and adapters this month. From trillion‑parameter LLMs to tiny on‑device translation apps, there’s something for everyone. Here’s your straight‑to‑the‑point breakdown.
Ring‑2.6‑1T is a trillion‑parameter model purpose‑built for continuous agent workflows and complex multi‑step tasks. Ling‑2.6‑1T brings another trillion‑parameter option, this time focused on coding and tool calling. Mistral‑Medium‑3.5‑128B arrives as a dense flagship that consolidates several previous Mistral releases into one model.
Command‑A‑Plus‑05‑2026‑Bf16 packs a 25‑billion‑parameter active engine that handles both text and images with a 128K context window. Step‑3.7‑Flash is a 198B vision‑language model that activates only 11B parameters per token thanks to sparse mixture‑of‑experts routing. NVIDIA‑Nemotron‑Labs‑3‑Elastic‑30B‑A3B‑BF16 is a single checkpoint that can serve three different reasoning model sizes on the fly.
Emo introduces a mixture‑of‑experts design where experts self‑organize into topics without human labels. ZAYA1‑8B uses only 760 million active parameters for deep long‑form reasoning tasks. Intern‑S2‑Preview is a 35B scientific assistant that understands text, images, and time‑series while calling external tools.
The refusal‑free movement is stronger than ever. Qwen3.6‑27B‑OBLITERATED surgically reduces safety refusals through direct weight editing, while Qwen3.5‑27B‑uncensored‑heretic‑v2‑Native‑MTP‑Preserved keeps all 15 Multi‑Token Prediction layers intact after censorship removal. Qwen3.6‑35B‑A3B‑Uncensored‑Genesis‑V2‑APEX‑MTP‑GGUF delivers a refusal‑free Qwen MoE as a ready‑to‑run quantized package.
Qwen3.6‑35B‑A3B‑uncensored‑heretic‑Native‑MTP‑Preserved cuts unwanted refusals by 88% while preserving 19 MTP layers. Qwen3.6‑27B‑AEON‑Ultimate‑Uncensored‑BF16 strips the “safety tax” completely for direct instruction‑following. Qwen3.6‑27B‑Heretic‑Uncensored‑FINETUNE‑NEO‑CODE‑Di‑IMatrix‑MAX‑GGUF packages the refusal‑free model in highly accurate compressed GGUF formats.
Gemma 4 models got the same treatment. Gemma‑4‑Ortenzya‑The‑Creative‑Wordsmith‑31B‑it‑uncensored‑heretic cuts refusals while boosting creative writing, G4‑MeroMero‑31B‑uncensored‑heretic strips refusals for storytelling, and Gemma‑4‑Gembrain‑31B‑it‑uncensored‑heretic uses abliteration to remove safety blocks. Gemma4‑26B‑A4B‑Uncensored‑HauhauCS‑Balanced scores zero refusals across 465 test prompts while keeping full capabilities.
Small models are punching above their weight. Supra‑50M is a tiny 50M‑parameter model trained from scratch that beats GPT‑2 on specific benchmarks. MiniCPM5‑1B runs entirely on personal devices, switching between a fast assistant and a deeper reasoning mode. Nandi‑Mini‑600M‑Early‑Checkpoint is an early preview of a compact model supporting English and 11 Indic languages.
Translation gets a major boost. Hy‑MT2‑30B‑A3B is an open‑source MoE translator covering 33 languages, while Hy‑MT2‑1.8B offers speedier translation for real‑world text. Hy‑MT1.5‑1.8B‑1.25bit shrinks the translation system to run entirely on a phone offline.
Domain‑specific models also appeared. AntAngelMed is a medical MoE model for clinical reasoning; Leanly_AI supports psychologists working with obesity patients. SmallCode is a terminal‑based coding agent that keeps code private on your hardware. needle is a 26M‑parameter model built exclusively for function calling and tool use. BitCPM4‑CANN‑8B compresses weights to three values, cutting memory by six times. Ettin‑Reranker‑1b‑V1 boosts search quality by scoring text pairs. HRM‑Text‑1B uses a dual‑timescale architecture instead of a standard transformer.
Fara‑7B is an open‑weight computer‑use agent that plans and executes web tasks by seeing screenshots. NuExtract3 extracts structured data from documents into Markdown using a 4B vision‑language model. Qwopus3.5‑9B‑Coder‑GGUF and MiMo‑V2.5‑coder‑Q2 bring compressed coding agents to local machines. Qwopus3.6‑27B‑v2‑MTP‑GGUF delivers a quantized reasoning model using multi‑token prediction.
Microsoft’s Lens is a 3.8B foundational model that outperforms many 6B+ alternatives with far less training compute. Its distilled sibling Lens‑Turbo generates high‑quality images in just four steps. Walkyrie‑1.3B‑v1.0 was rebuilt from a video model to produce crisp 1024×1024 images from prompts.
HiDream‑O1‑Image creates, edits, and personalizes pictures without needing separate compression tools. Nemotron‑Labs‑Diffusion‑14B can generate text either normally or with a faster diffusion‑based parallel method. Anima Base v1.0 is a 2B model focused on anime‑style and non‑photorealistic artwork.
Style control is expanding. Flux.2‑Klein‑Loras packs multiple style adapters for the Flux.2 Klein 9B model. AsymFLUX.2‑klein‑9B lets the same base model generate raw pixel images without a VAE. Qwen‑2512‑portrait sharpens human portraits with natural skin detail, and UltraReal_FineTune_Anima pushes the Anima generator toward realistic photo outputs.
Vision‑language AI is now faster and more capable. LocateAnything‑3B from NVIDIA marks objects or text in images based on plain prompts. Keye‑VL‑2.0‑30B‑A3B understands long videos and performs agent tasks like web search using sparse attention. MiniCPM‑V‑4.6 brings image and video understanding directly to smartphones with a cloud‑free experience.
SenseNova‑U1‑A3B‑MoT unifies image understanding, generation, and editing without separate visual encoders. Ovis2.6‑80B‑A3B examines high‑res images and long documents efficiently. NVIDIA also released Nemotron‑3‑Nano‑Omni‑30B‑A3B‑Reasoning‑NVFP4, a multimodal AI that processes video, audio, and text in a single pipeline.
Nvidia’s PiD is a pixel diffusion decoder that speeds up high‑resolution image generation by denoising directly in pixel space. Lance is a unified model that handles image and video understanding, generation, and editing in one place. IBM contributed Granite‑4.1‑8b for chat and instruction following, and Granite‑4.1‑30b for upgraded tool calling and long‑context tasks.
SANA‑WM Bidirectional creates minute‑long 720p videos from a single starting image. LongCat‑Video‑Avatar‑1.5 generates talking avatar videos with realistic characters or animations. LTX2.3‑10Eros turns a still image into a short motion clip by merging layers from different training steps.
Pixal3D transforms a single image into a detailed 3D asset. ScreenDiffusion instantly reimagines your desktop as living art through image‑to‑image AI. Phosphene is a free Mac panel that creates video clips with synced audio using LTX 2.3. studiomi300 chains multiple models to produce a 30‑second cinematic reel from a text prompt.
The LTX Video 2.3 ecosystem is exploding. LTX‑2.3 Upscale IC Lora turns soft clips into cleaner, sharper footage. VR‑360‑Outpaint‑LTX2.3‑IC‑LoRA converts widescreen to full 360‑degree equirectangular video. SYSTMS‑FLW‑IC‑LORA‑LTX‑2.3 creates smooth shot‑to‑shot transitions.
LTX‑2.3‑Dearchive‑Lora makes vintage footage look like it was shot yesterday. Obscura Remova removes haze, smoke, or foreground objects from clips. LTX‑2.3‑22b‑IC‑LoRA‑LipDub replaces speech and lip motion with synchronized audio. OmniNFT uses reinforcement learning to align audio and video generation better.
Marlin‑2B extracts structured descriptions and second‑precise timestamps from footage. Causal‑Forcing is a training method that distills large video models into efficient real‑time generators. Vlo v0.2.0 is a timeline‑based video editor with ComfyUI‑powered generative AI.
MOSS‑TTS‑v1.5 upgrades zero‑shot voice cloning with better quality. DramaBox turns scene descriptions into expressive speech with laughs, sighs, and pauses. Scenema‑Audio clones voices and performs scene‑aware emotional speech with ambient sounds.
Supertonic‑3 runs fully on‑device TTS with ONNX, expanding language support. Derpy‑Turtle‑The‑Kokoro‑Trainer blends Kokoro TTS with RVC voice conversion in a Windows GUI. Comfyui‑controlfoley generates synced foley sound effects directly from silent video.
Ace‑Step‑1.5‑XL‑Concept‑Sliders let you push a local music generator toward or away from specific audio traits. Ace‑Step‑1.5‑Api‑server‑UI wraps the model into a full‑featured local studio interface. MusiCue converts songs into timeline‑based cues for driving animation or show‑control software.
ComfyUI‑SmartPromptCrafter auto‑builds optimized prompt pairs for any checkpoint. RebelsPromptEnhancer rewrites short ideas into detailed prompts using a local 4B model. ErniePEUnleashed adds foreground‑background layering and lighting logic to descriptions. ComfyUI‑Anima‑Style‑Nodes lets you visually browse and apply anime artist tags, and Comfyui‑Anima‑Regional‑Conditioning routes attention so masked areas get specific prompts only.
Orion4D_generative_paint adds a full painting interface right in the browser. ComfyUI‑Olm‑Liquify brings Photoshop‑style warping with push, twirl, and pinch brushes. ComfyUI‑Angelo merges a sampler and inpaint refiner so you can paint fixes directly on outputs. ComfyUI_KleinTiledUpscaler upscales using tiled inpainting for creative detail.
ComfyUI‑PiD integrates NVIDIA’s pixel diffusion decoder, skipping the traditional VAE. ComfyUI‑FeatherOps accelerates diffusion inference on AMD RDNA3 GPUs with a custom HIP kernel. ComfyUI‑Safe‑Chunked‑Image‑Blend gives explicit control over batch resize and blending. ComfyUI‑ReferenceLatentPlus provides per‑image control over how references influence outputs. ComfyUI‑Untwisting‑RoPE brings training‑free style transfer to diffusion transformer models.
ComfyUI‑Yedp‑Action‑Director puts a full 3D viewport with path tracing into workflows. ComfyUI‑Magos‑Nodes adds a skeleton editor and retargeter for body and face keypoints. Comfyui_VideoCombine_Plus extends video combining with sound volume and extra controls. ComfyUI‑DramaBox brings ResembleAI’s expressive TTS into the node graph. ComfyUI‑XAV‑Google‑Sheets pulls text from a public Google Sheet to drive generations.
WorkflowX‑Configurator switches between workflow profiles without duplicating graphs. ComfyUI‑Workflow‑Finder searches local workflow collections with plain English descriptions. ComfyUI‑lora‑FindingLora replaces the stock LoRA loader with fuzzy search, bookmarking, and trigger word storage. BangtrixToolkit overlays a real‑time hardware monitor onto the canvas and includes a universal prompt translator.
ComfyUI‑ialhabbal bundles eight tools including interactive prompt review and batch loading. ComfyUI_ShowMe lets you draw annotation notes directly on the workflow canvas. ComfyUI‑gonztok_nodes replaces text inputs with visual pickers for images and LoRAs. ComfyUI‑SPEED nearly doubles sampling speed with Spectral Progressive Diffusion. Comfyui‑Mesh splits inference across two Nvidia GPUs using NVENC video encoder chips. ComfyUI‑PlagueKind‑Nodes unifies image and mask resizing in one step.
ComfyUI‑Fayens streamlines face swap pipelines with automatic face crops and masks. Comfyui‑Clippy‑Reloaded pastes images directly from your clipboard into a workflow. comfyui‑artius‑browser adds a fast sidebar asset manager for dragging images, videos, and 3D files. Comfyui‑node‑canvas gives a GUI app to build custom nodes without writing boilerplate code.
hipEngine is a new ROCm‑native inference engine for AMD RDNA3 GPUs that runs LLMs without PyTorch. MiniCPM‑V‑4.6‑OrangePi combines a from‑scratch C++ engine to bring the vision‑language model to a $100 edge board. Beellama.cpp forks llama.cpp with DFlash speculative decoding, TurboQuant KV‑cache compression, and better memory usage.
ExLlamaV3 introduces the EXL3 quantization format for very low bitrate performance. ds4.pinokio launches DeepSeek V4 Flash on Apple Silicon Macs with a native Metal engine. Deepseek‑V4‑GGUF shrinks the massive model to fit high‑end consumer GPUs. Qwen3.5‑9B‑DeepSeek‑V4‑Flash‑GGUF packs DeepSeek‑V4’s reasoning into a 9B package for local use.
Draft models speed up generation. Gemma‑4‑26B‑A4B‑It‑Assistant and Gemma‑4‑31B‑It‑Assistant are lightweight drafters that predict tokens ahead of the full model. Gemma‑4‑31B‑It‑DFlash works alongside Gemma 4 31B Instruct to accelerate text output. Lucebox‑hub adds DFlash speculative decoding and PFlash speculative prefill for AMD Ryzen GPUs. Kimi‑K2.6‑NVFP4 is a pre‑quantized version of the Kimi‑K2.6 model for Nvidia hardware.
FP16‑FP8‑to‑NVFP4 converts diffusion model files to NVFP4 format for Blackwell GPUs. Torch‑Nvenc‑Compress uses the GPU’s idle video encoder to compress ML data. Shrinking models is key; Qwen3.6‑27B‑GGUF‑MTP keeps multi‑token prediction layers intact in GGUF, while Qwen3.6‑27B‑MTP‑UD‑GGUF pairs Unsloth quantization with grafted MTP layers for speculative decoding.
TextGen morphs into a no‑install portable desktop app for local LLMs. Tokenspeed helps you feel what different tokens‑per‑second rates actually mean while working. Nexus‑BTA bundles image, video, and 3D generation into one local AI studio with an embedded ComfyUI runtime. EasyUI removes node‑graph editing with a clean open‑source web interface for local tools. somni‑comfyui delivers a polished ComfyUI frontend with a chat mode for quick generations.
OpenReader reads EPUB, PDF, and Markdown files aloud while highlighting words in sync. AI Metadata Viewer lets you drag an AI‑generated image to see all creation data instantly. Streamlined‑HF‑Model‑Search is a single HTML file that explores Hugging Face models and quantizations. Merlin‑community strips repeated text from prompts to improve output quality. Opendesk gives agents direct control over a desktop, including screenshot, mouse, and keyboard actions.
IMG‑Dataset‑Refiner turns messy folders into clean, balanced datasets with a visual editor. Caption‑Creator generates high‑quality image captions and tags locally. Diff‑forge automates video dataset preparation for diffusion model fine‑tuning. Cull scrapes, classifies, and sorts AI‑generated images into organized folders. Deepbooru‑tagwalker improves existing tags in image datasets without manual editing.
Bracket runs many short training experiments in parallel to find the best fine‑tuning hyperparameters. Anima‑TrainFlow is a single‑page desktop tool for training LoRAs on the Anima 2B model. ControlLight brightens low‑light photos with a simple slider while maintaining quality. FP‑Background_Obliterator pairs AI background removal with a full layer‑based editor. ShrinkComfy compresses ComfyUI PNG outputs to WEBP or JPG while preserving workflow metadata.
ComfyUI-SmartPromptCrafter is a new node that builds optimized prompt pairs for any checkpoint you load, automatically matching the correct token style. The tool reads your model’s architecture directly and turns […]
RebelsPromptEnhancer is a new ComfyUI node pack that rewrites short text ideas into detailed prompts using a lightweight 4-billion-parameter language model, running entirely on your own hardware. It works offline […]
Orion4D_generative_paint is a new custom node for ComfyUI that adds a full-featured painting interface you can open directly in your browser. It lets users draw, layer, mask, and compose images […]
Step-3.7-Flash is a 198-billion-parameter vision‑language model that uses a sparse mixture‑of‑experts design to activate only about 11 billion parameters per token. It handles images and text natively through a 1.8‑billion‑parameter […]
LocateAnything-3B is a new vision‑language model from NVIDIA that finds and marks objects, text, or interface elements in images based on simple text prompts. Instead of predicting coordinates word‑by‑word like […]
SupraLabs released Supra-50M, a tiny 50-million-parameter language model that punches above its weight class. Trained from scratch on 20 billion tokens, it beats much larger models like GPT-2 on specific […]
Qwen3.6-35B-A3B-Uncensored-Genesis-V2-APEX-MTP-GGUF is a fully quantized, refusal-free language model that packages the original Qwen3.6‑35B‑A3B MoE architecture into ready‑to‑run GGUF files. This release combines APEX and MTP‑APEX quantization formats with a numerical […]
hipEngine is a new local inference engine release for AMD RDNA3 GPUs that runs large language models without PyTorch. The v0.2.1 alpha from shisa-ai delivers fast, ROCm-native performance for Qwen […]
A new engine brings the MiniCPM-V-4.6 vision-language model to a $100 edge board. The MiniCPM-V-4.6-OrangePi project is a from-scratch C++ inference engine that runs the full 4.6B-parameter multimodal model on […]
MiMo-V2.5-coder-Q2 is a text-only GGUF build of the MiMo-V2.5 model, specifically quantized and tested for English-language coding and tool-calling tasks. This Q2_K_S quant was iteratively calibrated to preserve syntax precision, […]
The newly released Qwen3.5-27B-uncensored-heretic-v2-Native-MTP-Preserved is a modified version of Alibaba’s Qwen3.5-27B model that removes most content restrictions while keeping its performance nearly identical. This release preserves all 15 Multi-Token Prediction […]
Keye-VL-2.0-30B-A3B is a new open-source multimodal model designed to understand long videos and perform agent tasks like code execution and web search. It uses a sparse attention mechanism called DSA […]
MOSS-TTS-v1.5 is an upgraded open-source text-to-speech model from the OpenMOSS team, building on their earlier 1.0 release. It keeps zero-shot voice cloning, long-form generation, and multilingual capabilities while delivering more […]
ErniePEUnleashed is a fine-tuned prompt enhancement model that transforms short ideas into richly detailed, spatially structured descriptions for AI image generation. It pays attention to foreground-midground-background layering, lighting logic, camera […]
BangtrixToolkit is a custom node collection for ComfyUI that puts a real-time hardware monitor overlay directly onto the canvas and adds a universal prompt translator. The overlay shows GPU load, […]
ComfyUI-ialhabbal is a new all-in-one custom node pack for ComfyUI that bundles eight distinct tools into a single install. The release adds capabilities like interactive prompt review, batch image loading, […]
ComfyUI_KleinTiledUpscaler is a custom node for ComfyUI that performs creative image upscaling using a tiled, inpainting-based approach designed specifically for the Flux2.Klein model. Rather than just sharpening existing pixels, the […]
IMG-Dataset-Refiner version 4.3 is a free local application that turns messy image folders into clean, training-ready datasets. It provides a visual workspace for editing captions, removing duplicates, and balancing image […]
The comfyui-artius-browser extension adds a fast, sidebar-based asset manager directly to ComfyUI. It lets users preview, search, and drag images, videos, audio, 3D models, and workflow files straight into native […]
ComfyUI-Yedp-Action-Director is a custom node that puts a fully interactive 3D viewport right inside a ComfyUI workflow. The V9.3 update delivers physical path tracing, HDRI lighting, and native Gaussian splatting […]
Nexus-BTA is a local AI studio that bundles image, video, workflow, and experimental 3D generation into one interface powered by an embedded ComfyUI runtime. The update tightens template handling, expands […]
ComfyUI-FeatherOps is a new custom node for ComfyUI that accelerates diffusion model inference on AMD RDNA3 and RDNA3.5 graphics cards like the Strix Halo. It uses a hand-written HIP kernel […]
ComfyUI-PiD is a new custom node that brings NVIDIA’s Pixel Diffusion Decoder (PiD) directly into ComfyUI workflows. Instead of using a traditional VAE to decode latent images, the tool performs […]
ScreenDiffusion is an open-source tool that instantly reimagines your desktop screen as living art through image-to-image AI rendering. Version 0.2 brings a major code refactoring that improves stability and real-time […]
ComfyUI-Anima-Style-Nodes is a new ComfyUI custom node that lets you visually browse and apply anime artist tags, character tags, and style references directly inside your generation workflow. It replaces the […]
Comfyui-Anima-Regional-Conditioning is a custom node for ComfyUI that brings regional text conditioning to Anima image generation models. It works by routing cross-attention so that masked parts of the image only […]
Vlo v0.2.0 is a free, open-source video editor that brings ComfyUI-powered generative AI directly into a timeline-based editing workspace. This new alpha release adds capabilities like mask algebra, draggable motion […]
ControlLight is a new open-source model that lets you brighten low-light photos using a simple slider, adjusting the enhancement strength from subtle to full correction. The system builds on the […]
Qwen3.6-27B-OBLITERATED is a modified version of the Qwen3.6 language model where the built-in refusal behaviors have been surgically reduced through direct weight editing. This 26.9-billion-parameter release from OBLITERATUS uses a […]
Microsoft has released Lens-Turbo, a distilled version of its Lens text-to-image model that can generate high-quality pictures in just four processing steps. Lens is a 3.8-billion-parameter foundational model designed from […]
Microsoft has released Lens, a 3.8-billion-parameter text-to-image model that generates high-quality images with much lower training compute requirements than larger alternatives. It outperforms or matches 6B+ parameter models on standard […]
Nvidia has released PiD, a pixel diffusion decoder that speeds up high‑resolution image generation from latent models. It reformulates the standard decoder as a conditional diffusion model, denoising directly in […]
Nemotron-Labs-Diffusion-14B is a 14-billion-parameter language model that can generate text using standard autoregressive (AR) decoding or a faster diffusion-based parallel method, all within the same model. Switching attention patterns lets […]
Jackrong has released Qwopus3.6-27B-v2-MTP-GGUF, a quantized version of the new Qwopus reasoning model that uses multi-token prediction to speed up text generation. The original Qwopus3.6-27B-v2-MTP model was fine-tuned from Qwen3.6-27B […]
Hy-MT2-1.8B is a new open-source translation model from Tencent that handles 33 languages and can follow detailed instructions. It belongs to a family of fast-thinking translators built for real-world text […]
Tencent has released Hy-MT2-30B-A3B, an open-source multilingual translation model that uses a mixture-of-experts design to deliver fast, high-quality results. It handles translation among 33 languages and can follow complex instructions […]
NuExtract3 is a new 4-billion-parameter vision-language model that extracts structured data from documents and converts them into Markdown. It handles text, images, or both at once, making it suitable for […]
MiniCPM5-1B is a small 1-billion-parameter language model designed to run locally on personal devices and in low-resource settings. The same checkpoint can work as a fast everyday assistant or a […]
LongCat-Video-Avatar-1.5 is a new open-source model for generating talking avatar videos from audio paired with text or image references. It produces realistic human characters, stylized animations, and coordinated multi-person conversations […]
BigStationW has released ComfyUi-Untwisting-RoPE, a custom node for ComfyUI that brings training-free style transfer to diffusion transformer (DiT) models. The tool tackles a common problem where shared attention mechanisms accidentally […]
OpenReader v3.0.0 is a self-hosted web application that reads EPUB, PDF, TXT, Markdown, and DOCX files aloud while highlighting each word in sync. It functions as a private document reader […]
The Gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic is a fine-tuned version of Google’s Gemma 4 31B instruct model that cuts content refusals dramatically while sharpening its writing style. It starts from an already decensored base, […]
G4-MeroMero-31B-uncensored-heretic is a newly released language model that strips away almost all refusal behaviors. It builds on a fine-tuned version of Google’s Gemma 4 31B designed for storytelling and roleplay. […]
Gemma-4-Gembrain-31B-it-uncensored-heretic is a stripped-down version of Google’s Gemma 4 31B instruct model that says “no” far less often. It was built with the Abliteration technique to remove most safety refusals […]
SmallCode is a terminal-based coding agent for local language models between 8B and 35B parameters. It operates fully on your own hardware, keeping code private and avoiding cloud costs. Its […]
BitCPM4-CANN-8B is a new 8-billion-parameter language model that compresses its weights to just three possible values, cutting memory use by roughly six times compared to full-precision versions. The model was […]
The new cross-encoder, Ettin-Reranker-1b-V1, scores pairs of text to reassess search results and boost retrieval quality. It is a 1-billion-parameter transformer that handles sequences up to 7,999 tokens long. The […]
FP-Background_Obliterator is a new open-source tool that combines advanced AI background removal with a full layer-based editing environment, all running locally on your machine. It provides both a rich desktop […]
The new Streamlined-HF-Model-Search is a single browser-based HTML file that acts as a 4-level explorer for Hugging Face models and their quantizations. It queries the public Hugging Face API with […]
The open-source release of Command-A-Plus-05-2026-Bf16 delivers a massive 25‑billion‑parameter active model (218B total) that processes both text and images. It supports a 128K‑token context window and can generate up to […]
The ComfyUI-Workflow-Finder is a new desktop tool that lets you search through your local collection of ComfyUI workflow files using plain English descriptions. It indexes workflows by their node structure, […]
ComfyUI-Magos-Nodes is a professional node pack that adds a full skeleton editor, retargeter, and renderer directly into ComfyUI. It lets you extract body, hand, and face keypoints from video, adjust […]
The somni-comfyui project by searcc delivers a polished, streamlined web interface for ComfyUI that works on desktop and mobile. It offers a Gemini-style chat mode for quick generations and a […]
Comfyui-Mesh is a new open-source ComfyUI node pack that splits diffusion model inference across two Nvidia GPUs. The system compresses the model’s internal activations using each GPU’s dedicated NVENC video […]
Comfyui-controlfoley brings the ability to generate synchronized foley sound effects directly into the ComfyUI node-based interface. It can produce time-matched audio like footsteps or door slams from silent video, still […]
WorkflowX-Configurator is a ComfyUI custom node package that brings selectable workflow profiles to complex image and video generation graphs. Instead of duplicating entire workflows or manually swapping parameters for each […]
ComfyUI-Olm-Liquify is a custom node that adds an interactive image warping editor to ComfyUI, modeled after Photoshop’s Liquify tool. It enables push, pull, twirl, pinch, expand, and smooth brushes on […]
ComfyUI-SPEED is a new custom node for ComfyUI that speeds up image generation by using a technique called Spectral Progressive Diffusion. The node can nearly double sampling speed by starting […]
ComfyUI-Safe-Chunked-Image-Blend is a new custom node for ComfyUI that gives users explicit control over how image batches are resized and blended. It replaces the standard blend to prevent hidden CPU […]
AsymFLUX.2-klein-9B is an adapter that lets the FLUX.2 klein Base 9B model create images in raw pixel space, bypassing the usual VAE (decoder) step. It uses an asymmetric flow method […]
ComfyUI-Angelo merges an image sampler and inpaint refiner into a single custom node for ComfyUI. The tool lets users click or paint directly on a generated image to fix specific […]
Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced is a version of Google’s Gemma 4-26B model with all refusal mechanisms removed while keeping the original capabilities fully intact. This release candidate scored zero refusals across 465 standard […]
A new open-source model called SANA-WM Bidirectional can generate minute-long, 720p videos from a single starting image and a text prompt. It uses a 2.6B-parameter diffusion transformer to synthesize smooth […]
Nandi-Mini-600M-Early-Checkpoint is an early-stage preview of a compact 600-million-parameter language model trained from scratch with strong support for English and 11 Indic languages. The checkpoint shows the model’s progress after […]
Intern-S2-Preview is a 35-billion-parameter scientific multimodal model that analyzes text, images, and time-series data while calling external tools. It continues pretraining from Qwen3.5 and undergoes a full training chain from […]
Ring-2.6-1T is a newly released trillion-parameter reasoning model designed for complex, multi-step tasks in real-world settings. It moves beyond simple question answering to handle continuous agent workflows, tool use, and […]
Fara-7B is a new open-weight computer use agent that understands screenshots and text to complete multi-step web tasks. It takes a high-level goal like “book a restaurant” and plans and […]
Pixal3D is a new open-source model that turns a single image into a detailed 3D asset with high fidelity. It goes beyond typical generation methods by creating a direct pixel-to-3D […]
Marlin-2B is a new open-source video language model that extracts structured descriptions and second‑precise timestamps from video footage. It answers the two questions developers most often ask about a video: […]
Qwopus3.5-9B-Coder-GGUF is a compressed, ready-to-run model file that brings an experimental 9‑billion‑parameter coding agent to local machines. It specializes in writing, debugging, and refactoring code, and can call tools like […]
Sapient Intelligence has released HRM-Text-1B, a 1-billion-parameter language model that uses a new dual-timescale architecture instead of a standard transformer. The model processes information through two recurrent loops — a […]
Lance is a new open-source AI model that handles image and video tasks like understanding, generation, and editing all in one place. It was trained entirely from scratch with only […]
Nvidia has released Kimi-K2.6-NVFP4, a pre-quantized version of Moonshot AI’s massive Kimi-K2.6 language model that runs efficiently on Nvidia GPUs. This is a ready-to-deploy inference model that handles text, images, […]
LTX-2.3 Upscale IC Lora is a generative refinement adapter for the LTX Video 2.3 model that turns soft or low-resolution clips into cleaner, more detailed footage. Instead of simply stretching […]
The VR-360-Outpaint-LTX2.3-IC-LoRA is a proof-of-concept adapter that turns standard widescreen video clips into full 360-degree equirectangular footage for VR viewing. It operates as an IC-LoRA on top of the LTX-2.3 […]
SYSTMS-FLW-IC-LORA-LTX-2.3 is a new LoRA adapter that gives the LTX Video 2.3 model the ability to create smooth shot-to-shot transitions. The method works by placing a plain gray frame between […]
The LTX-2.3-Dearchive-Lora is a LoRA adapter for the LTX-2.3 video model that transforms real archive footage into footage that looks like it was shot recently. It learned to undo many […]
Obscura Remova is a new video-to-video LoRA adapter that removes visual obstructions from existing footage. The model clears away haze, smoke, foreground objects, and partial occlusions to reveal the scene […]
SenseNova-U1-A3B-MoT is a new open-source vision-language model that handles image understanding, generation, and editing through a unified architecture without relying on separate visual encoders. This release belongs to the SenseNova […]
The Opendesk framework gives any AI agent direct control over a desktop computer—screenshots, mouse, keyboard, and app interaction—just like a real person. It works across macOS, Linux, and Windows, turning […]
The studiomi300 pipeline turns a single text prompt into a complete 30-second cinematic reel, complete with consistent characters, music, and voice-over. It strings together multiple large AI models—a director, image […]
MusiCue is an open-source tool that converts songs into typed, timeline-based cues for driving animation and show-control software. Developer cedarconnor built it to break audio into separated stems, beats, drum […]
The open-source project comfyui-node-canvas introduces a local GUI app called ComfyUI Node Builder that helps people create custom nodes for ComfyUI without manually writing all the repetitive code. It provides […]
OmniNFT is a set of LoRA adapters that fine-tune the open-source LTX Video model to produce better-aligned audio and video. Using reinforcement learning, it guides the generation process so that […]
DramaBox is a text-to-speech system that turns scene descriptions and dialogue into expressive speech, complete with laughs, sighs, and pauses. It can clone a speaker’s timbre from just a 10-second […]
ComfyUI-PlagueKind-Nodes is a custom node for ComfyUI that unifies image and mask resizing in a single step. It offers multiple scaling modes, preserves aspect ratios, and ensures masks stay perfectly […]
ComfyUI-DramaBox is a new custom node pack that brings ResembleAI’s expressive text-to-speech system directly into ComfyUI workflows. It turns text prompts into spoken audio using the LTX-2.3 audio diffusion model, […]
Anima-TrainFlow is a simple, single-page desktop tool for training LoRA adapters on the Anima 2B image generation model. It puts every setting you need right in front of you, skipping […]
A new quantized file for DeepSeek V4 Flash, called Deepseek-V4-GGUF, shrinks the massive AI model so it can run on high-end consumer hardware. It’s a set of GGUF format files […]
Emo is a new mixture-of-experts language model designed so groups of experts naturally specialize in specific topics during training, rather than requiring human labeling. The main release from the Allen […]
Llmfan46 has released Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved, a modified version of Qwen3.6-35B-A3B that cuts unwanted refusals by 88% while keeping all 19 multi-token prediction (MTP) layers fully intact. The model uses an abliteration […]
Beellama.cpp is a fork of the popular llama.cpp project that squeezes extra speed and memory efficiency out of local GGUF model inference. It adds DFlash speculative decoding, TurboQuant KV‑cache compression, […]
ds4.pinokio is a new launcher and browser interface that brings the massive DeepSeek V4 Flash AI model to Apple Silicon Macs. It builds on the ds4.c Metal-only inference engine created […]
The NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B-BF16 release packs three distinct reasoning model sizes — 30 billion, 23 billion, and 12 billion parameters — into a single checkpoint file. Rather than requiring separate training runs, […]
Coming across tokens-per-second benchmarks is easy, but truly understanding what "47 tok/s" feels like while you work is much harder. A new open-source tool called Tokenspeed solves this problem by […]
ExLlamaV3 is an inference library that lets you run large language models on consumer graphics cards. It introduces the EXL3 quantization format, which compresses models to very low bitrates while […]
Unsloth has released Qwen3.6-27B-GGUF-MTP, a quantized model file that preserves the multi-token prediction (MTP) layers from Qwen’s latest 27-billion-parameter language model. This GGUF format makes it possible to run the […]
The new release, needle, is a tiny 26-million parameter open-source AI model purpose-built for function calling, or tool use. It interprets a user's plain text query and outputs a structured […]
Lucebox-hub is a collection of hand-tuned LLM inference servers that push consumer GPUs to their limits. The latest release adds DFlash speculative decoding and PFlash speculative prefill for AMD Ryzen […]
Derpy-Turtle-The-Kokoro-Trainer is a Windows GUI that blends Kokoro’s text-to-speech with RVC voice conversion to build better local voice clones. It lets you search for and refine Kokoro voice tensors, train […]
AntAngelMed is a new open-source medical language model designed to assist with clinical reasoning and diagnosis. The model uses a mixture-of-experts architecture, activating only 6.1 billion of its 100 billion […]
Ovis2.6-80B-A3B is a new multimodal AI that pairs vision and language through a mixture-of-experts design, keeping it fast and efficient. It can examine high-resolution images, long documents, and even videos, […]
TextGen is a desktop application that runs large language models locally on your own computer. The latest update transforms the project from a web interface into a no-install portable app […]
Merlin-community is the free, open-core release of a deduplication engine that strips repeated text chunks from AI prompts before they reach the model. The tool now ships with a transparent […]
The ComfyUI-lora-FindingLora custom node replaces ComfyUI’s stock LoRA loader with fuzzy search, bookmarking, trigger word storage, and one-click LoRA stacking. It allows you to find any LoRA from a large […]
ComfyUI_ShowMe is a new canvas overlay extension that lets you draw notes directly onto your ComfyUI workflow without affecting how it runs. Created by developer SKBv0, this tool solves a […]
A small, privacy-first web tool called AI Metadata Viewer now gives anyone a quick way to read all the hidden creation data tucked inside AI-generated images. You simply drag a […]
The Ace-Step-1.5-XL-Concept-Sliders are a new set of directional LoRA files that let you nudge an AI music generator toward or away from specific audio traits. These sliders work with the […]
Cull is a single-machine curation engine that automatically scrapes, classifies, and sorts AI-generated images into organized folders. It runs entirely on your local hardware without needing Docker, Redis, or a […]
A new open-source tool called Bracket aims to replace guesswork with hard numbers when fine-tuning image-generation models. It automates the trial-and-error loop by running many short training runs at once, […]
Lightricks released LTX-2.3-22b-IC-LoRA-LipDub, a new open-source adapter that replaces speech and lip motion in existing videos. The adapter works with the LTX-2.3-22b video model to generate synchronized audio and lip […]
Scenema-Audio is a new open-source model that clones voices and generates speech with emotional acting, scene sounds, and zero-shot identity transfer. It doesn’t just read text aloud—it interprets stage directions […]
The FP16-FP8-to-NVFP4 tool by developer Thenotrealuser is a Windows-based converter that turns FP16 or BF16 diffusion model files into NVFP4 format for Blackwell GPUs. It targets popular image generation models […]
Gemma-4-31B-It-DFlash is a drafter model that works alongside Google’s Gemma 4 31B Instruct to speed up text generation for local deployments. Instead of generating tokens one by one, it uses […]
An AI model specifically trained to support psychologists and physicians working with obesity patients has been released on Hugging Face. Leanly_AI is a large language model that provides structured, evidence-informed […]
Anima Base v1.0 is a 2 billion parameter text-to-image model built to generate anime-style and other non-photorealistic artwork. It creates illustrations with a focus on anime concepts, characters, and styles, […]
Havenoammo’s new Qwen3.6-27B-MTP-UD-GGUF package combines Unsloth Dynamic 2.0 XL quantization with grafted Multi-Token Prediction (MTP) layers for the Qwen3.6 27B model. This format enables speculative decoding, where the model predicts […]
Supertonic-3 is a lightweight text-to-speech system that runs entirely on your device using ONNX Runtime, with no cloud calls needed for synthesis. This open-weight release expands language support from 5 […]
HiDream-O1-Image is an open-source image generation model that creates, edits, and personalizes visuals without relying on separate compression tools. It uses a Pixel-level Unified Transformer to process raw pixels, text, […]
MiniCPM-V-4.6 is a new open-source multimodal model that brings image and video understanding directly to smartphones and small computers. It answers questions about photos and video clips without a cloud […]
ComfyUI-XAV-Google-Sheets is a new custom node package for ComfyUI that lets you pull text directly from a public Google Sheet. It loads a shared spreadsheet as a data table and […]
Flux.2-Klein-Loras is a fresh bundle of style adapters for the Flux.2 Klein 9b distilled image model. It packs multiple LoRA files that let users generate or edit pictures in distinct […]
Causal-Forcing is a new training method that distills large autoregressive video models into efficient ones that can generate video in real-time. The approach bridges a structural mismatch between teacher and […]
ComfyUI-ReferenceLatentPlus is a custom node that completely replaces ComfyUI’s original ReferenceLatent tool. It gives creators per-image control over how reference images influence the final output during image and video generation. […]
The Comfyui-Clippy-Reloaded add-on by Shootthesound lets you paste images directly from your computer’s clipboard into a ComfyUI workflow. It grabs whatever image you’ve copied — a screenshot, a browser image, […]
The Qwen3.5-9B-DeepSeek-V4-Flash-GGUF is a compressed language model that packs DeepSeek-V4’s advanced reasoning into a 9-billion-parameter package for local use. It converts the full model into the GGUF format, so it […]
The Qwen3.6-27B-Heretic-Uncensored-FINETUNE-NEO-CODE-Di-IMatrix-MAX-GGUF package delivers an uncensored, performance-enhanced version of Qwen’s latest 27B model in highly accurate compressed formats. This release strips away the original model’s refusal behavior, cutting the refusal […]
Google just dropped a new tool that makes its open-source AI models run much faster. The Gemma-4-26B-A4B-It-Assistant is a lightweight draft model that predicts tokens ahead of the main AI, […]
Zyphra’s new ZAYA1-8B is a compact mixture-of-experts (MoE) language model with only 760 million active parameters drawn from a total of 8.4 billion. It handles detailed long-form reasoning, especially for […]
The Gemma-4-31B-It-Assistant is a lightweight draft model built to speed up text generation when paired with Google’s full Gemma 4 31B instruction-tuned model. It uses a technique called speculative decoding […]
ShrinkComfy is a small Windows application that shrinks ComfyUI PNG outputs into WEBP or JPG files without breaking drag-and-drop workflow recovery. It copies all prompt and workflow metadata from the […]
ComfyUI-Fayens is a new collection of custom nodes for ComfyUI that streamlines face swap workflows from start to finish. It automatically extracts clean face crops, generates accurate masks, and prepares […]
EasyUI is a newly released open-source web interface that removes the need to edit complex node graphs when working with local AI tools. Instead of clicking through spaghetti-like connections, you […]
Torch-Nvenc-Compress is a new open-source library that uses a GPU’s idle video encoding chip to compress machine learning data. Instead of serving video streams, the normally dormant NVENC hardware compresses […]
Deepbooru-tagwalker is a lightweight desktop tool that improves the accuracy of existing tags in image datasets. Instead of opening images and editing tags one by one, you select a single […]
Diff-forge is a new open-source tool that automates the tedious work of preparing video datasets for diffusion model fine-tuning. It runs entirely on your own machine, providing a visual browser-based […]
Caption-Creator is a fast, portable GUI tool that runs entirely on your local machine to generate high‑quality image captions and tags. It helps users build custom datasets for AI model […]
Tritant has released Ace-Step-1.5-Api-server-UI, a visual interface that turns the ACE-Step 1.5 music generation model into a full-featured local studio. The tool wraps the model’s API server in a single […]
Walkyrie-1.3B-v1.0 is a new text-to-image model that turns written prompts into 1024×1024 pixel images. It was rebuilt from an existing video-generation model after its language-understanding component was trimmed down to […]
UltraReal_FineTune_Anima is an experimental full model fine-tune of the Anima_preview1 image generator, aimed at delivering more realistic photo-style outputs. It produces strikingly varied visuals, from analog film grain to clean […]
NVIDIA has released Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4, an open multimodal AI model that simultaneously processes video, audio, images, and text. The 31-billion-parameter system uses a hybrid Mamba2-Transformer design that activates only about 3 […]
LTX2.3-10Eros is a new image-to-video model merge that turns a single still image into a short motion clip. Unlike standard weight blending, it combines layers from different training steps to […]
The new release Hy-MT1.5-1.8B-1.25bit is a heavily compressed language translation model designed to run entirely on your phone, no internet needed. It shrinks a powerful 1.8 billion parameter system down […]
IBM has released Granite-4.1-30b, a 30‑billion parameter instruct model that brings upgraded tool calling and long‑context abilities to the open‑source community. It can summarize text, answer questions, write code, and […]
Qwen-2512-portrait is a LoRA adapter for the Qwen-2512 image model that sharpens human portraits with lifelike facial detail and natural skin. It addresses the common plastic-smoothing problem by preserving realistic […]
ComfyUI-gonztok_nodes is a new suite of custom nodes that swaps clunky text inputs for fast, visual pickers in ComfyUI. Version 1.2.0 introduces modal popups for selecting images and LoRA models, […]
Phosphene is a free desktop panel that turns text or images into video clips with synchronized audio, running entirely on Apple Silicon Macs. It wraps the LTX 2.3 model through […]
A new custom node called Comfyui_VideoCombine_Plus gives users extra tools when combining images into video files inside ComfyUI. The node extends the built-in video combine feature with sound volume control, […]
Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16 is a high-precision, uncensored large language model designed to follow instructions without refusal. It removes the "safety tax" found in standard models, allowing for more direct reasoning and compliance. […]
Mistral-Medium-3.5-128B is a dense flagship model designed to handle complex reasoning, coding, and instruction-following tasks. It serves as a unified replacement for several previous models released by the company. The […]
The Ling-2.6-1T is a new open-source language model from inclusionAI that packs a trillion parameters. It is designed to handle complex tasks like coding, reasoning, and tool calling while keeping […]
IBM has released Granite-4.1-8b, a new large language model designed for instruction following and chat tasks. This 8B parameter model is built to handle long contexts and complex requests. It […]