The AI & Creative Tools Magazine
New models, image and video generation, music generation, 3D generation, and the workflows creative people are building around them.
LFM3 is Liquid AI's new foundation model family: six 2B-class models for text, vision, audio, omni and calibrated decision-making, all branching from one 1.7B backbone, built to be fine-tuned locally and deployed on-device.
Signal Board
Hugging Face models and datasets, 2026-10-27 against 2026-10-26. Today’s snapshot was taken locally at 12:40 UTC because the scheduled run did not fire at all — both of its cron times had passed with no snapshot on disk. OpenRouter text is a rolling window labelled 2026-10-26. The image table still reports the week of 19 October against 12 October, unchanged from yesterday.
625,013 downloads, 604 likes
The CC BY 4.0 3D asset library whose licence is invisible to anyone filtering by licence tag. One like per 1,034 downloads, against Event6DBlender at one per 56,000 - volume and engagement rising together, which is what adoption looks like.
35 likes on 155 downloads
One like per four downloads. Created 29 September, almost nobody has found it, and nearly everyone who has marked it - the shape of a repository reaching people who were already looking. Suggests simready is a programme, not an upload.
750,694 downloads, 1,790 likes
The ComfyUI-packaged build of FLUX 3, created 14 October. The convenience distribution is where the attention has landed rather than the base weights, which is the same pattern driving the GGUF repacks up the model table.
9,509 likes, 2.16M downloads
The highest like count on either board, still climbing three months after creation. Video generation accumulating long-run engagement rather than a launch spike.
Tess-4.5-27B +111%, Sage-A1.5-Q8_0 +111%, Spark-X2.6-4B +51%
The fastest-growing models by downloads are quantisations of other people’s weights, not new models. The download column is measuring distribution convenience; read the likes column for interest.
Unchanged from yesterday
Still comparing 19 October against 12 October. The glm-5.3-flash fall to 21.8% and the rise of Others to 35.8% are yesterday’s findings; nothing in this table moved today, and the week of 26 October is still open so its shares are not quotable.
Zero in all 260 rows
Still collected so the day it changes is visible, but it has not been evidence of anything for a fortnight.
`nvidia/simready-assets` — the 2,575-asset simulation-ready 3D library whose CC BY 4.0 licence is buried one directory deep in every asset rather than declared on the repository — has gone from **235,040 downloads on 24 October to 625,013 today**. That is **+166% in three days**, including **+59% in the last twenty-four hours alone**. Likes over the same window: **68 → 134 → 373 → 604**, a nine-fold rise.
The reason to trust this one is the ratio we introduced two days ago to kill a different story. At 625,013 downloads against 604 likes, simready is running **one like per 1,034 downloads**. `mickeykang/Event6DBlender`, which posted a larger percentage spike on Monday, was at one per 56,000 — fifty times thinner. For a closer comparison, `scenario-labs/3d-collection` sits at one per 90, which is what a small, engaged audience looks like. A dataset acquiring both volume and likes in proportion is being used, not crawled.
A second NVIDIA repository is now visible behind it. **`nvidia/simready-central-library`** gained **+12 likes (23 → 35, +52.2%)** on just **155 downloads** — one like per four downloads, which is almost nobody downloading it and nearly everyone who does marking it. Created 29 September. That is the signature of a repository being found by people who already know what they are looking for, and it suggests the simready push is a programme rather than one upload.
On the model side the creative-tooling rows moved quietly and in the right direction. **`black-forest-labs/flux-3-image-klein-comfy`** added **+137 likes (1,653 → 1,790)** at 750,694 downloads, thirteen days after it was created — a ComfyUI-packaged distribution of FLUX 3, which is to say the convenience build is where the attention is. **`Lightricks/LTX-2.5`** added **+114** to reach **9,509 likes** against 2.16M downloads, the largest like count on either board.
Everything else at the top of the download table is quantisation. `migtissera/Tess-4.5-27B-GGUF` **+111%**, `InternScience/Sage-A1.5-Q8_0-GGUF` **+111%**, `InternScience/Sage-A1.5-Q4_K_M-GGUF` **+73%**, `unsloth/Ornith-1.5-122B-GGUF` **+58%**, `XHToken/Spark-X2.6-4B-GGUF` **+51%**. Six of the nine fastest-growing models today are GGUF repacks of somebody else’s weights — the sixth consecutive day this pattern has held, and a reminder that the download column measures distribution convenience at least as much as it measures interest in a model.
Two things we are deliberately **not** reporting as news. The **image-generation table has not advanced**: it still compares the week of 19 October against 12 October, the same window we analysed yesterday, so `z-ai/glm-5.3-flash-20260826` at 21.8% and `Others` at 35.8% are yesterday’s figures and not today’s movement. And the OpenRouter text file is a **rolling window labelled by date**, so `minimax/minimax-m3.1-flash-20261019` reading **+86% to 1,467B tokens** describes the window sliding forward as much as the model climbing; we quote it as a direction, not a measurement.
Ollama’s pulls column has now been zero in all 260 rows for **fifteen** days. We keep collecting it so the day it changes is visible, but it has stopped being evidence of anything.
Genie 4, announced October 27, is DeepMind's first real-time world model to generate audio natively alongside interactive, explorable environments. A research preview for trusted testers, it leans on Gemini for scene reasoning and carries the same drift and consistency limits every world model still has over long sessions.
We said yesterday that nvidia/simready-assets had no licence. That was wrong, and the way it is wrong matters: the licence is CC BY 4.0, 2,575 times over, one directory deep inside each asset. The repository itself declares nothing — so anyone filtering Hugging Face by licence will never find the most permissively licensed industrial 3D library on the platform.
Gradium, a Paris voice-AI company spun out of Kyutai, has released the Phonon family: tiny streaming text-to-speech models that run in a browser tab, on a phone or on a laptop CPU. Coverage called it open-sourced under Apache 2.0. Checking every repository in the family gives a more interesting picture — the permissive licence covers exactly one of seven models, and that one has more downloads than the other six combined.
FIBO 2.0's pitch is clean provenance: every training image licensed, indemnified output, no scraped web. That is a real answer to the question hanging over image generation. But the model you would use to act on it is non-commercial, gated, and reliant on a Google API for its own prompt format.
mcp-for-blender-pro puts a chat workspace inside Blender with scene inspection, material edits and asset search, and lets you configure a team chain where a manager splits the work, workers execute and reviewers approve. The release is a single Windows x64 archive, the honest caveats are in the README, and the fork count is doing something strange.
MOSS-Image is a 5B DiT built around two things most open image models treat as afterthoughts: legible rendered text, and explicit control over where things go. You can prompt it in natural language, in tag sequences, or by handing it a JSON layout with bounding boxes. The ComfyUI plugin ships inside the repo.
The interesting number is not 5.6 billion parameters. It is that pairing INT8 ConvRot quantisation with the step-distilled checkpoint makes 1024×1024 generation 4.3× to 9.3× faster than the BF16 baseline, with the biggest gains on consumer cards.
Drama 3 is built for the thing most TTS models handle badly: a scene with several people in it. Speaker tags sit inside the text, up to sixteen per passage, and previous takes can be passed back in as context so a fix inherits the delivery of the line it replaces.
ComfyUI shipped NVIDIA's RTX video effects as native nodes on 13 October. The upscaler and the frame interpolator are the headline, but the HDR node is the one that puts a colour-managed display pipeline inside a generative workflow.
OpenBMB's MiniCPM-V 4.7 35B-A3B claims to be the first open-source MoE model that runs on-device without compromising performance: 13+ tokens/s in 3 GB of memory on an iPhone 17 Pro Max. Apache 2.0, with day-zero support for Rockchip, Houmo, Intel, AMD and Arm, and 302 likes on 630 downloads six days after release.
In the week of October 5, z-ai/glm-5.3-flash-20260826 moved from 7.3% of OpenRouter's image-generation tokens to 55.9%, a gain of 48.6 points. Every other named model lost share, and the unnamed long tail lost 23.4 points. Consolidation on this scale, this fast, has not happened in the tracked period.
Odyssey has opened a public research preview of Odyssey-3. Type a description, choose first or third person, and move through a world the model generates frame by frame in response to what you do. The free demo runs Odyssey-3 Flash; the Physics-IQ record claim comes from a best-of-eight selection on a single run.
Qwen-Image-2.1-Turbo is an accelerated checkpoint of Qwen-Image-2.1 that does text-to-image and image editing in eight denoising steps. It ships its own sampling schedule so there is no scheduler to configure, runs at CFG=1, and reuses text and reference-image context across steps. It needs a Diffusers build from source.
Phocinae-Largha-150M-v1 appeared on Hugging Face on October 8, 2026: a 144.3M-parameter ModernBERT-based decision model that takes a state plus typed questions and returns calibrated answers with no generation at all. 21 ms per decision on a 5090, 288.6 MB of weights, and a confidence gate the authors measure as cutting LLM calls 55%.
Google Labs launched Playground on October 7, 2026: an experimental platform where you describe a game in text, test it immediately, and publish it to a community gallery with leaderboards and multiplayer. The interesting part is the announced integration with Unity Spark, which is the only stated path from a prompt to something with professional-grade mechanics.
An embedding turns a file into a list of numbers positioned so that similar things sit near each other. That is the whole idea, and it is enough to make a decade of your own work searchable by description, by reference image, or by hummed melody. This covers the model, the index, the chunking decision, and the mistakes that waste a weekend.
Searching a mixed-media archive normally means a separate embedding model per modality and no way to compare across them. This one embeds text, code, images, video and audio into a single shared space — 270M text backbone plus modular 170M vision and 300M audio encoders, with Matryoshka truncation for 6x smaller vectors.
The open decision models that have arrived in the last fortnight are 26 to 27 billion parameters and need a real GPU. laya is 421M, runs in 33ms on a Tesla T4, speaks 100+ languages, and was trained with strictly proper scoring rules so that honest probability reporting is the only way to maximise its reward. Apache 2.0, pip-installable.
AutoTrust's new decision model entered our tracker tonight at trending rank #9 with 446,527 downloads. It is built on Gemma-4-26B rather than Qwen, which means the open decision-model category has spread across base models — and its adaptive-thinking gate is the first idea in the category that solves a real problem for unattended work.
Our tracker has carried space-bunny-alpha for six days as a number without a story. This week it doubled its share, which makes the stealth-model arrangement itself the story: free anonymous previews that accumulate real usage and real dependency, then vanish on a date nobody chose but the provider.
Kolibri-1 is 78 billion total parameters with 3.46 billion active — 384 experts per layer, six routed plus one shared. German and English, low/medium/high reasoning modes, 262K native context extendable to 1M, and Apache 2.0 on the weights. The licence has a carve-out worth reading carefully.
Upscaling has split into two incompatible jobs — faithfully enlarging what is there, and generating plausible detail that was never there. Picking the wrong one is why upscaled portraits look like strangers. This covers Real-ESRGAN, SUPIR-class generative upscalers, tiling, video temporal consistency, and when not to upscale at all.
Convolution reverb normally means loading someone else's impulse response of a real room. This inverts the workflow: you design the IR, from audio you drag in or play live, and hear the result immediately. Free to use, with WAV export behind a Patreon membership.
JEV-27B-VL's downloads rose 72% today to 308,531, and clef and clef-flash both entered Ollama's popular list. A category that did not exist a month ago now has several entrants and a shared benchmark — the Jev Decision Index, which Typesafe AI built and Cloudflare reports against. That provenance matters, and it corrects something we wrote yesterday.
Clef and Clef-Flash went up on Hugging Face on September 30 and Clef's likes rose from 19 to 632 in a day. It is a Qwen3.8-27B with a joint schema head that scores answer options against evidence, and Clef-Flash answers in 38.8ms. For anyone who needs a model to decide rather than to write, this is the shape of the thing.
The Engram is a sampler groovebox on Kickstarter with small neural audio models built into the hardware. It needs no internet connection, is driven by voice rather than typed prompts, and the developer's stated technique is pushing the models past their intended limits. $770, delivery expected May 2027.
Most document-parsing models are built for clean scans or for phone photographs, not both. TeleOCR — 1.2B parameters, Apache 2.0, from XingChen-AGI — does both with geometry-aware modelling for camera captures. It has been climbing our tracker for two days and was updated yesterday.
PrunaAI released distilled LoRA adapters for Qwen-Image-2.1 on September 23 that generate in 5 or 8 steps rather than 40, with no classifier-free guidance. Downloads jumped 217% in the last 24 hours on Hugging Face. Here is what the distillation actually trades away, and why the step count matters more than the wall-clock number.
Demucs separates a mix into drums, bass, vocals and other, runs entirely on your own machine, costs nothing, and holds up against paid tools. This walks through getting it running — GUI or command line — picking the right model, and the parts of its output you need to listen for before you build on it.
VirtualDJ 2026 adds AI features including ShoutOut, which generates spoken shout-outs and announcements, and Vibescripting. It's a useful window into where generative audio is actually landing in working software — not as a compositional tool but as a labour-saving device for performers, with the questions that raises.
Supersonic Labs' Julia-1 is a 144.3M-parameter model built for classification, routing and scoring rather than language understanding. It takes a state, a question and 2–20 options and returns a choice, an ordered score, or a boolean. Apache 2.0, built on JHU CLSP's mmBERT-small, and explicitly not for knowledge retrieval or multi-step reasoning.
inclusionAI's Ming-Image-0.1-Design is a 6B open-weight text-to-image model aimed at text-heavy visual design: UI screens, infographics and posters. It writes real RGBA transparency and a sibling checkpoint decomposes a design into editable layers. Our tracker has it climbing three days running — but an unmodified ComfyUI still can't load it.
Comfy Router went live September 23: one API for frontier image, video, 3D and audio models, with the same model string and arguments, no new SDK and no new key. Day one covers Seedance 2.5, MiniMax H3, Nano Banana Pro, GPT Image 2, Kling and Black Forest Labs — and every job reports back which provider actually ran it.
ByteDance Seed's SeedVR2 is a one-step diffusion video restoration model, adversarially trained against real data, now natively supported in ComfyUI. It entered our tracker this week at 471,279 downloads. The reason to care isn't resolution — it's temporal consistency, which is where ESRGAN-class upscalers fail on video.
Pixal3D turns a single image into a 3D asset with PBR materials, aligning the generated front face 1:1 with your input via back-projection. Multi-view inference shipped September 1 and ComfyUI core support landed September 8 — feed up to four orbit shots and the Pixal3DMultiViewConditioning node fuses them into one mesh. Comfy-Org's repackaged weights entered our tracker this week at 612,886 downloads.
Apple released LensVLM-9B on Hugging Face: a vision-language model built on Qwen3.5-9B that scans a compressed image representation of text, then selectively expands only the relevant pages back to uncompressed form. Our tracker caught its likes jumping 1,123% in 24 hours — from 13 to 159 — the day after it first appeared.
Qwen-Image-2.1 shipped on September 20 as a ~14.2GB BF16 diffusion model. By September 23, community GGUF quantisations had it running in ComfyUI on cards with as little as 8GB of VRAM — and download data shows the quantised repos, not the official weights, are where the adoption is happening.
Most open models chase math and code. Altworld fine-tuned a 27B model to write the email you've been putting off, released the weights under Apache-2.0, and published a benchmark that measures whether judges can tell a human wrote it.
Alibaba open-sourced Qwen-Image-2.1 on September 20: a 7-billion-parameter unified generation-and-editing model that natively produces and edits RGBA layers for compositing, accepts up to 10 reference images, and supports selection-based local editing and inpainting. Weights are downloadable under a research licence permitting non-commercial use only.
Google released Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking on Sept 15. They hold real-time spoken conversation, understand a camera feed, run tool calls while still talking, and handle 97 languages, at $0.005/min for audio in and $0.018/min out. For artists building conversational installations, that pricing changes what's feasible.
M-A-P released YuE2-3B on September 9 — an open-weights song generation model built around 'white-box' symbolic planning. It writes a composition you can read, play, and change before any audio is rendered, handles zero-shot covers and agentic editing from the same checkpoint, and scores 6.9632 on WildSongBench at best-of-8 against Suno v6's 6.5562. That benchmark is M-A-P's own, which is worth holding in mind.
Released September 4, Viggle-Animate is Viggle Research's first open-weight model: a 33.1B full finetune of MiniMax-H3's ref2va transformer, distilled to three forward passes. It needs no pose skeleton, mask, or prompt at inference — only a driving video and one of its own frames, repainted. 124 frames render in 26 seconds on a B200.
OpenAI's ChatGPT Images 2.5 rolled out September 8 with 50% faster generation, more consistent multi-turn editing, a Sketch tool that converts hand-drawn layouts into finished images, and comment-based edits that let you annotate a specific part of an image instead of describing it in a prompt. Two new API models, GPT-Image-2.5 Flare and Sunburst, ship alongside it.
Runway's GWM Worlds 2 generates a continuous, explorable 720p video-and-audio stream in real time, steered by text actions and camera movement as you go — a step beyond video generation toward simulation you can actually walk through, released just two days after the company's Solaris interface model.
Runway's new Solaris is an 'interface world model' that renders interactive software as a continuous stream of generated video frames instead of executing code — drags, clicks, and voice commands act as prompts, and a 250-participant study found people preferred its responses to a traditional coded interface's 61% to 71% of the time.
ElevenLabs' new Composer tool breaks AI-generated songs into independently editable sections, so changing a chorus no longer means re-rolling the entire track and hoping the rest survives intact.
FLUX Video Upscale, Black Forest Labs' new tool, doesn't just stretch a video to a higher resolution — it regenerates each frame at the target resolution, repairing the smudged faces and gridded textures that generative video is known for.
The Phoney Club, a Rotterdam design studio, uses generative AI to place licensed digital twins of real models into entirely fabricated worlds — while holding a hard line against retouching the model's actual face, arguing authenticity is the one thing AI can't manufacture on its own.
LTX-2.5 adds native multishot generation, auto duration, 4K HDR output, and a cinema-grade rendering pass — and like its predecessor, ships as open weights rather than an API-only product.
Meta's pitch with Glimmer isn't a bigger model. It's a model built to sit on your own machine indefinitely, breaking down tasks and recovering when a step goes wrong, without ever phoning home.
A magic wand to edit just one part of an image, five reference photos composited at once, and now a dedicated meme mode — Grok Imagine's newest update reads like a checklist of what people actually do with image tools.
A model's real adoption curve shows up in what a tracker catches, not in a launch-day headline. MiniMax H3's curve, one week out, still hasn't flattened.
A new model's real adoption signal isn't the download count. It's whether strangers start building tools on top of it before the launch week is even over.
Voice AI has mostly competed on how natural it sounds. xAI's newest model is competing on how little dead air there is between you finishing a sentence and it starting one.
Instead of upscaling a low-res video after the fact, H3 regenerates the whole scene at 2K in-context — recovering fine detail a bolted-on super-resolution pass usually can't.
The rule that's been on the calendar for two years just became enforceable today. Every generative tool covered on this site now has a legal labeling obligation if it touches EU users.
Object removal is a crowded field now — Ideogram just showed up with the top benchmark score and the lowest price, in the same week.
Our daily Hugging Face tracker flagged the ComfyUI repack of Mage-Flow climbing 35% in downloads overnight. The model behind it is small, fast, permissively licensed, and aimed squarely at people who generate images for a living.
Instead of shipping separate models for image, video, and audio, Black Forest Labs trained one network on all three at once — and threw in robot action-prediction as a fourth. Early access opened July 23.
The gap between 'that model exists' and 'that model runs on my GPU' just narrowed. Nunchaku's 4-bit inference is now native to Hugging Face Diffusers — no separate library, no local compilation.
Describe a game in a sentence; get a playable 3D world back. Roblox's Build tool just entered public alpha — mobile-native, AI-generated, and aimed at collapsing the distance between having an idea and shipping it.
Black Forest Labs made its name on FLUX image models. FLUX 3 is a different animal: one architecture that does images, video-with-audio, and robot actions — and, eventually, an open-weight release.
Texturing is one of the slowest steps in 3D. Abstract's InstaMAT update pairs generative materials with a full 2D painting toolset — the bet being that artists want AI to speed the work up, not take the brush away.
The fight over AI music is really a fight over training data. BandLab just bought Aiode — an AI studio whose whole pitch is that 100% of its training audio is licensed and traceable to the musicians who played it.
The interesting move in AI-for-art right now isn't another image generator — it's agents that reach into the tools artists already use. NVIDIA is wiring them into Blender, Houdini, Unreal, and Adobe through an open protocol.
Meta's first in-house image model, Muse Image, can blend several of your photos into one polished result and drop it straight into a chat or story. The launch was fast; the backlash over training data was faster.
Type a prompt or drop in a photo, and get back a 3D world you can move through — walls that stay put, depth you can step into. World Labs is making scene generation a creator tool, and it changes what 'generate an image' can mean.
Most generative AI tools produce an asset. Adobe's creative agent is explicitly built to leave the asset-making to specialized tools it already owns, and instead handle the orchestration between them.
Every long AI video you've seen until now was clips stitched together, seams and all. Seedance 2.5 generates the full 30 seconds in a single pass — and takes 50 reference inputs while doing it.
Sora launched to enormous hype and a viral app-store run. Eighteen months later, OpenAI is shutting the whole thing down — and Disney found out its billion-dollar partnership was dead less than an hour before the rest of us did.
Not every 'open' model is open the same way. FLUX.2's three-tier release is a good case study in reading the license before you build on top of any of them.
Most open video models still leave sound as someone else's problem. LTX-2.3 doesn't, and that one decision is why it's become a reference point for the rest of the field.