Open model trends, October 2026: what creators download

A character feed lives or dies on the model underneath it, and in the week to 18 October 2026 the most downloaded open model for turning a still of your character into a clip was LTX-2.5, at 1,993,003 downloads in the last month on Hugging Face. The most downloaded base image model in the same snapshot, Qwen-Image 2.1, sat at 171,381. That gap is the first of six patterns in this report.

This is a trend report, so every claim is dated and sourced. The numbers come from one snapshot of Hugging Face trending and Civitai's most downloaded list taken on 2026-10-18, and the voices come from public discussion threads. Download counts measure files fetched, not people, not quality and not revenue. Where a number cannot carry a conclusion, the text says so.

Trend 1: video models now lead the download counts

Put the three leading downloads side by side and the order is not close. LTX-2.5 and lightx2v's 8-step build of MiniMax H3 are each above one million downloads in 30 days, while the best base image checkpoints are in the low hundreds of thousands. Both video models are image-to-video capable, which is the job a creator with an approved face needs: start from a still, add motion.

Read the chart with care. Hugging Face counts are rolling 30-day figures, and a video model ships as several large files plus workflow assets, so one user can register more than one download. A video model's count is also inflated relative to an image model that most creators already have. What the chart supports is narrower: attention this month sits with video, and with models that run in ComfyUI. Our own LTX-2.5 walkthrough and the MiniMax H3 Turbo breakdown cover how to keep the face steady in each.

Trend 2: every model now ships a fast twin

Look at the names in the snapshot: Qwen-Image 2.1 and Qwen-Image 2.1 Turbo, Krea 2 Raw and Krea 2 Turbo, MiniMax H3 and its 8-step Turbo build, a 4-step Lightning LoRA for Qwen, and on Civitai a Turbo LoRA for Anima. The pattern is a slow base that you train against and a distilled copy that you render with. The differences matter for a recurring face because distilled models trade variety across seeds for speed, which we explain in distilled vs base checkpoints.

The practical reading for a creator: expect to keep two files per model, and decide per task which one runs. A Turbo draft to pick the pose, the base for the final, is the habit this trend rewards.

Trend 3: a new open model can be a nightly-build problem

MOSS-Image from OpenMOSS is the newest example. Its Hugging Face card lists Apache 2.0 for the weights and inference code, 5B parameters, and a sampling table of 50 steps with CFG 3 and a shift of 5. ComfyUI Wiki dates the release to 16 October and says core support merged after the latest release tag, so an existing install has to be updated. The card gives no VRAM figure and no release date, and ModelScope's launch post describes the Base checkpoint as 8.56B, which does not match the card, so treat the size as unsettled until the lab reconciles them.

The first thread on Comfy-Org's repackaged weights reads like every launch week. One user could not load the workflow by dragging it in. Another got desaturated images with curls until they removed the --use-sage-attention flag. Others needed the nightly update script or the master branch, while one reported it working on stable but not on nightly. If your feed depends on a pinned ComfyUI, a launch-week model is a test environment, not a production dependency. The missing nodes guide covers the usual first fix.

Trend 4: Civitai and Hugging Face are measuring different crowds

A September analysis by one Japanese author, who reads Hugging Face trending daily for a week and compares it with Civitai's weekly ranking, found Hugging Face led by Krea 2 while Civitai's top spots stayed with Anima, Illustrious and NoobAI. The author's explanation is that accumulated anime LoRAs and tags make switching expensive. It is one person's observation, not a dataset, and the author notes the counting windows differ.

Today's Civitai list agrees in shape. Among the top 20 safe-for-work entries, the base models named are MOSS Image Base, Pony, Anima, Illustrious, Krea 2, MiniMax H3, ZImageTurbo and SDXL 1.0, with Krea 2 appearing only as one checkpoint, one workflow and two LoRAs. A caution on the two biggest entries: Lenovo UltraReal (198,002 downloads all time) and NiceGirls UltraReal (139,922) list MOSS Image Base on their pages, but those are all-time counts for the model page, so they say nothing about demand for MOSS-Image itself, which the Hugging Face card shows at 958 downloads last month for Base.

Trend 5: workflows are a download category of their own

Three of the table rows are not models at all. They are packaged ComfyUI graphs, with names that promise speed boosts, low VRAM options and multi-reference editing. For people who run a character, this is the real shape of demand: not another checkpoint, but a graph that already wires the reference images, the upscaler and the detailer. A downloaded graph also brings its own risk, which is missing custom nodes and unpinned versions. Check what a graph needs before you trust it with a feed, and read our guide to keeping a consistent character in ComfyUI for what a safe graph keeps fixed.

Trend 6: hosted models are arriving inside ComfyUI

The newest Krea release is not a download at all. Comfy's documentation says Krea 3 runs in ComfyUI as a Partner Node, with no weights to download and nothing running on your own GPU: the node sends your prompt and reference images to Krea's servers. It accepts up to 8 reference images, outputs 1K or 2K, and each run is billed to your Comfy account. For a creator this is a trade: strong multi-reference editing in the same graph editor, against sending your character's reference photos to a third party and paying per run.

docs.comfy.orgOpen the page

What the six trends mean for a character feed

Three habits follow from the data. First, pin your ComfyUI version for the daily feed and test any launch-week model in a separate install. Second, keep a base file for training and a distilled file for rendering, and keep notes on which one made each approved image. Third, when a graph or a model asks for your character's reference photos, know where they go: local means your own machine, a partner node means someone else's server.

Gaps and caveats

• Download counts are not users. A model that ships as many files, or that bots and mirrors fetch, can look larger than it is.
• Hugging Face counts are 30-day rolling values and Civitai counts on these pages are all-time, so the two cannot be added or compared directly.
• The Civitai base label shows what the model page lists, not which files the downloads were for.
• The Japanese analysis of September is one author's reading of snapshots; we cite it for the pattern, not for its numbers.
• We could not read Reddit from our network, so the loudest English-language threads are missing from the voices.
• Items labelled uncensored carry the largest counts on Hugging Face this week. We left them out and do not draw any conclusion from them.

Our verdict

Watch video, not the base image models, for where attention is going this month, and treat every new launch as a nightly-build experiment until the threads settle. The sturdy choice for a monetized account is the one whose license you have read and whose version you have pinned: LTX-2.5 for open, local image-to-video, Apache-licensed builds where the parent allows it, and a base plus distilled pair for stills. The trend we would bet on is the packaged workflow, because it is the first thing in the list that saves a creator time instead of adding another file to manage.

If you would rather not manage that by hand, the Musetta studio runs on your own ComfyUI and keeps each character's approved reference photos, so a draft on a fast checkpoint and a final on the base start from the same face. The advice above holds without it, and our comparison with working in ComfyUI directly is candid about where the manual route is enough.

Musetta studioDemo data