duongve · Released, not on our provider
MossYume (2026): what the anime and realistic text-to-image base model is, how to run it, its license and what its card leaves out
MossYume is an independent community text-to-image model by Hugging Face user duongve, uploaded on November 13, 2026 as v0.1, fine-tuned from OpenMOSS’s MOSS-Image-Base on 20 million curated anime and realistic images, and released under Apache-2.0 as a base checkpoint without supervised fine-tuning and with a thin card that gives no benchmarks.
Use it now
Chat
This chat starts with Qwen3.8 Flash, not MossYume.
≈ 1 credit per reply · See all models
New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.
Start chatting — 5 free credits- Maker
- duongve
- Type
- Image
- Released
- No date given
- OpenRouter id
- None
Checked
MossYume: facts, with sources
- Hugging Face’s commit history for duongve/MossYume starts with an initial commit on November 13, 2026 (15:30 UTC); the diffusion model, VAE and text encoder were uploaded within the next six minutes, a demo image later that day, and the README was edited on November 13 and again on November 16, 2026. We found no release announcement, so this is the repository creation date, not a stated release date.Source: huggingface.co · checked 2026-11-19
- The author’s Civitai page for MossYume lists a single version, v0.1, created on November 13, 2026 and published at 16:33 UTC the same day, with the base model shown as “MOSS Image”.Source: civitai.com · checked 2026-11-19
- The model card says MossYume is a text-to-image model fine-tuned from MOSS-Image-Base, the image generation foundation model of OpenMOSS-Team, as “an independent, community-driven experiment” not affiliated with any research lab or company. Its stated goal is to fix anatomical inconsistencies and stabilise irregular weight distributions in the original base model. The card’s title spells the name “MosYume” while the repository is MossYume.Source: huggingface.co · checked 2026-11-19
- For version v0.1 the card states: 20 million curated images spanning anime and realistic styles, with no AI-generated images and no model distillation; a 1:1 ratio between anime and realistic images; training at dynamic resolutions from 1024px to 2048px; native inference up to 2048×2048; and a clean base checkpoint without supervised fine-tuning, meant for further training or LoRA creation. The anime data sources it names are Danbooru (updated to September 4, 2026), e621, Gelbooru, Pixiv and Konachan, with other datasets updated at least to July 30, 2026. It does not say how the data was filtered.Source: huggingface.co · checked 2026-11-19
- The card lists the parts as a pre-trained Qwen3.5-9B text encoder, a pre-trained Flux.2 VAE and a fine-tuned MOSS-Image-Base backbone, and recommends CFG 3 to 5, 30 sampling steps, the er_sde sampler and the simple scheduler. It gives no parameter count, no benchmark, no example prompt and no workflow.Source: huggingface.co · checked 2026-11-19
- Hugging Face’s file listing for the repository shows three files that make a working set: diffusion_models/MossYume_v01.safetensors at 10,164,655,808 bytes, text_encoders/qwen3.5_9b_bf16.safetensors at 19,306,312,328 bytes and vae/flux2-vae.safetensors at 336,211,292 bytes, plus a 22 MB demo image. Together that is about 29.8 GB to download. The folder names follow ComfyUI’s models layout.Source: huggingface.co · checked 2026-11-19
- The Civitai page lists two files for v0.1: MossYume_v01.safetensors in bf16 (9,926,421 KB, about 10.2 GB) and MossYume_v01_int8_convrot.safetensors in int8 (5,286,815 KB, about 5.4 GB). It says the file holds only the diffusion model, with no VAE or text encoder, and thanks SumomoLee for the int8 conversion. It adds that the captions mixed tags, prose and structured captions.Source: civitai.com · checked 2026-11-19
- Civitai’s API lists permission flags for the page as printed: allowCommercialUse “Rent” and “RentCivit”, allowDerivatives true, allowDifferentLicense true and allowNoCredit true. These are Civitai’s own usage fields for the listing and are not the same wording as the Apache-2.0 on Hugging Face; the page we read does not explain how they relate. On November 19, 2026 Civitai shows 1,337 downloads and 123 thumbs up.Source: civitai.com · checked 2026-11-19
- On November 19, 2026 Hugging Face’s API shows the MossYume repository with 16 likes, 0 downloads counted, no Spaces and no discussions. The repository has no model-index. These are Hugging Face counters, not statements by the author.Source: huggingface.co · checked 2026-11-19 · second-hand
- MossYume is not on OpenRouter’s model list (checked November 19, 2026; the list has no model by this author or from the MOSS-Image family), and this site only chats with text models, so it cannot generate images. A text model such as Qwen3.8 Flash or Claude Haiku 5.5 can be chatted with here.Source: openrouter.ai · checked 2026-11-19 · second-hand
Our table
MossYume next to its base model MOSS-Image-Base
- About the base model: the MOSS-Image-Base card calls it the base text-to-image model of the OpenMOSS MOSS-Image family, taking natural-language prompts, tag sequences and JSON layouts, with weights and bundled code under Apache 2.0. Its quick start needs Python 3.10 or later, an NVIDIA GPU with CUDA, PyTorch 2.9 or later and diffusers 0.38 or later, and uses 50 steps at CFG 3.Source: huggingface.co · checked 2026-11-19
- About the base model: OpenMOSS’s GitHub page introduces MOSS-Image-Base and MOSS-Image-SFT on October 16, 2026 and calls it a 5B backbone. It also gives a low-memory profile that peaked at 3.65 GiB of GPU memory allocated (3.85 GiB reserved) and 14.00 GiB of system RAM for one 1024×1024 image at 50 steps, and says the DiT stays in CPU RAM in that mode. It also notes day-0 ComfyUI support for MOSS-Image, and community trainers, one of which advertises adapter training with as little as 6 GB of VRAM. These figures are for MOSS-Image-Base, not for MossYume, which does not state any.Source: github.com · checked 2026-11-19
Model cards
MossYume v0.1 and MOSS-Image-Base: who made them, license, data, settings
People who search for MossYume want to know what it is and whether it is worth the 30 GB of downloads. The honest answer from the cards: it is a young community fine-tune of an open base model, with a card that states its data and settings but no results to compare. If you want an image model whose maker publishes speed, memory and license details, read FLUX 3 Image Klein Comfy first.
| Item | MossYume v0.1 | MOSS-Image-Base |
|---|---|---|
| Made by | duongve, independent and not affiliated with any lab | OpenMOSS-Team |
| Hugging Face repository created | November 13, 2026 | October 11, 2026 |
| License | Apache-2.0 | Apache 2.0 (weights and bundled code) |
| What it is | A fine-tune of MOSS-Image-Base, a clean base checkpoint without SFT | The base model of the MOSS-Image family; a separate SFT model exists |
| Training data stated | 20 million curated images, 1:1 anime and realistic | Not stated on the card we read |
| Resolution | Trained from 1024px to 2048px, native up to 2048×2048 | Python API sizes must be positive multiples of 16; the example uses 1024×1024 |
| Recommended sampling | CFG 3 to 5, 30 steps, er_sde sampler, simple scheduler | Flow-matching Euler, 50 steps, CFG 3, shift 5 |
| VAE | Flux.2 VAE | Bundled VAE uses the FLUX.2 small decoder |
| JSON layout prompts | Not mentioned on the card | Supported |
Source: huggingface.coSource: huggingface.coSource: huggingface.coSource: huggingface.co
Our table
What the MossYume card says and what it leaves out
- Hugging Face’s metadata and the card both give Apache-2.0 as the license; the card has no separate license section and no LICENSE file in the repository. The card has no intended-use statement, no limitations section, no content policy or safety statement and no statement about what the model should not be used for. We found no discussions or community posts on the repository on November 19, 2026.Source: huggingface.co · checked 2026-11-19
Model card
MossYume card check: items a reader would look for and whether the card has them
The card is short, so this table says which parts of a normal model card are there. It is not a quality judgement on the model, only a list of what a reader can and cannot check.
| Item | On the card? |
|---|---|
| License | Yes: Apache-2.0, no license text |
| Training data size and sources | Yes: 20 million images, five named sources |
| Data filtering method | No |
| Recommended settings | Yes: CFG, steps, sampler, scheduler |
| Benchmarks or comparisons | No |
| Parameter count and hardware needs | No (file sizes are in the repository) |
| Step-by-step run instructions or workflow | No |
| Intended use, limitations, content policy | No |
Source: huggingface.coSource: huggingface.coSource: huggingface.co
Get told when it is available
Sign in and press the button: we store your account and this model, nothing else. No emails are sent yet.
Nearest available
Alternatives to MossYume you can use now
Qwen3.8 Flash
Alibaba$0.15 / $0.47 (per million tokens)
- Alibaba’s price for qwen3.8-flash is $0.15 / $0.47 per million input / output tokens (Singapore, EU, Hong Kong and US regions).Source: docs.qwencloud.com · checked 2026-11-13
- Listed on OpenRouter as qwen/qwen3.8-flash at $0.15 / $0.47 per million tokens.Source: openrouter.ai · checked 2026-11-13
Claude Haiku 5.5
Anthropic$0.10 / $0.50 (per million tokens)
- Anthropic’s model page says Claude Haiku 5.5 was released on October 7, 2026, is the latest Haiku, and will not be retired sooner than October 7, 2027. Anthropic’s launch post is dated the same day.Source: platform.claude.com · checked 2026-11-17
- In its launch post Anthropic calls Haiku 5.5 "the cheapest, fastest, and most capable small model we’ve ever released", for high-volume, cost-sensitive tasks such as summaries, compactions, database queries and classification, and as a coding subagent alongside Opus 5.5 and Sonnet 5.5. It calls Haiku 5.5 its first Haiku-class model with an adjustable effort setting.Source: anthropic.com · checked 2026-11-17
- Anthropic’s model page says it is built for high-volume, latency-sensitive work such as classification, routing, extraction and subagent tasks, and rates its latency "Fastest" in the current lineup (Fable 5.2 and Opus 5.5 are "Moderate", Sonnet 5.5 is "Fast").Source: platform.claude.com · checked 2026-11-17
FAQ
MossYume FAQ
What is MossYume?
MossYume is a text-to-image model fine-tuned from MOSS-Image-Base, trained on 20 million curated images split evenly between anime and realistic styles. Version v0.1 is a clean base checkpoint without supervised fine-tuning, which its author says makes it suited to further training or LoRA creation. It can render up to 2048×2048, according to the card.
Who made MossYume, and is it from OpenMOSS?
No. The card says it is an independent, community-driven experiment by the Hugging Face user duongve and is not affiliated with any research lab or company. It builds on MOSS-Image-Base, the open model of OpenMOSS-Team, which is why the name starts with Moss. Any question about the base model belongs with OpenMOSS; questions about this fine-tune belong with the author.
How do I run MossYume?
The card gives no step-by-step instructions or workflow. It recommends CFG 3 to 5, 30 steps, the er_sde sampler and the simple scheduler, and the repository holds the three parts in ComfyUI-style folders: the diffusion file, a Qwen3.5-9B text encoder and the Flux.2 VAE. The card does not say which loader to use, so check the MOSS-Image ComfyUI notes for the base model.
What hardware does MossYume need?
The card does not say. The diffusion file is 10,164,655,808 bytes in bf16 (Civitai also lists an int8 file of about 5.4 GB), the text encoder is 19.3 GB, and the full set is about 29.8 GB to download. For the base model, OpenMOSS reports a low-memory profile that peaked at 3.65 GiB of GPU memory and 14 GiB of RAM for one 1024×1024 image; that figure is not MossYume’s.
What license does MossYume use, and can I use it commercially?
Hugging Face lists Apache-2.0 for MossYume, and the base model is also Apache 2.0, but the card has no license text. Civitai’s page uses its own permission flags (commercial use as “Rent” and “RentCivit”, derivatives allowed) and does not explain how they relate. For anything commercial, read the license files yourself and ask the author if unsure.
What was MossYume trained on, and what is it for?
The card names five anime sources (Danbooru, e621, Gelbooru, Pixiv and Konachan) and says nothing about how the data was filtered. It states no intended use, no limitations and no content policy, so these are yours to judge before you use it. Its author describes the release as a base model for further training, with the aim of fixing anatomy problems in the base.
Can I use MossYume on this site?
No. It is not on OpenRouter’s model list, and this site only chats with text models, so it cannot generate images. For an image model with fuller documentation, see FLUX 3 Image Klein Comfy. For a text model to help word prompts, on this site a typical reply to Qwen3.8 Flash costs about 1 credit (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit. Hy4 Flash is another small chat model here.
Is this the official MossYume website?
No. This page is unofficial and not affiliated with or endorsed by duongve. Facts link to the pages we read.