Meituan's LongCat team (technical report arXiv 2512.07584, 8 December 2025)
LongCat-Image
Meituan's open 6B bilingual image model in your browser: photoreal pictures and posters with English or Chinese text, plus single-image edits. 8 sizes up to 1536×1536, no 17 GB GPU needed.
- Model
- A 6B core diffusion model — the report contrasts it with the 20B-plus mixture-of-experts image models common today
- Text rendering
- Text wrapped in single or double quotes (English or Chinese quote marks) gets character-level encoding; without quotes, text quality drops sharply, per the model card
- Versions
- LongCat-Image (text to image), LongCat-Image-Edit, LongCat-Image-Edit-Turbo (3 February 2026, about 10× faster) and LongCat-Image-Dev (a mid-training checkpoint for fine-tuning)
- Weights
- Open under Apache 2.0, with Diffusers (16 December 2025) and ComfyUI (22 March 2026) support; about 17 GB VRAM for text to image and 18 GB for edits with CPU offload
- Model-card settings
- Example run: 1344×768, 50 steps, guidance 4.0, prompt rewrite on
- On Kenerate
- kenerate/longcat-image — text to image in 8 sizes from 720×1280 to 1536×1536, or an edit of 1 uploaded image; seed control
- Real specs
- Official sources
- Free to start
- No watermark
Updated by the Kenerate AI team
On this page
What is LongCat-Image?
What is LongCat-Image?
LongCat-Image is an open 6-billion-parameter text-to-image model from Meituan's LongCat team, released on 5 December 2025 under the Apache 2.0 licence. It is bilingual (Chinese and English) and built for text inside images: put the words you want in quotation marks and it renders them character by character. A companion LongCat-Image-Edit model handles edits. On Kenerate AI it runs in the browser for 45 credits per image or edit.
Key facts
- Official name
- LongCat-Image
- Hugging Face ID
- meituan-longcat/LongCat-Image
- Maker
- Meituan LongCat team
- Released
- 5 December 2025
- Type
- Diffusion · text to image + edit
- Size
- 6B parameters
- Licence
- Apache 2.0
- Languages
- English and Chinese
- Sizes on Kenerate
- 8, up to 1536×1536
Models
LongCat-Image on Kenerate
Every version in the generator, and what each is good at.
Meituan
LongCat Image
Baidu
ERNIE Image
- Free plan
Alibaba
Qwen Image 3.0
- Reference images
- Up to 3
- Output
- 1K or 2K
Open Qwen Image 3.0Best for
Text in 12 languages, free plan
Black Forest Labs
Flux 2 Klein Base 9B
- Reference images
- Up to 3
- Output
- 1536 px
Open Flux 2 Klein Base 9BBest for
Open undistilled 9B with edits
BitDance
BitDance 14B
From our tests
Strengths and limits
Good at
- Made for text in pictures — signs, posters, packaging — in English and in Chinese, including rare characters, per Meituan.
- Small for what it does: a 6B model with open Apache 2.0 weights, so self-hosted copies can be used commercially.
- Generates and edits in one place: upload one image and describe the change.
- One flat price — 45 credits — for every size up to 1536×1536 and for edits.
Limits
- No real LongCat-Image outputs on this page yet — we only show a model's own images, so run a test prompt first.
- Text renders well only inside quotation marks; leave them out and the lettering suffers.
- Pricier than other open models on Kenerate (45 credits) and not on the free plan — for free drafts use Qwen Image 3.0.
- Edits take one reference image, and there is no negative prompt.
How to
How to use LongCat-Image on Kenerate
Step 1
Open the model
Go to /app/image and pick LongCat Image (the button on this page selects it for you). Sign in free to get started.
Step 2
Write the prompt, text in quotes
Describe the scene in English or Chinese and put every word that should appear in the image in quotation marks, e.g. a shop sign that reads "Night Market".
Step 3
Pick a size, or upload a photo
Choose one of 8 sizes up to 1536×1536 for a new image, or upload one picture and describe the change to edit it. Fix the seed to compare prompt tweaks.
Step 4
Generate and check the lettering
Each image or edit costs 45 credits. Zoom in on the text; if a word is off, shorten it or move it into its own quoted phrase and run again.
See it in the app
Open Image Studio and type what you want to see — subject, style, lighting and mood.
Watch the tour · 1:23
Prompts, styles, references, edit, Arena & chat
Chapters
FAQ
LongCat-Image questions
9 questions
What is LongCat-Image?
LongCat-Image is an open-source text-to-image model from Meituan's LongCat team, released on 5 December 2025. It has 6 billion parameters, understands Chinese and English, and is tuned for photorealism and for rendering text inside images. The weights are on Hugging Face as meituan-longcat/LongCat-Image under Apache 2.0.
Link to this answerIs LongCat-Image free?
The weights are free under Apache 2.0, but you need a GPU with about 17 GB of VRAM (with CPU offload) to run them. On Kenerate, LongCat Image is a paid model at 45 credits per image or edit, from one-time credit packs with no subscription. New accounts get 25 credits on sign-up, and the free plan includes Qwen Image 3.0 and Flux 2 Dev.
Link to this answerHow do I make LongCat-Image render text?
Put the exact words in single or double quotation marks — English "…" or Chinese “…” both work. The model card says quoted text switches on character-level encoding, and that text quality drops a lot without quotes. Keep each quoted phrase short and say where it goes: on a sign, a label or a poster title.
Link to this answerLongCat-Image vs LongCat-Image-Edit vs Dev — what's the difference?
LongCat-Image makes images from text. LongCat-Image-Edit changes an existing image from an instruction, and LongCat-Image-Edit-Turbo (February 2026) is a distilled Edit model about 10 times faster. LongCat-Image-Dev is a mid-training checkpoint meant for fine-tuning, not for everyday use. On Kenerate, LongCat Image covers both text to image and single-image edits.
Link to this answerIs LongCat-Image on Hugging Face?
Yes. Meituan publishes it at huggingface.co/meituan-longcat/LongCat-Image, alongside LongCat-Image-Edit, Edit-Turbo and Dev. The code and training toolchain are at github.com/meituan-longcat/LongCat-Image, and it works in Diffusers (LongCatImagePipeline) and ComfyUI.
Link to this answerLongCat-Image vs FLUX — which is better?
They aim at different things. LongCat-Image is a 6B bilingual model whose strength is text, especially Chinese characters. FLUX.2 [klein] Base 9B is an open 9B model with up to 3 reference images for edits, and it is far cheaper on Kenerate (4.5 credits vs 45). Run the same prompt on both to see which suits your image.
Link to this answerIs LongCat-Image the same as LongCat Avatar?
No. Both come from Meituan's LongCat team, but LongCat-Image makes still images, while LongCat Avatar turns a photo and a voice track into a talking video. Kenerate runs LongCat Avatar, LongCat Avatar 1.5 and the two-speaker LongCat 1.5 Duo in its talking-avatar tool.
Link to this answerCan I use LongCat-Image online without a GPU?
Yes. Kenerate AI runs it in the browser as LongCat Image — no download, Diffusers setup or 17 GB GPU. Pick it in /app/image, write a prompt and choose one of 8 sizes up to 1536×1536, or upload a photo to edit. It costs 45 credits per image.
Link to this answerCan I use LongCat-Image images commercially?
The model weights are Apache 2.0, which allows commercial use of self-hosted copies under that licence's terms. Images you make on Kenerate follow Kenerate's terms of service. Check any brand names, logos or real people in your prompt separately.
Link to this answerIn depth
LongCat-Image, in depth
Background, how to prompt it, and when to pick something else.
On this page
The quotation-mark rule
LongCat-Image reads quoted text differently from the rest of the prompt: anything inside single or double quotes — English or Chinese quote marks — is encoded character by character [1][2]. That is how it gets small and rare Chinese characters right. Practical rules:
- Quote every word that must appear in the image, and nothing else.
- One short phrase per quote — a title, a sign, a price — beats a quoted paragraph.
- Say where the text goes and how it looks: “a red neon sign above the door that reads "Open Late"”.
Running LongCat-Image locally vs online
Locally you install Diffusers from source (or the ComfyUI nodes) and need about 17 GB of VRAM with CPU offload, 18 GB for the Edit models [2]. The model card's example uses 50 steps and guidance 4.0 at 1344×768 [1]. On Kenerate nothing is installed: pick LongCat Image, choose a size or upload a photo, and pay 45 credits per result.
Why a 6B model
The technical report argues that a compact 6B diffusion model can match much larger 20B-plus mixture-of-experts models on text rendering and photorealism, while using less memory and running faster [3]. Meituan also released mid-training checkpoints and its training toolchain, so researchers can fine-tune from LongCat-Image-Dev [2][3].