What is Hunyuan Image 3.0?
Hunyuan Image 3.0 is Tencent's open image model. Where most open image models are diffusion transformers, it is one autoregressive mixture-of-experts model that both reads and draws images, with 80B parameters and 13B active per token. It fills out short prompts from its world knowledge, and the Instruct version reasons over the prompt first, edits pictures and merges up to three reference images.
It is still Tencent's newest open-weights image model. Tencent Cloud also sells it by API as Hy-Image-3.0, while the newer Hy-Image 3.5 Preview, with 4K output and multi-turn editing, is API-only.
Which Hunyuan Image 3.0 version should I use?
Use Instruct-Distil if you have eight 80 GB GPUs and want editing, the base model if you only need text to image on three, and the API if you have none.
- HunyuanImage-3.0
- Text to image only; the lightest to host at three 80 GB GPUs, and it does not rewrite prompts itself.
- HunyuanImage-3.0-Instruct
- Adds prompt rewriting, step-by-step reasoning, editing and three-image fusion, on eight 80 GB GPUs.
- HunyuanImage-3.0-Instruct-Distil
- The same abilities as Instruct in 8 sampling steps instead of 50, on the same eight GPUs.
- Hy-Image-3.0 API
- No hardware to run: $0.032 per image on Tencent Cloud, with output up to an area of 1024 x 1024.
Who is Hunyuan Image 3.0 best for?
Hunyuan Image 3.0 is best for teams with multi-GPU servers who want an open model with strong prompt following, for photo editing with the Instruct version, and for apps that can start on Tencent Cloud's API.
Teams with large GPU servers
The base model wants three 80 GB GPUs and the Instruct versions eight, so it suits labs and studios rather than a single workstation.
Prompts with many objects and details
People who test it locally rate its prompt following above other open models, and it adds fitting detail when a prompt is sparse.
Editing and combining images
The Instruct version adds or removes objects, restyles pictures and swaps backgrounds from a written instruction, and merges up to three images into one.
Apps that can start on an API
Tencent Cloud serves it for $0.032 per image, so you can build on it before deciding whether to host the weights yourself.
What's included with Hunyuan Image 3.0?
One 80B mixture-of-experts model that reads and draws images, prompt expansion, instruction editing, three-image fusion, a faster distilled version, vLLM support and a hosted API.
- One model for reading and drawing
- An autoregressive mixture-of-experts design with 64 experts, 13B of its 80B parameters active per token.
- Prompt expansion
- It fills out short prompts from world knowledge, and the Instruct version can rewrite the prompt and reason step by step before drawing.
- Instruction editing
- Instruct adds elements, removes objects, changes style and replaces backgrounds while keeping the rest of the picture.
- Multi-image fusion
- Instruct combines up to three reference images into one new picture.
- Distilled version
- Instruct-Distil needs 8 sampling steps instead of the default 50, for faster output.
- Faster self-hosting
- vLLM support, plus FlashInfer kernels that Tencent says make inference up to 3x faster after a one-time compile.
- Hosted API
- Tencent Cloud TokenHub serves it as hy-image-v3 with 37 preset sizes, optional prompt rewriting and a custom watermark line.
How much does Hunyuan Image 3.0 cost?
Hunyuan Image 3.0 is free to download and run under Tencent's community license, or $0.032 per image through Tencent Cloud's API; the real cost of self-hosting is the three or more 80 GB GPUs it needs.
- Text to image
- At least 3 GPUs with 80 GB each
- Tencent Hunyuan Community License; weights on Hugging Face, code on GitHub
- Adds editing and merging up to 3 images
- Rewrites and reasons over your prompt first
- At least 8 GPUs with 80 GB each
- Tencent Hunyuan Community License
- Instruct abilities in 8 sampling steps
- At least 8 GPUs with 80 GB each
- 20,000 tokens per image at $1.60 per million tokens
- Text to image with up to 3 reference images
- Output area up to 1024 x 1024
The license does not cover the European Union, the United Kingdom or South Korea, and companies whose products had more than 100 million monthly active users need a separate license from Tencent. Outputs may not be used to improve other AI models, and AI-made content you publish must be labelled as machine-generated; Tencent claims no rights in the images you make. API prices are in US dollars on Tencent Cloud's international site, billed hourly after use. Tencent's newer Hy-Image 3.5 Preview costs $0.024 per image at 1K or 2K and $0.032 at 4K. HunyuanImage-3.0-Instruct's open weights come under the Tencent Hunyuan Community License. Full pricing on tencentcloud.com. We checked these prices on .
How do readers rate Hunyuan Image 3.0?
No one has rated Hunyuan Image 3.0 yet. Be the first.
Sign in to rate. One rating per account.
Hunyuan Image 3.0 pros and cons
Users rate its prompt following and realism close to the best hosted models, but it needs server-class GPUs, runs slowly on anything smaller, and its license excludes the EU, UK and South Korea.
Pros
- People testing it locally say no other open model beats its prompt following, with only ChatGPT and Nano Banana ahead, and that it handles scenes with many objects better than FLUX.2 dev.
- Users praise the look of its images, from easy photorealism to varied faces, and some rate it above Qwen-Image and Z-Image.
- In Tencent's blind test of 1,000 prompts, raters preferred it slightly over Seedream 4.0, Nano Banana and GPT-Image, by net margins of 1.17%, 2.64% and 5.00%.
- Open weights under a royalty-free license, plus a $0.032-per-image API when you do not want to host it.
Cons
- The hardware bar is far above one consumer GPU: Tencent recommends three 80 GB cards for the base model and eight for Instruct, and Reddit users with an RTX 5090 report out-of-memory errors even with 4-bit versions.
- Squeezed onto smaller hardware it is slow and fiddly: users describe long waits with 4-bit weights split across GPU and system memory, and ComfyUI setups that take more debugging than generating.
- Some users say the gain over rival open models is not worth running a model about ten times their size, and Tencent's API caps output at an area of 1024 x 1024, while Nano Banana Pro goes to 4K.
- The license excludes the EU, the UK and South Korea, and bars using its outputs to improve other AI models.
What are the best Hunyuan Image 3.0 alternatives?
Three tools on Recaply do Hunyuan Image 3.0's job. Z-Image and GLM-Image are open source, so you can run them on your own server.
Z-ImageOpen SourceChoose it when you need an open model for one consumer GPU; its 6B Turbo version fits in 16 GB of memory, needs 8 steps and uses the Apache 2.0 license.
GLM-ImageOpen SourceChoose it when your images carry a lot of text; Z.ai built it for accurate text rendering, releases it under MIT, runs it in about 23 GB with CPU offload, and sells API access at $0.015 per image.
Nano Banana ProFreeChoose it when you want hosted 4K output and no GPUs to manage; it costs $0.134 per 1K or 2K image on the Gemini API, and Tencent's own editing test puts it slightly ahead of the Instruct version.
Questions about Hunyuan Image 3.0
What is Hy-Image 3.5 Preview?
It is Tencent's newer image model, sold only through Tencent Cloud's API. It adds multi-turn editing, up to 20 reference images and output up to 4096 x 4096, and costs $0.024 per image at 1K or 2K, or $0.032 at 4K. Tencent has not published its weights.
Who is Hunyuan Image 3.0 for?
Hunyuan Image 3.0 is best for teams with multi-GPU servers who want an open model with strong prompt following, for photo editing with the Instruct version, and for apps that can start on Tencent Cloud's API.
Is Hunyuan Image 3.0 free?
Yes. Hunyuan Image 3.0 has a free HunyuanImage-3.0 plan: text to image.
How much does Hunyuan Image 3.0 cost?
Hunyuan Image 3.0's paid plans are Hy-Image-3.0 API at $0.032 per image on Tencent Cloud TokenHub.
Is there an open source alternative to Hunyuan Image 3.0?
Yes. Z-Image and GLM-Image are open source tools that do the same job, so you can run your own.
What are the best Hunyuan Image 3.0 alternatives?
The closest alternatives on Recaply are Z-Image, GLM-Image and Nano Banana Pro.
Where can I try Hunyuan Image 3.0?
On its official website, huggingface.co.
Can I use Hunyuan Image 3.0 commercially?
Yes in most places. The Tencent Hunyuan Community License is royalty-free, but it does not apply in the European Union, the United Kingdom or South Korea, and businesses with more than 100 million monthly active users must ask Tencent for a license. You also may not use its outputs to train other AI models.
What hardware does Hunyuan Image 3.0 need?
Tencent recommends at least three GPUs with 80 GB of memory each for the base model and eight for the Instruct and Distil versions, since each has 80 billion parameters with 13 billion active. The code is tested on Python 3.12 with CUDA 12.8.
Is there a Hunyuan Image 3.0 API?
Yes. Tencent Cloud's TokenHub serves it as Hy-Image-3.0 (model ID hy-image-v3) at $0.032 per image. It takes prompts up to 8,192 characters and up to three reference images, returns images up to an area of 1024 x 1024, and keeps result links for 12 hours.
How does Hunyuan Image 3.0 compare with Nano Banana?
In Tencent's blind test of 1,000 prompts, raters preferred Hunyuan Image 3.0 over the original Nano Banana by a net 2.64%. For editing, Tencent's own test puts the Instruct version slightly behind Nano Banana Pro. Nano Banana models are hosted only, while Hunyuan's weights are open.
Our verdict
Hunyuan Image 3.0 is one of the strongest open image models, but it is built for GPU servers, not desktops.
Tencent's blind tests put it level with or slightly ahead of Seedream 4.0 and Nano Banana, and local testers rate its prompt following highly, yet the base model wants three 80 GB GPUs.
- Skip it for
- Anyone with a single consumer GPU, and anyone in the EU, UK or South Korea, where the license does not apply.
- Check first
- Try the $0.032-per-image Hy-Image-3.0 API before buying hardware, and make sure its 1024 x 1024 output limit is enough.
GPT Image 2
FLUX 3 Image Klein
Ideogram 4.5
Seedream 4.5
MAI-Image-2.7