# HunyuanImage 3.0: Architecture, Access, and Deployment Guide

> Published 2026-04-03, updated 2026-09-06 · https://www.promptzone.com/miles_fischer/tencent-unveils-hunyuanimage3-for-advanced-image-generation-373j

---
title: "HunyuanImage 3.0: Architecture, Access, and Deployment Guide"
published: true
description: "HunyuanImage 3.0 is Tencent's 80B-parameter image model; learn its architecture, official download path, inference setup, and license limits."
tags: ai,imagegeneration,deeplearning
---
HunyuanImage 3.0 is Tencent's image generation model built around a unified multimodal autoregressive framework. Tencent publishes its inference code and weights through GitHub and Hugging Face, including a text-to-image checkpoint and separately named Instruct releases. The original model has 80 billion total parameters, with 13 billion activated per token across a mixture-of-experts architecture. [Official repository][repo]

## What are the key facts about HunyuanImage 3.0?

| Field | Verified information |
| --- | --- |
| Developer | Tencent Hunyuan. [Official repository][repo] |
| Released | September 28, 2025 for the original public weights and inference code. [Release record][repo] |
| Type | Native multimodal image generation using an autoregressive framework and mixture-of-experts architecture. [Model card][card] |
| Size or parameters | 80B total parameters, 13B activated per token, and 64 experts. [Model card][card] |
| License and access | Downloadable weights under the Tencent Hunyuan Community License, with territory and use restrictions. [Model license][license] |
| Where it runs | User-managed GPU inference through the documented CUDA/PyTorch environment and Transformers or repository code. [Setup instructions][repo] |
| Recommended VRAM | Original 3.0: at least 3 × 80 GB; Instruct and Instruct-Distil: at least 8 × 80 GB. These are Tencent's recommendations for the documented models. [Model table][repo] |

## What can HunyuanImage 3.0 generate, and which checkpoint fits?

Tencent presents HunyuanImage 3.0 as combining language and image understanding with generation in a unified framework. Its model documentation emphasizes following detailed instructions and drawing on world knowledge when composing an image. The public weights and inference implementation make it possible to inspect the model's setup and evaluate it in a controlled environment. [Model card][card]

The original release is particularly useful for a test set focused on semantic requirements. Ask for a scene with named objects, materials, and spatial relationships, then inspect whether those requirements survive generation. Keep photorealism and instruction following as separate review criteria so an attractive image does not hide a missing object.

Tencent's repository describes machine and human evaluation procedures and supplies generation examples. Those results provide developer evidence about the intended capabilities, but your own deployment still needs a representative prompt set. Do not convert a developer evaluation into a claim that every production prompt will succeed. [Evaluation documentation][repo]

The distinction between original and Instruct checkpoints matters. Tencent's release record dates the Instruct and Instruct-Distil releases to January 26, 2026, and describes them as adding reasoning-related behavior and image-to-image capabilities. Choose the checkpoint that matches the operation you intend to test. [Release history][repo]

## What hardware and license limits apply to HunyuanImage 3.0?

The published parameter figures describe different aspects of the model. Total parameters indicate the full model capacity, while activated parameters describe the subset used per token. For deployment planning, do not size the entire checkpoint as though it contained only the activated subset. Record the actual memory usage of the loading and inference configuration you choose. [Architecture description][card]

Tencent recommends at least 3 × 80 GB of GPU memory for the original model and at least 8 × 80 GB for Instruct and Instruct-Distil. Its setup instructions specify PyTorch and CUDA versions and optional acceleration components. Use the recommendation for your checkpoint when planning the documented deployment; measure any alternative configuration separately. [Environment and model table][repo]

The community license also has material restrictions. Its grant excludes the European Union, United Kingdom, and South Korea; it includes separate terms for organizations above its specified user threshold and restricts using the works or outputs to improve unrelated AI models. Review the actual agreement before selecting this checkpoint for deployment. [License terms][license]

The original text-to-image checkpoint does not automatically rewrite prompts, according to Tencent's usage notes. Its example prompt-enhancement path uses a separate service, while the Instruct model has its own documented reasoning workflow. Keep that difference explicit when comparing outputs or deciding whether an experiment is fully local. [Original checkpoint instructions][repo]

## How do you download and run the original HunyuanImage 3.0?

First choose the original text-to-image checkpoint or the separately documented Instruct family, and confirm that your intended use fits the model license. For the original checkpoint, use Tencent's official repository and the `tencent/HunyuanImage-3.0` model repository. The following steps follow its text-to-image instructions. [Setup guide][repo]

1. Clone the official repository and open its directory.
2. Prepare the documented environment, including the specified PyTorch installation and repository requirements.
3. Download the original weights with `hf download tencent/HunyuanImage-3.0 --local-dir ./HunyuanImage-3`.
4. Load that local directory using the supplied Transformers-compatible code.

Tencent specifically documents using a local directory without a dot in its name for this loading path. The model's public identifier and local folder name therefore differ intentionally. After installing dependencies and downloading the checkpoint, its loading pattern is:

```python
from transformers import AutoModelForCausalLM

model_path = "./HunyuanImage-3"
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    attn_implementation="sdpa",
    trust_remote_code=True,
    torch_dtype="auto",
    device_map="auto",
    moe_impl="eager",
)
model.load_tokenizer(model_path)
image = model.generate_image(
    prompt="A blue ceramic bowl on a wooden table, soft side lighting.",
    stream=True,
)
image.save("hunyuan-image.png")
```

The custom-code setting is required by the documented loading example, so use Tencent's published model code and review the selected revision. This snippet performs the original text-to-image operation and does not configure Tencent's optional external prompt-enhancement service. [Transformers example][repo]

Treat the first successful image as an installation check. Confirm that the model loads, produces an output, and saves it at the intended location. Then add representative prompts, recording the checkpoint revision, environment, settings, elapsed time, and peak memory. These measurements belong to that particular configuration.

Build a small acceptance sheet around your intended use. For a product concept, check shape, material, color, and object placement. For a scene illustration, check required subjects and relationships. Keep failed outputs as part of the evaluation so later changes can be assessed against the same set of requirements.

For a preceding Tencent model with a different architecture, see the sibling [HunyuanImage 2.1 guide](/florence_liu/tencents-hunyuanimage-21-ai-image-generator-5hcm). Use the [cloud GPU pricing guide](/cloud-gpu-pricing) when comparing infrastructure for an environment you have measured.

## How does HunyuanImage 3.0 compare with HunyuanImage 2.1?

| Model | Published architecture | Relevant distinction |
| --- | --- | --- |
| HunyuanImage 3.0 | Multimodal autoregressive framework with an 80B mixture-of-experts model. [Model card][card] | Evaluate semantic instructions and the deployment configuration together |
| HunyuanImage 2.1 | A 17B diffusion-transformer base stage with a separate refinement stage. [2.1 model card][previous] | A different pipeline and a separately documented deployment path |

Tencent's 2.1 card documents an FP8/offloading configuration for image generation. Those resource figures apply to that model and setup, so they should not be reused as a requirement for 3.0. [2.1 deployment notes][previous]

For general workflow context, the [ComfyUI pillar](/tomas_novak/comfyui-2026-the-complete-guide-to-power-user-ai-image-generation-1g17) is a useful companion. Confirm support for the exact checkpoint and operation before assuming a workflow prepared for another model will load it.

## What else should you know before deploying HunyuanImage 3.0?

### How many parameters does HunyuanImage 3.0 have?

Tencent publishes 80 billion total parameters and 13 billion activated per token, with 64 experts. The version number is separate from those architecture figures. [Model card][card]

### Are the HunyuanImage 3.0 weights available?

Tencent publishes the original HunyuanImage 3.0 checkpoint and inference code. Their use is subject to the Tencent Hunyuan Community License and its restrictions. [Repository][repo], [license][license]

### Does the original checkpoint edit uploaded images?

Tencent documents the original HunyuanImage 3.0 checkpoint for text-to-image generation. For image-to-image work, follow the separately named Instruct checkpoint and its specific instructions. [Checkpoint usage guide][repo]

### Can I assume it fits on a consumer GPU?

Tencent's documented recommendation for original HunyuanImage 3.0 is at least 3 × 80 GB of GPU memory, with multi-GPU inference recommended. That recommendation does not establish support for a single consumer GPU; evaluate an alternative backend or compressed checkpoint using its own documentation. [Deployment instructions][repo]

## Sources

[repo]: https://raw.githubusercontent.com/Tencent-Hunyuan/HunyuanImage-3.0/main/README.md
[card]: https://huggingface.co/tencent/HunyuanImage-3.0
[license]: https://huggingface.co/tencent/HunyuanImage-3.0/blob/main/LICENSE
[previous]: https://huggingface.co/tencent/HunyuanImage-2.1/raw/main/README.md

- [Tencent HunyuanImage 3.0 release and inference documentation][repo]
- [Tencent HunyuanImage 3.0 model card][card]
- [Tencent Hunyuan Community License][license]
- [Tencent HunyuanImage 2.1 architecture and deployment comparison][previous]

<!-- pz-related-guides -->
## Related guides on PromptZone

- [Best SDXL Models in 2026](/tara_suzuki/best-sdxl-models-in-2026-realistic-anime-and-all-purpose-checkpoints-116)
- [ComfyUI 2026: The Complete Guide](/tomas_novak/comfyui-2026-the-complete-guide-to-power-user-ai-image-generation-1g17)
- [AI Model Releases Timeline](/ai-model-releases)