Images · Tencent
Hunyuan Image
Tencent's large open-weight model, run in private mode. Pick it for long, knowledge-heavy prompts in English or Chinese.
Private Moderate costWhat sets it apart
Tencent's large open-weight image model: 80B parameters with 13B active per token in a unified autoregressive design. It reasons over long prompts and renders long text strings, in private mode here.
- 80B parameters, 13B active (MoE)
- Open weights, run in private mode
- Handles long, knowledge-heavy prompts
Strengths and limits
Where it shines
- Private mode with a large open model (80B parameters, mixture of experts)
- Reasons over long prompts and renders long text strings
Think twice when
- About 4.6 credits, dear for a private model
- First released in 2025: newer families often follow prompts better
- Prompts up to 3,000 characters
Use cases
A long, detailed brief
A cutaway illustration of a Song dynasty river boat showing the cabin, the crew quarters and the cargo hold, labeled in English and Chinese.
It reasons over long prompts and renders long text.
A private concept image
A concept image for an unreleased board game: a dragon sleeping on a pile of maps, ink and watercolor style.
Private mode keeps unreleased ideas from being retained.
Against its rivals
- Grok ImagineGrok Imagine is private too, ranks far higher and costs less.#5 of 37 on Artificial Analysis Text to Image Arena (1,155 Elo)
- Qwen ImageQwen Image 3 ranks higher on dense text; Hunyuan is the open-weight option.#11 of 37 on Artificial Analysis Text to Image Arena (1,090 Elo)
- Z-Image TurboZ-Image is far smaller and faster at a fraction of the price; Hunyuan follows long prompts better.#35 of 37 on Artificial Analysis Text to Image Arena (941 Elo)
Version history
Each version, dated, with what it changed over the one before.
Hollow points are earlier versions no longer offered in Sub Rosa.
See the numbers
| Version | Released | Elo |
|---|---|---|
| HunyuanImage 3.0 | 2025-09-28 | 944 |
| HunyuanImage 3.0 Instruct | 2026-01 | 995 |
- HunyuanImage 3.0Latest
- Open-source release of the largest image MoE model at the time, with 64 experts
- Native multimodal understanding and generation
- HunyuanImage 3.0 InstructEarlier, not in Sub Rosa
- Adds instruction reasoning and image-to-image editing
- HunyuanImage 3.5 (Preview)Earlier, not in Sub Rosa
- Preview of the next generation, 1087 on the edit arena
Benchmarks
How to read these scores: what a benchmark tells you.
Artificial Analysis Text to Image Arena: #34 of 37 models measured in the catalog. Thousands of people compare two images made from the same prompt without knowing the models, and pick the better one. An Elo rating, as in chess: higher is preferred more often. 100 points apart means the higher one wins about 64% of duels.
- 1. GPT Image 2.5 SunburstGPT Image1,197
- 2. GPT Image 2.5 FlareGPT Image1,191
- 3. GPT Image 2GPT Image1,172
- 4. Nano Banana 2.1Nano Banana1,172
- 5. Grok Imagine 2.0Grok Imagine1,155
- 6. Nano Banana 2Nano Banana1,125
- 7. Muse ImageMuse Image1,115
- 8. GPT Image 1.5GPT Image1,107
- 9. Nano Banana ProNano Banana1,102
- 10. Nano Banana 2 LiteNano Banana1,096
- 34. Hunyuan Image 3.0Hunyuan Image944
See the numbers
| Model | Elo |
|---|---|
| 1. GPT Image 2.5 Sunburst | 1,197 |
| 2. GPT Image 2.5 Flare | 1,191 |
| 3. GPT Image 2 | 1,172 |
| 4. Nano Banana 2.1 | 1,172 |
| 5. Grok Imagine 2.0 | 1,155 |
| 6. Nano Banana 2 | 1,125 |
| 7. Muse Image | 1,115 |
| 8. GPT Image 1.5 | 1,107 |
| 9. Nano Banana Pro | 1,102 |
| 10. Nano Banana 2 Lite | 1,096 |
| 34. Hunyuan Image 3.0 | 944 |
The model in this family
| Model | Use | Privacy | Price |
|---|---|---|---|
Hunyuan Image 3.0hunyuan-image-v3 | Image | Private | $0.0405 per image |
Specifications
| Specification | HunyuanImage 3.0 |
|---|---|
| Released | September 28, 2025 |
| Parameters | 80B total, 13B active |
| Weights | Open, Tencent Hunyuan Community |
| Accepts | text, images |
| Produces | images |
| Price | $0.0405 per image |
| Privacy | Private |
Prices, context and lengths from the live catalog on October 16, 2026; parameters and licenses from the maker.
Sources
Everything on this page comes from these pages, read on October 16, 2026 or as dated on each score.