Tencent’s Hunyuan Image 3 is now available on Novita AI as an async text-to-image API at https://api.novita.ai/v3/async/hunyuan-image-3. The model uses an 80-billion-parameter mixture-of-experts architecture — the largest open-source image generation MoE to date — and is a practical choice for developers who need strong multilingual text rendering, fine-grained prompt following, and flexible output sizing at $0.10 per image. This guide covers what the model does well, how to integrate it, and where it fits in a production image pipeline.
Key Takeaways
- Hunyuan Image 3 is available on Novita AI at
POST https://api.novita.ai/v3/async/hunyuan-image-3, priced at $0.10/image (checked 2026-07-06). - The model uses a 80B MoE architecture (13B parameters active per token) with an autoregressive unified generation framework.
- It natively supports Chinese and English prompts with strong in-image text rendering.
- The API is asynchronous: you submit a request, receive a
task_id, then poll the Task Result API for output. - Practical fit: complex multilingual scenes, text-heavy imagery, and pipelines that need reasoning-level prompt interpretation.
What Is Hunyuan Image 3?
Hunyuan Image 3 (HunyuanImage 3.0) is Tencent’s third-generation text-to-image model, open-sourced in September 2025. Unlike most image generation models built on diffusion transformer (DiT) architectures, Hunyuan Image 3 uses a unified autoregressive framework that treats text and image generation as a single modeling problem — the same architectural approach used in large language models.
The result is a model that reasons about image content before generating it, using a chain-of-thought schema with “thinking tokens” to interpret sparse or ambiguous prompts, infer missing details, and produce outputs that match intent rather than just matching token frequency.
Key architecture facts (source: HunyuanImage 3.0 technical report, arXiv 2509.23951, September 2025):
- 80B total parameters, 64 experts, ~13B parameters activated per token
- Autoregressive unified multimodal framework (not DiT-based)
- Chain-of-thought prompt reasoning via thinking tokens
- Reinforcement learning post-training for prompt alignment
- Native Chinese and English support
Compared to Hunyuan Image 2.1, the 3.0 release shows a 14.1% relative win rate improvement in professional human evaluation on image quality and text alignment.
Hunyuan Image 3 API Access on Novita AI
Novita AI hosts Hunyuan Image 3 as an async task API. The workflow is:
- POST a generation request to
https://api.novita.ai/v3/async/hunyuan-image-3 - Receive a
task_idin the response - Poll
https://api.novita.ai/v3/async/task-resultwith thetask_iduntil the task status isTASK_STATUS_SUCCEED - Extract the image URL from the result
Get your API key from the Novita AI console. New accounts receive $1 in free credits.
Hunyuan Image 3 Specs and Pricing Summary
| Field | Details | Source / Date checked |
|---|---|---|
| Display name | Hunyuan Image 3 | Novita AI docs / 2026-07-06 |
| Endpoint | /v3/async/hunyuan-image-3 | Novita AI docs / 2026-07-06 |
| Base URL | https://api.novita.ai | Novita AI docs / 2026-07-06 |
| Architecture | 80B MoE, autoregressive, 13B active | arXiv 2509.23951 / Sep 2025 |
| Supported languages | English, Chinese | Tencent HunyuanImage-3.0 repo / Sep 2025 |
| Output format | JPEG (default) | Novita AI docs / 2026-07-06 |
| Pricing | $0.10 per image | novita.ai/models / 2026-07-06 |
| API type | Async (submit + poll) | Novita AI docs / 2026-07-06 |
| Best fit | Multilingual scenes, text-heavy imagery, complex prompt interpretation | Judgment |
Key Capabilities for Developers
Prompt reasoning. The chain-of-thought mechanism means the model can take a short prompt like “a product photo for a high-end Japanese tea brand” and elaborate it with appropriate visual context — lighting, composition, style cues — without requiring a long detailed prompt. This is useful for pipelines where prompts come from end users rather than prompt engineers.
Text rendering. Hunyuan Image 3 handles in-image text in both Chinese and English significantly better than most open-source T2I models. If your use case involves generating images with legible signage, labels, UI mockups, or localized content, this is one of the few models where you don’t need a separate post-processing step for text.
Flexible sizing. The API accepts a size parameter as width*height, covering a broad range of aspect ratios for web, social, and print output.
Bilingual pipelines. Chinese-language prompts generate Chinese-appropriate content without romanization artifacts or translation overhead. Useful for apps targeting Chinese-speaking audiences.
When to Use Hunyuan Image 3
- Your prompts are short or user-provided and need intelligent elaboration before generation
- You need legible Chinese or English text rendered within the image
- You’re generating localized marketing assets, product imagery, or editorial visuals
- Your prompt describes complex scenes with multiple semantic elements that need coherent composition
When Not to Use Hunyuan Image 3
- You need the lowest cost per image: at $0.10/image, alternatives like Qwen Image on Novita AI ($0.02/image) or Seedream 3.0 ($0.03/image) are significantly cheaper for high-volume pipelines where text rendering and reasoning are not critical.
- You need image editing or inpainting: Hunyuan Image 3 is text-to-image only. For image editing workflows, use a dedicated editing model.
- You need sub-second response times: the async architecture adds latency from polling. It is not suitable for real-time interactive applications.
API Quick Start
Python
import requests
import time
API_KEY = "YOUR_NOVITA_API_KEY"
BASE_URL = "https://api.novita.ai"
def generate_image(prompt: str, size: str = "1024*1024", seed: int = -1) -> str:
# Submit generation request
response = requests.post(
f"{BASE_URL}/v3/async/hunyuan-image-3",
headers={
"Authorization": f"Bearer {API_KEY}",
"Content-Type": "application/json",
},
json={
"prompt": prompt,
"size": size,
"seed": seed,
},
)
response.raise_for_status()
task_id = response.json()["task_id"]
# Poll for result
while True:
result = requests.get(
f"{BASE_URL}/v3/async/task-result",
headers={"Authorization": f"Bearer {API_KEY}"},
params={"task_id": task_id},
)
result.raise_for_status()
data = result.json()
status = data.get("task", {}).get("status")
if status == "TASK_STATUS_SUCCEED":
return data["images"][0]["image_url"]
elif status == "TASK_STATUS_FAILED":
raise RuntimeError(f"Generation failed: {data}")
time.sleep(2)
image_url = generate_image(
prompt="A minimalist product photo of a premium Japanese matcha tin on a white marble surface, studio lighting",
size="1024*1024",
)
print(image_url)cURL
# Step 1: Submit request
curl -s -X POST https://api.novita.ai/v3/async/hunyuan-image-3 \
-H "Authorization: Bearer $NOVITA_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"prompt": "A neon-lit street food stall in Hong Kong at night, cinematic, rainy reflections",
"size": "1280*720",
"seed": -1
}'
# Returns: {"task_id": "..."}
# Step 2: Poll for result
curl -s "https://api.novita.ai/v3/async/task-result?task_id=YOUR_TASK_ID" \
-H "Authorization: Bearer $NOVITA_API_KEY"Request Parameters
| Parameter | Type | Required | Notes |
|---|---|---|---|
prompt | string | Yes | Text description of the image. Supports English and Chinese. |
size | string | No | Output dimensions as width*height (e.g., "1024*1024", "1280*720", "768*1024") |
seed | integer | No | -1 for random; set a fixed value to reproduce results |
Task Result Status Values
| Status | Meaning |
|---|---|
TASK_STATUS_QUEUED | Request received, waiting to run |
TASK_STATUS_RUNNING | Generation in progress |
TASK_STATUS_SUCCEED | Complete — image URL available in images[0].image_url |
TASK_STATUS_FAILED | Generation failed |
How Hunyuan Image 3 Fits Your API Workflow
The async pattern used by Hunyuan Image 3 on Novita AI is the same as other image and video generation endpoints on the platform — submit a task, receive a task_id, poll for results. If you’re already using FLUX.1 Dev on Novita AI or the Kontext family, switching to Hunyuan Image 3 is a one-line endpoint change plus updating the request body.
One practical difference: Hunyuan Image 3 takes a flat request body (prompt, size, seed) rather than nested input/parameters objects. Check the Novita AI docs for the canonical schema before integrating.
Final Recommendation
Hunyuan Image 3 is the right choice on Novita AI when prompt reasoning, multilingual text rendering, or complex scene composition matter more than cost. At $0.10/image it is priced at the higher end of the Novita AI image catalog — roughly 3× Seedream 3.0 and 5× Qwen Image — but it delivers a meaningfully different capability: the model interprets sparse prompts, generates legible Chinese and English text in images, and handles semantically dense scenes where simpler models tend to produce incoherent compositions.
For high-volume commodity image generation, cheaper alternatives are the better starting point. For production workflows where image quality, text accuracy, or bilingual content are requirements, Hunyuan Image 3 is worth the price difference.
FAQ
What is the Novita AI endpoint for Hunyuan Image 3?
POST https://api.novita.ai/v3/async/hunyuan-image-3. The API is asynchronous — you submit a request, receive a task_id, then poll /v3/async/task-result until the status is TASK_STATUS_SUCCEED.
How much does Hunyuan Image 3 cost on Novita AI?
$0.10 per image as of 2026-07-06 (source: novita.ai/models). Verify current pricing on the Novita AI pricing page before building a cost model for production workloads.
What makes Hunyuan Image 3 different from FLUX.1 or Stable Diffusion?
Hunyuan Image 3 uses an autoregressive unified multimodal framework rather than the diffusion transformer (DiT) architecture used by FLUX.1 and most Stable Diffusion variants. The practical effect is stronger prompt reasoning via chain-of-thought thinking tokens, better multilingual text rendering, and more coherent handling of complex multi-element scenes.
Does Hunyuan Image 3 support image editing or inpainting?
No. The model on Novita AI is text-to-image only. For image editing, use a dedicated editing model such as FLUX.1 Kontext or Qwen-Image-Edit.
Can I use Chinese prompts with Hunyuan Image 3?
Yes. The model natively supports Chinese and English prompts. Chinese prompts generate Chinese-appropriate visual content without translation artifacts.
How do I reproduce a specific image output?
Set the seed parameter to a fixed integer value. The same seed, prompt, and size combination will reproduce the same image.
