2026-09-29

HunyuanImage 3.0 API: open weights, no listed Sume model

HunyuanImage 3.0 is an 80B-parameter open-weights image model from Tencent. Sume's image docs do not list it; here is what the card says and the hosted route.

4 min readSume

All posts

HunyuanImage 3.0 is published by Tencent as downloadable weights with inference code. Its Hugging Face card describes an 80 billion parameter Mixture of Experts model with 13 billion parameters activated per token. Sume's image docs do not list it, so there is no Sume model id for it.

Tencent's facts are from the HunyuanImage-3.0 card, read 2026-09-29. Sume's side is from Image generation.

What does the HunyuanImage 3.0 card say?

From the HunyuanImage-3.0 model card, read 2026-09-29.
Point	What the card says
Size	64 experts and a total of 80 billion parameters, with 13 billion activated per token
Weights release	September 28, 2025: inference code and model weights publicly available
Instruct variant	January 26, 2026: HunyuanImage-3.0-Instruct, with image-to-image generation including editing and multi-image fusion
License label	`tencent-hunyuan-community`
Software	Python 3.12+ (recommended and tested), CUDA 12.8

How do I run it?

The card's usage section installs PyTorch for CUDA 12.8 and downloads the weights, and its Gradio demo section sets GPUS with a default of 0,1,2,3. The card stresses that the CUDA version PyTorch uses must match the system's CUDA version. The card text read for this post states no GPU memory minimum, so this post gives none; check the card again before you size hardware.

Is there a hosted HunyuanImage 3.0 API on Sume?

No. The image docs name models such as openai/gpt-image-2.5, bytedance-seed/seedream-4.5 and black-forest-labs/flux.2-pro, and they say GET /v1/images/models returns the models available with their capabilities. HunyuanImage is not among the models the docs name. Read the endpoint for the live list:

curl "https://api.sume.com/v1/images/models" \
  -H "Authorization: Bearer $SUME_API_KEY"

What should I check before using the weights?

The tencent-hunyuan-community license text, which the card links; this post does not summarize it.
Which checkpoint you need: the base text-to-image model and the Instruct variants are listed separately on the card.
The hardware you have: the card gives no memory figure, and its Gradio demo defaults to four GPUs.

What is the Instruct-Distil checkpoint?

The card's Jan 26, 2026 news lists HunyuanImage-3.0-Instruct-Distil as a distilled checkpoint for efficient deployment with 8 steps of sampling recommended. It says the base text-to-image model needs the tencentcloud-sdk only for Prompt Enhancement, and that the Instruct model does not.

What else does the card list?

The news list also mentions vLLM acceleration (October 30, 2025), described as significantly faster inference. The card's stated goal is a native multimodal model that unifies understanding and generation in one autoregressive framework, instead of the DiT-based designs it says are prevalent.

What does a hosted image API change?

A hosted call needs no GPU on your side, and a catalog lists what you can call. The trade-off is that you use the catalog's models, not your own weights. AI image API price list shows how Sume prices its listed image models.