What it is
HunyuanImage 3.0 Instruct was released on January 26, 2026, four months after the original HunyuanImage 3.0 text-to-image model. It adds image-to-image generation, covering editing and multi-image fusion, plus a reasoning step that rewrites and plans from the prompt before drawing. A distilled version that samples in 8 steps instead of 50 was released the same day; the weights have not been updated since February 2026.
Unlike most image models, it is an autoregressive mixture-of-experts model that handles understanding and generation in one network: 80 billion parameters in total, 13 billion active per token, across 64 experts. Tencent lists at least eight 80GB GPUs for both the Instruct and the distilled model, against three for the base text-to-image model.
The weights use the Tencent Hunyuan Community Licence, which does not apply in the European Union, the United Kingdom or South Korea, bars using outputs to improve other AI models, and requires a separate licence from Tencent for services with more than 100 million monthly active users.
Our review
HunyuanImage 3.0 Instruct is a strong editor wrapped in impractical terms. On Artificial Analysis's image-editing board it scores 1064 Elo, twenty-eighth overall and third among open-weights models behind Fibo 2.0 (1089) and Qwen-Image-2.1 (1082). It sits about 70 points clear of Qwen-Image-Edit-2511 at 992. As a text-to-image model it ranks far lower, fifty-fifth at 987, so its value is in editing.
The hardware bill is the first obstacle. Tencent's minimum is eight 80GB GPUs, and the distilled 8-step version needs the same eight; it only cuts time per image, not memory. At Lambda's $3.99 per H100 per hour, an always-on node runs to about $23,000 a month. There is no hosted API for the Instruct model, so that node is the only route.
The licence is the bigger problem for many readers. It does not apply at all in the EU, the UK or South Korea, and Clause 5(c) forbids using the model or its outputs there, which rules out a European marketing team even if the GPUs sit in Virginia. A second clause bars using outputs to train or improve any non-Hunyuan model, a real constraint for teams generating synthetic training data.
Tencent's own pipeline has moved on. Artificial Analysis now lists a hosted HunyuanImage 3.5 Preview at $24 per 1,000 images, while the 3.0 Instruct weights have not changed since February. Downloads, at about 41,000 a month on Hugging Face, are under a quarter of Qwen-Image-Edit-2511's.
Where it earns its place
An ML platform team at a Singapore e-commerce group with spare H100 nodes
Run a self-hosted editor for catalogue images that beats Apache-licensed options on quality
Serve the distilled checkpoint on one 8×80GB node with vLLM and FlashInfer. It edits about 70 Elo above Qwen-Image-Edit-2511; the cost is the node, roughly $23,000 a month at Lambda's on-demand H100 rate if bought rather than owned.
A creative-tools startup in the US with under 100 million users
Offer reasoning-led edits ('make this poster feel like a winter sale') without a closed-model API bill
The model rewrites vague instructions before drawing, which suits non-expert prompts. The startup must geo-block the EU, UK and South Korea and pass Tencent's use restrictions into its own terms, as the licence requires.
A research group studying unified multimodal models
Inspect a single network that both understands and generates images
The autoregressive mixture-of-experts design (80B total, 13B active) is unusual among image models. Research use is allowed in the licensed territory, but distilling it into another model is barred.
Before you commit
- Setup
- Tencent's guide needs CUDA 12.8, PyTorch, FlashAttention and FlashInfer; the first run with FlashInfer spends about 10 minutes compiling kernels. Default sampling is 50 steps, or 8 with the Distil checkpoint.
- Integrations
- Hugging Face Transformers-style loading, Tencent's vLLM path, a Gradio demo app. No hosted endpoint for Instruct.
- Data and privacy
- Self-hosted, so images stay on your hardware. Licence: void in the EU, UK and South Korea including for outputs; outputs may not improve non-Hunyuan models; services above 100 million monthly users need Tencent's permission; an acceptable-use policy applies.
- Cost at scale
- Eight 80GB GPUs minimum: at Lambda's $3.99 per H100 per hour that is about $32 an hour, or about $23,000 a month always on. For comparison, hosted HunyuanImage 3.0 on Tencent Cloud is $32 per 1,000 images, without Instruct's editing.
- Limits
- Tencent publishes no throughput or latency figures; no API; text-to-image quality ranks fifty-fifth on Artificial Analysis; weights last updated February 2026.
What's good
- Third-best open model on Artificial Analysis's editing board at 1064 Elo
- Reasoning step plans and rewrites the prompt before generating
- Generation, editing and multi-image fusion in one model
- vLLM support and a distilled 8-step checkpoint
What's not
- Minimum eight 80GB GPUs, even for the distilled version
- Licence void in the EU, UK and South Korea, including for outputs
- Outputs may not be used to improve other AI models
- No hosted API; fifty-fifth on text-to-image
Best for
Teams outside Europe and South Korea that already run eight-GPU inference nodes and want a top open editing model with reasoning.
Alternatives
HunyuanImage 3.0 Instruct: common questions
- Is HunyuanImage 3.0 Instruct worth it?
- HunyuanImage 3.0 Instruct edits like a paid model and deploys like a research project: eight 80GB GPUs, no API, and a licence that is void across the EU and UK. We rate it 3.5 out of 5.
- How much does HunyuanImage 3.0 Instruct cost?
- Free weights under the Tencent Hunyuan Community Licence (not valid in the EU, UK or South Korea). No hosted API for Instruct; Tencent Cloud's HunyuanImage 3.0 is $32 per 1,000 images per Artificial Analysis.
- What are the downsides of HunyuanImage 3.0 Instruct?
- Minimum eight 80GB GPUs, even for the distilled version; licence void in the EU, UK and South Korea, including for outputs; outputs may not be used to improve other AI models; no hosted API; fifty-fifth on text-to-image.
- Who is HunyuanImage 3.0 Instruct best for?
- Teams outside Europe and South Korea that already run eight-GPU inference nodes and want a top open editing model with reasoning.
- How much does HunyuanImage 3.0 Instruct cost for a team or in production?
- Eight 80GB GPUs minimum: at Lambda's $3.99 per H100 per hour that is about $32 an hour, or about $23,000 a month always on. For comparison, hosted HunyuanImage 3.0 on Tencent Cloud is $32 per 1,000 images, without Instruct's editing.
- Is HunyuanImage 3.0 Instruct safe for company data?
- Self-hosted, so images stay on your hardware. Licence: void in the EU, UK and South Korea including for outputs; outputs may not improve non-Hunyuan models; services above 100 million monthly users need Tencent's permission; an acceptable-use policy applies.
- What are the best HunyuanImage 3.0 Instruct alternatives?
- The closest alternatives we have reviewed are Qwen-Image-2.1, Qwen-Image-Edit-2511, FLUX 3 Image Klein and Nano Banana 2.1.
Answers checked Oct 22, 2026.