Open weights · Card read 2026-11-06 · Not loaded here
Hunyuan Video 1.5 runs on your GPU, not in this form
Tencent published the weights, the install steps, and a 8.3B text-to-video and image-to-video model. Nano AI does not host that stack.
This generator runs Kling 2.6 for a silent 5-second clip at 55 credits, with sound: false. Sign-up credits are 64, enough for one such clip. It does not load Hunyuan Video 1.5.
Duration
Ratio
Generated Videos
No videos yet
Fill out the form and click Generate Video to create your first AI video.
Tips: Use detailed descriptions for better results
- Made by
- Tencent
- Released
- November 20, 2025
- Run on Nano AI
- Kling 2.6 · from 55 credits per video
- Page updated
- November 6, 2026
Source: huggingface.co
Cards, not clips
Both images were downloaded on 2026-11-06. The Hugging Face card is the repository name. The GitHub card's counters are GitHub's at download time, not a Nano AI measurement, and neither image is a Hunyuan Video 1.5 frame.
What Hunyuan Video 1.5 publishes
The Hugging Face model card, read on 2026-11-06, says the inference code and weights were released on 20 November 2025. It describes Hunyuan Video 1.5 as an 8.3B Diffusion Transformer with a 3D causal VAE, compression of 16× in space and 4× in time, and selective and sliding tile attention. It says that attention is 1.87× faster than FlashAttention-3 on a 10-second 720p clip in their comparison. A super-resolution network upscales toward 1080p. The license name on the card is tencent-hunyuan-community. This page does not summarize the license terms beyond that name and the link.
The card's hardware line is an NVIDIA GPU and 14 GB minimum with offloading. Software lines are Linux and Python 3.10 or higher. Defaults on the card include video_length 121, 50 steps, and super-resolution on unless you pass --sr false. A 480p image-to-video step-distilled model is described as 8 or 12 steps, about 75% faster on an RTX 4090. The card still prints Comming soon, their spelling, for HunyuanVideo-1.5-720P-T2V-cfg-distill and HunyuanVideo-1.5-720P-T2V-sparse-cfg-distill. Evaluation widgets on the card showed 78.2 and 74.3. Those are their figures, not a Nano AI benchmark. The paper linked from the card is arXiv 2511.18870.
checkpoints-download.md says the latest HunyuanMLLM is not released and recommends Qwen2.5-VL-7B-Instruct. It also lists google/byt5-small, Glyph-SDXL-v2 from ModelScope, and a SigLIP download from black-forest-labs/FLUX.1-Redux-dev that needs an approved Hugging Face token. This site does not request that access and does not supply a token. The form below is Kling 2.6 only.
Hunyuan Video 1.5 at a glance
- Weights date
- 20 November 2025, on the model card
- Size on the card
- 8.3B, 16× space, 4× time
- Their speed line
- 1.87× vs FlashAttention-3, 10-second 720p
- Their minimum
- 14 GB with offloading, Linux, Python 3.10
- Their defaults
- 121 frames, 50 steps, 480p or 720p, SR toward 1080p
- Step distill
- 480p I2V, 8 or 12 steps, 75% on an RTX 4090
- This form
- Kling 2.6, 5 seconds, 55 credits, sound: false
Install Hunyuan Video 1.5 on your own machine
Commands are copied from the model card and from checkpoints-download.md as read on 2026-11-06. Running them does not happen on Nano AI.
- 1
Clone the repository
The model card's first step. This page does not run the clone.
git clone https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5.git cd HunyuanVideo-1.5 - 2
Install the Python packages named on the card
The second line uses the Tencent pip mirror printed on the card.
pip install -r requirements.txt pip install -i https://mirrors.tencent.com/pypi/simple/ --upgrade tencentcloud-sdk-python - 3
Download the DiT and VAE checkpoints
checkpoints-download.md also says to install the Hugging Face CLI and ModelScope first. The hf download line is theirs.
pip install -U "huggingface_hub[cli]" pip install modelscope hf download tencent/HunyuanVideo-1.5 --local-dir ./ckpts - 4
Download the text encoders they recommend
The doc says HunyuanMLLM is not released yet, so it points at Qwen2.5-VL-7B-Instruct, byt5-small, and Glyph-SDXL-v2.
hf download Qwen/Qwen2.5-VL-7B-Instruct --local-dir ./ckpts/text_encoder/llm hf download google/byt5-small --local-dir ./ckpts/text_encoder/byt5-small modelscope download --model AI-ModelScope/Glyph-SDXL-v2 --local_dir ./ckpts/text_encoder/Glyph-SDXL-v2 - 5
Vision encoder, only with your own token
The doc says SigLIP is taken from FLUX.1-Redux-dev after Hugging Face approves your access. Leave the placeholder until you have a token. This site does not provide one.
hf download black-forest-labs/FLUX.1-Redux-dev --local-dir ./ckpts/vision_encoder/siglip --token <your_hf_token>
Download targets named in the checkpoint doc
Sizes are not printed on the card read on 2026-11-06, so this table does not invent them. The two 720p rows marked Comming soon are omitted.
| File | Role | Note |
|---|---|---|
| tencent/HunyuanVideo-1.5 | DiT and VAE | hf download into ./ckpts |
| Qwen2.5-VL-7B-Instruct | Text encoder they recommend | HunyuanMLLM not released, per the doc |
| google/byt5-small | byT5 encoder | Paired with Glyph-SDXL-v2 |
| Glyph-SDXL-v2 | Glyph encoder | ModelScope download |
| FLUX.1-Redux-dev | SigLIP vision encoder | Gated. Your token, not ours |
What this Hunyuan Video 1.5 page will not run
No weights on this server
The clone, the hf download, and the GPU stay on your machine. Nano AI does not install Hunyuan Video 1.5.
One priced clip
Kling 2.6, 5 seconds, 55 credits, sound: false. 64 sign-up credits cover one clip. That is not 480p or 720p Hunyuan output.
Gated vision encoder
The checkpoint doc says FLUX.1-Redux-dev needs your own approved token. The command below keeps the placeholder <your_hf_token>.
Coming-soon rows stay blank
Two 720p distill rows are still marked Comming soon on the card. This page does not pretend those files downloaded.
Hunyuan Video 1.5 beside this form
Left column is the model card and checkpoints-download.md read on 2026-11-06. Right column is what Nano AI actually runs.
| Hunyuan Video 1.5 | This page | |
|---|---|---|
| Where it runs | Your NVIDIA GPU | This website |
| What you get | 480p or 720p, SR toward 1080p | A 5-second Kling clip |
| Audio | Not described as a soundtrack on the card | sound: false |
| Extra downloads | Qwen2.5-VL-7B-Instruct, byt5-small, Glyph-SDXL-v2, FLUX.1-Redux-dev | None |
| Price here | Electricity and your own GPU | 55 credits; 64 at sign-up |
Pages this install follows
Same links in every language. Read dates are 2026-11-06.
How to use this page
- 1
Treat the install as their stack
The commands further down are copied from the model card and from checkpoints-download.md. They do not run when you press the button on this page.
- 2
Use the form for a different model
Kling 2.6 makes one silent 5-second clip at 55 credits. It is not Hunyuan Video 1.5 and it does not use your GPU weights.
- 3
Bring your own token if you install
FLUX.1-Redux-dev is gated. Replace <your_hf_token> yourself after Hugging Face approves access. This site has no token to give you.
Presets for the Kling clip
These prompts were written for the Kling form. They are not the model card's demo prompts, and they do not ask for a person.
Lantern
A paper lantern sways once in an empty stairwell, cool wall, warm light, camera still, no people, no text.
Tea set
Rain beads on a tea set beside a window, one drop falls, no hands, no logos.
Ferry
A small ferry crosses an empty harbor at dawn, gulls distant, no readable names on the hull.
Ink
Ink spreads on wet paper into a mountain shape, then settles, no letters, overhead view.
When the install does not match the card
These are limits printed by Tencent, not errors Nano AI can clear.
- Out of memory under 14 GB
- The card says 14 GB is the minimum with offloading. It also says to turn offloading off only when the GPU has room. This page did not test a GPU.
- A 720p distill file is missing
- HunyuanVideo-1.5-720P-T2V-cfg-distill and the sparse T2V twin are still marked Comming soon. Do not invent a filename for them.
- The vision download asks for a token
- FLUX.1-Redux-dev is gated. A lockfile error mentioned in the doc, ckpts/.huggingface/.gitignore.lock, is something they say you can ignore and rerun. This site does not proxy that download.
- The clip on this page is the wrong model
- That is expected. The button runs Kling 2.6 at 55 credits with sound: false. It is a different product from Hunyuan Video 1.5.
FAQ
Questions
Answers from the model card and the checkpoint doc read on 2026-11-06, plus the price of the form on this site.
Can I run Hunyuan Video 1.5 here?
No. This form runs Kling 2.6, 5 seconds, 55 credits, sound: false. The weights stay in the install section for your own machine.
What does 8.3B refer to?
The model card's parameter count for the Diffusion Transformer. It also states 16× spatial and 4× temporal compression. Nano AI did not remeasure that.
Why is Qwen2.5-VL-7B-Instruct in the steps?
checkpoints-download.md says the latest HunyuanMLLM is not released and recommends that Qwen model as the text encoder. This page repeats that recommendation and does not claim it is Tencent's own MLLM.
What is still marked coming soon?
The card prints Comming soon for HunyuanVideo-1.5-720P-T2V-cfg-distill and HunyuanVideo-1.5-720P-T2V-sparse-cfg-distill. Those two are not listed as files you already have.
Do 64 credits load the 14 GB model?
No. 64 sign-up credits cover one 55-credit Kling clip. 14 GB is their offloaded GPU minimum, not a Nano AI price.
Is Nano AI affiliated with Tencent?
No. Nano AI is not affiliated with Tencent or the Hunyuan Video 1.5 repository. The card was read on 2026-11-06.
More from Nano AI
Related pages
Nano AI is an independent service and is not affiliated with Tencent or Hunyuan Video 1.5. The model card, checkpoints-download.md, and the paper link were read on 2026-11-06. Speed, memory, and the scores 78.2 and 74.3 are theirs. The form runs Kling 2.6.