Tencent's HunyuanVideo, a 13B open-weights text-to-video diffusion transformer. Produces high-motion, photorealistic clips with smooth temporal consistency and was one of the first open models to rival closed systems on motion quality.
Real outputs from the public examples of this model on Replicate, with the prompt and settings that produced them. They were not generated live on this page.
Close-up, A little girl wearing a red hoodie in winter strikes a match. The sky is dark, there is a layer of snow on the ground, and it is still snowing lightly. The flame of the match flickers, illuminating the girl's face intermittently.
a stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage. She wears a black leather jacket, a long red dress, and black boots, and carries a black purse. She wears
sunglasses and red lipstick. She walks confidently and casually. The street is damp and reflective, creating a mirror effect of the colorful lights. Many pedestrians walk about.
Camera gliding through a vibrant coral reef teeming with tropical fish, sunlight filtering through the crystal-clear water creating dancing light patterns on the ocean floor, documentary style
A tiny glowing fairy emerging from an opening flower bud in an enchanted forest, sparkles trailing behind her wings as she takes flight, magical atmosphere with bioluminescent plants
HunyuanVideo from Tencent is a large open-source text-to-video model built on a DiT architecture with a unified full-attention design. It is known for realistic physics, large-scale motion and good text alignment, and ships its own weights, making it a popular base for fine-tuning and LoRA training.
Tencent Holdings is one of the largest technology and entertainment conglomerates in the world, founded in 1998 by Ma Huateng (Pony Ma) and four co-founders in Shenzhen. Tencent's AI Lab, founded in 2016, and the Hunyuan team (the company's foundation-model group) developed the Hunyuan family of text, image, 3D and video models. HunyuanVideo (released December 2024 with the 13B foundation model open-sourced) was the largest publicly released open-weight video diffusion model at the time, and rapidly became a popular base for community fine-tunes (LoRAs, control nets, audio-driven extensions). The model is hosted on Hugging Face and GitHub under a custom licence that allows research and limited commercial use.
Diffusion Transformer with dual-stream design (separate text/video streams)
HunyuanVideo is a 13B-parameter Diffusion Transformer (DiT) operating on a 3D causal VAE latent. Its denoiser uses a dual-stream / single-stream design inspired by FLUX-1: dedicated text and video streams first process their modalities separately and then concatenate for joint self-attention. Text conditioning combines a CLIP-like vision-language encoder and a multilingual large-language-model text encoder (MLLM) for stronger prompt understanding, especially on long captions. The model is trained with Flow Matching and 3D Rotary Position Embeddings, and uses a progressive curriculum from images to short videos to long videos at 720p / 24 fps. The accompanying paper details a meticulously curated multi-billion-clip dataset with hierarchical filtering and dense bilingual captions. HunyuanVideo supports text-to-video natively and the open release includes a high-quality 5-second / 720p model; community follow-ups added image-to-video, audio-driven and longer-duration variants.
13B open-weight text-to-video model (one of the largest released openly)
Generates ~5-second clips at 720p / 24 fps natively
Bilingual Chinese/English prompt handling via MLLM text encoder
Dual-stream DiT architecture inspired by FLUX-1
Strong prompt adherence on long, dense captions
Active community ecosystem (image-to-video, LoRAs, audio-sync)
Runs on 60-80 GB GPU memory (FP16), 40 GB with optimisations
Permissive licence for non-commercial research and limited commercial use
Best for: open-source video pipelines, research, on-prem creative tooling.
Entraînement et licence
Multi-billion-clip curated video corpus with hierarchical aesthetic, motion and caption-quality filtering, plus dense bilingual captions produced by an in-house MLLM captioner.
Licence: Tencent Hunyuan Community Licence: weights free for research and limited commercial use with attribution; restrictions apply above thresholds and for certain jurisdictions.
curl https://railwail.com/api/v1/videos/generations \
-H "Authorization: Bearer $RAILWAIL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "hunyuan-video",
"prompt": "A slow drone shot over a misty pine forest at sunrise"
}'
# Response: {"job_id": "...", "status": "queued", ...}
# Poll until status is completed, failed or cancelled:
curl https://railwail.com/api/v1/jobs/JOB_ID \
-H "Authorization: Bearer $RAILWAIL_API_KEY"
import os
import time
import requests
API = "https://railwail.com/api/v1"
headers = {"Authorization": f"Bearer {os.environ['RAILWAIL_API_KEY']}"}
job = requests.post(
f"{API}/videos/generations",
headers=headers,
json={
"model": "hunyuan-video",
"prompt": "A slow drone shot over a misty pine forest at sunrise",
},
).json()
while True:
status = requests.get(f"{API}/jobs/{job['job_id']}", headers=headers).json()
if status["status"] in ("completed", "failed", "cancelled"):
break
time.sleep(5)
print(status["status"], status.get("output_url"))
const API = "https://railwail.com/api/v1";
const headers = {
Authorization: `Bearer ${process.env.RAILWAIL_API_KEY}`,
"Content-Type": "application/json",
};
const job = await fetch(`${API}/videos/generations`, {
method: "POST",
headers,
body: JSON.stringify({
model: "hunyuan-video",
prompt: "A slow drone shot over a misty pine forest at sunrise"
}),
}).then((r) => r.json());
let status;
do {
await new Promise((r) => setTimeout(r, 5000));
status = await fetch(`${API}/jobs/${job.job_id}`, { headers }).then((r) => r.json());
} while (!["completed", "failed", "cancelled"].includes(status.status));
console.log(status.status, status.output_url);
Tencent Hunyuan Community Licence: weights free for research and limited commercial use with attribution; restrictions apply above thresholds and for certain jurisdictions.
