Tencent dropped their Hunyuan video model today, with 13 billion parameters and full weights available open source. This makes it the largest open source video model out there right now.
The model can generate film-quality videos with native scene switching between realistic and virtual styles. It handles high-dynamic motion scenes smoothly and can execute multiple actions in sequence.
People are already excited about testing it. Some worry that 13B parameters seems small for a video model, but for context – LTX is 2B and Cog is 2-5B parameters. So this is actually quite large.
Tencent built this specifically for commercial use cases like advertising and creative videos. The model already beat out competitors in blind tests with thousands of questions.
The technical specs look solid. It uses a spatial-temporal compression setup with Causal 3D VAE, text prompts encoded through a large language model, and output latents decoded to images/videos.
Right now it needs 45GB VRAM to run, but the open source community will likely optimize that down significantly. We’ve seen this pattern before with other large models.
You can access it through:
– Tencent Yuanbao APP
– Tencent Cloud service interface
– Hugging Face
– GitHub
This release continues the trend we saw with OpenAI’s Sora leak and subsequent artist backlash (see my coverage here: https://adam.holter.com/openai-sora-leak-artists-fight-back-against-tech-giant/). But Tencent took a different approach by open sourcing their model completely.
I’ll be testing this out myself and will report back with hands-on results. Let me know if you’d like to see anything specific in my testing.
