Overview
Lance is a research-oriented, unified multimodal model with approximately 3 billion active parameters. It uses a dual-stream mixture-of-experts architecture to support image and video understanding, generation, and editing within one framework. It is a research artifact rather than a polished product model. (github.com)
Capabilities and uses
- Generate images and videos from text, and generate video from an image.
- Edit images and videos; understand visual inputs for tasks such as captioning and question answering.
- Support research and experimentation with unified multimodal models. The repository provides inference and fine-tuning code; its task examples use separate image and video checkpoints. (github.com)
Limitations and deployment
The released checkpoint was trained for image generation up to 768 × 768 and video generation up to 480p at 12 frames per second. Output quality can vary with the prompt, resolution, duration, motion complexity, and editing scenario. The documented inference environment requires Python 3.10+, CUDA 12.4+, and a GPU with at least 40 GB of VRAM. The model card lists an Apache 2.0 license. (github.com)
Release
The project homepage, initial inference code, and model weights were released on May 18, 2026. Image-to-video support and fine-tuning code were added later. (github.com)
Pricing
As of November 16, 2026, no official Lance-specific API input or output prices could be verified in either CNY on the domestic Volcano Engine pricing documentation or USD on the international BytePlus ModelArk pricing documentation. No billing unit or rate is therefore stated. Self-hosting may still incur infrastructure costs. The international pricing page required JavaScript and its available text did not list Lance. (docs.volcengine.com)