ByteDance Research · 2026-05-18 · notable
ByteDance Lance — 3B Unified Multimodal Model Handles Image and Video Generation, Editing, and Understanding in One Stack
Lance is a 3B-active-parameter unified multimodal model from ByteDance Research that does text-to-image, text-to-video, image and video editing, and image/video understanding in a single network.
One 3B-active model that generates, edits, and understands both images and video — trained from scratch on 128 A100s.
Key specs
| Active params | 3B |
|---|---|
| Training budget | 128 A100 GPUs |
| Vbench score | 85.11 |
| Gen eval score | 0.90 |
| Gedit bench avg | 7.30 |
What is it?
Lance is ByteDance Research's unified multimodal model with 3B active parameters, fine-tuned from Qwen2.5-VL-3B. It supports six tasks in a single network: text-to-image, text-to-video, image editing, video editing, image VQA/reasoning, and video VQA/captioning. Weights are released under Apache 2.0.
How does it work?
The team trained Lance from scratch on a 128-A100-GPU budget using a staged multi-task recipe that lets generation, editing, and understanding heads share representations. The 3B model hits 0.90 on GenEval and 85.11 on VBench — the highest VBench score among unified models in the paper — and averages 7.30 on GEdit-Bench. Inference fits on a single 40GB+ GPU.
Why does it matter?
Most unified models lean on much bigger backbones or specialise in one modality. Lance shows you can hit strong numbers across generation, editing, and understanding with a compact open model, lowering the cost of running multimodal research and shipping products on consumer-class hardware.
Who is it for?
Multimodal researchers, indie video-AI builders, anyone running diffusion or VLM workloads on a single GPU
Try it
git clone https://github.com/bytedance/Lance