ByteDance · 8× · vu le 30 juin 2026
Lance is an Open Source unified multimodal model with 3 billion activated parameters released by ByteDance (Intelligent Creation Lab), which combines image and video understanding, generation, and editing in a single architecture. It utilizes a Dual-Stream-Mixture-of-Experts architecture (initialized from Qwen2.5-VL-3B) and modality-aware rotatory positional encoding (MaPE) to reduce computational load and minimize signal interference between image and video Token. Despite its small size, the model was trained from scratch with a limited budget of up to 128 A100 GPUs and achieves competitive to leading results on Benchmarks for image/video generation, editing, and video understanding. Lance is released under the Apache 2.0 license.
Fonctionnalités
| Key Benchmark (%) | GenEval: 0.90 (overall score, top among unified models); DPG-Bench: 84.67 |
| License | Apache 2.0 |
| Multimodality | Image & video: understanding (captioning, VQA, OCR, grounding, reasoning), generation (text-to-image, text-to-video, image-to-video), and editing (image/video editing incl. multi-turn consistency) |
| Platform | Open source: GitHub repo (bytedance/Lance) and Hugging Face (bytedance-research/Lance), local inference via CLI/Gradio, supported in vLLM-Omni since 2026-06-03 |
| Release Date | May 18, 2026 (project page, GitHub, Hugging Face launched; technical report on arXiv May 19, 2026) |