ByteDance Lance 3B tops unified multimodal benchmarks

Lance is a masterclass in architectural efficiency, proving that active parameters matter more than total scale for complex multimodal tasks.

• –Dual-stream MoE architecture effectively decouples understanding and generation while maintaining a shared underlying context
• –Native unified design eliminates the "franken-model" approach of stitching separate visual encoders and generative heads
• –3B size is ideal for edge experimentation but requires 40GB VRAM, suggesting high memory overhead for multimodal state
• –Apache 2.0 license makes it highly accessible for commercial use and community-driven video editing workflows
• –Exceptional scores on GenEval and VBench confirm its competitive edge against models double its size