ByteDance released an open-source model that attempts broad multimodal tasks with 3B parameters

bytedance released an open source model that attempts to do just about anything with only 3b parameters

ByteDance released Lance, an open-source unified multimodal model with 3B active parameters that supports image and video understanding, generation, and editing, and the post says it was trained from scratch with a staged multi-task recipe under a 128-A100-GPU budget.

Why it matters: HKR-H/K/R all pass: ByteDance’s open Lance has a compact multimodal hook, concrete 3B/128-A100 facts, and clear cost/deployment resonance. Reddit-sourced details lack benchmarks, license terms, and official context, so it stays featured, not P1.