Pelican1.0-VL-7B

Pelican1.0-VL-7B is the entry-level build of Pelican-VL 1.0, the open-weights family of embodied brain models from the Beijing Innovation Center of Humanoid Robotics — the state-backed unit that also builds the Tiangong humanoids. It went public on 13 November 2025, the same day as the 72B flagship, under Apache 2.0. Like the rest of the family, it is not a chat assistant. It reads images and video together with text and answers the questions that come before a robot moves: what is where, what blocks what, where a plausible grasp point sits, and in what order a multi-step job should run. It inherits the Qwen2.5-VL architecture, keeps the full 128,000-token context of the larger builds, and was trained by the same DPPO loop — reinforcement learning, refinement, diagnosis and supervised fine-tuning, cycling so that each pass generates fresh hard examples aimed at the model's latest mistakes. The practical argument for this size is where it can run. The 72B build needs a multi-card server; this one fits a single high-memory accelerator, which puts the same perception stack on a robot's own computer rather than behind a network link. Two things are worth stating plainly, because the naming hides them. The checkpoint labelled 7B actually carries 8.29 billion parameters — the label follows the base model's class, not the count. And the developer's own report frames the family as spanning 7B to 72B; the 3B build that appeared two weeks later sits below the range the authors described. Downloads are in the single digits per month on Hugging Face, in line with the rest of the family.

▸Videos

No videos yet.

▸Reviews