Pelican1.0-VL-3B

Pelican1.0-VL-3B is the smallest member of the Pelican-VL 1.0 family from the Beijing Innovation Center of Humanoid Robotics, published on 1 December 2025 under Apache 2.0. It is an afterthought in the literal sense. The centre's own report, released on 30 October 2025, describes Pelican-VL 1.0 as a family "with parameter scales ranging from 7B to 72B"; the 7B and 72B checkpoints duly appeared on 13 November. This build, at 3.75 billion parameters, went up eighteen days later and sits below the range the authors had set out. The report was never updated to include it. What it does is unchanged from its larger siblings: it reads images and video with text and answers the questions that precede a robot's movement — spatial relations, occlusion, plausible grasp points, the order of a multi-step task. It shares the Qwen2.5-VL base architecture, the DPPO training loop and, notably, the full 128,000-token context. Cutting the parameter count did not cut the window, so a long video sequence still fits in a single pass. The size is the point: at under four billion parameters the model is a candidate for on-board computers with modest memory, where the 72B flagship is out of the question and even the 7B build is tight. No separate benchmark table has been published for this build. The percentages quoted in the family's report describe the 72B model, and applying them to this one would be unwarranted. Downloads on Hugging Face run at single digits a month.

▸Videos

No videos yet.

▸Reviews