Sign in
Pending benchmarksTracking since Nov 4, 2026

Is GroundingPI state of the art?Pending benchmarks.Announced Sep 30, 2026; no independent results yet.

XPENG's roughly 4B-parameter perception model for embodied AI, built with Peking University, the University of Hong Kong and other institutions. It grounds language in images, answering with boxes and points, and is trained only on perception data without action labels. It pairs a MoonViT-V2 vision encoder from Kimi K3 with Qwen3-4B-Instruct-2507. Open weights on Hugging Face and ModelScope.

Announced facts

The details so far

Lab
XPENG
Announced
Sep 30, 2026
API
Not yet
Open weights
Published · Hugging Face GroundingPI/GroundingPI, not gated; ModelScope from Oct 14
Parameters
4,843,664,384 stored parameters (BF16)
Task
Visual grounding for embodied perception: detection, referring expressions, OCR, layout, UI elements and robot spatial localization, output as boxes and points
Architecture
MoonViT-V2 vision encoder (from Kimi K3) with Qwen3-4B-Instruct-2507
Training data
Perception-only data in 11 capability categories, with no robot action labels
Context
16,384 tokens default serving context (model card)
License
Apache-2.0 for XPENG's contributions; Kimi K3 License for Kimi-derived parts
Paper
GroundingPI: A Grounding Foundation Model towards Physical Intelligence with Visual Primitives (arXiv 2609.39601)
Partners
Peking University, the University of Hong Kong and other institutions

Tags marked unsure are not stated on the lab's own pages.

Next steps

What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

  • Publicly announced with a source
  • Usable by the public (API or weights)
  • Listed on an independent board we track
  • Enough results for a rank in a category
  • Holds across two snapshots (crowns only)
Self-reported

What the lab claims

Self-reported. Shown for context; not used in any verdict.

Lab claim: 73.68 average across 34 perception benchmarks in 11 categories, vs 71.54 for GPT-6 Astra.

Lab claim: up to +24.8% success on RoboTwin 2.0 and 0.296 m L2 on nuScenes planning as a perception backbone.

Receipts

Sources

PAPERGroundingPI (arXiv 2609.39601)arxiv.org/abs/2609.39601 · Sep 30, 2026LABGroundingPI project pagegroundingpi.github.io/LABXPENG article on Xx.com/XPENG_Global/status/2117924611514069366 · Nov 4, 2026REPOGroundingPI/GroundingPI on Hugging Facehuggingface.co/GroundingPI/GroundingPIREPOgroundingpi/GroundingPI on GitHubgithub.com/groundingpi/GroundingPI

Tracking continues until an independent board lists it.