GroundingPI wins this comparison on localization. XPeng's open-source 4B model averaged 73.68 across 34 perception benchmarks, ahead of GPT-6 Astra at 71.54 (XPeng; arXiv). If you need precise boxes and points in cluttered scenes - robotics, GUI agents, document pipelines - it is the open-weights model to try first; keep your existing vision-language model for general reasoning, because GroundingPI is a specialist perception layer.
TL;DR
- GroundingPI: a 4B-class open-source grounding model announced by XPeng on 4 November 2026 - paper, weights and code all public.
- 73.68 on 34 perception benchmarks - highest of 44 baselines - against 71.54 for GPT-6 Astra; these are the research team's own numbers.
- Trained without robot action labels; outputs points and boxes as integer coordinate tokens.
- Layered licence: Apache 2.0 for original work, Kimi K3 licence terms apply to Kimi-derived material.
- Our estimate: about 9.7 GB of BF16 weights, but the model card lists only data-centre GPUs as tested.
What is GroundingPI and how does it work?
GroundingPI is a visual perception foundation model built by researchers affiliated with XPeng, Peking University, The University of Hong Kong, UC Berkeley, Princeton, NUS, Tsinghua and HKUST(GZ), per the XPeng pressroom. Grounding ties a phrase such as "the red car" to a precise location in an image.
Objects and positions are represented as visual primitives: points and bounding boxes written with a shared vocabulary of quantised coordinate tokens, the integers <0> through <999>, per the Hugging Face model card. Because the output is a short run of ordinary tokens, one language-model head handles detection, pointing and text recognition.
Training is perception-first and uses no robot action labels. The recipe has multimodal and spatial pretraining, supervised fine-tuning, then reinforcement learning with GRPO (arXiv).
How does GroundingPI compare with GPT-6 Astra?
On the headline average GroundingPI leads by 2.14 points - 73.68 against 71.54 - across 34 benchmarks in 11 categories covering object detection, dense localisation, referring grounding, OCR, document layout, UI recognition and robotic spatial localisation, with 44 baselines evaluated (XPeng).
The gap widens on crowded scenes: on Dense200 the paper's ablation tables give GroundingPI 74.53, GPT-6 Astra 65.04 and Qwen3-VL-4B 14.02 (arXiv PDF) - grounding-specific training matters more than raw size.
Treat all of these as vendor-side results. Several outlets repeated 73.68 versus 71.54 on announcement day - Sina Finance, ifeng, d1ev - but that corroborates the reporting, not a replication; no independent evaluation of the released weights exists yet.
What is under the hood?
The model pairs a MoonViT-V2 vision encoder, derived from the Kimi K3 vision stack, with a Qwen3-4B-Instruct-2507 language backbone. A 2x2 spatial-aggregation projector feeds a two-layer MLP, and the total is 4.84B parameters stored in BF16 (model card). For background on the Qwen lineage, see our comparison of Qwen, Claude and Kimi.
The ablations in the paper show where the gains come from: quantised coordinates lift the 34-benchmark average from 71.21 to 73.68, MoonViT-V2 scores 73.68 against 73.17 for MoonViT and 71.65 for Qwen3-ViT, and dropping the final training stage leaves 72.86.
Does it help robots and driving systems?
Somewhat, with caveats. As a visual backbone on RoboTwin 2.0 in simulation, it beat every mainstream backbone tested in all four out-of-distribution settings, by up to 24.8 percent relative to the strongest (XPeng); a RoboCasa-GR1 policy trained on 50 percent of demonstrations outperformed baselines trained on 75 percent, and on nuScenes it reached an average open-loop L2 error of 0.296 metres (arXiv).
XPeng states the research is exploratory, downstream results come from simulation and offline evaluation, and nothing commits it to a product. Open-loop error says little about closed-loop driving, so read these as feature-quality signals.
How do you run GroundingPI?
The repo serves the model via a vLLM 0.18.x integration plus a native Transformers reference service using a bundled Transformers 5.7.0 fork (model card). From the card:
- Download the weights:
hf download GroundingPI/GroundingPI - Prepare the environment:
python3 run.py setup serve - Start the service:
python3 run.py serve - Query it from Python:
client.predict("example.jpg", "the red car", task="bbox")
The default vLLM service uses BF16, eager execution and tensor parallel size 1. The card describes several task styles.
| Task style | What you ask for | Output |
|---|---|---|
| Category or dense grounding | Every instance of a named class | Boxes |
| Referring box | One object from a description | A box |
| Pointing | A location on an object | A point |
| OCR | Text in a region or image | Text, with location |
| Layout | Structure of a document page | Regions |
| GUI pointing | An interface element | A point |
| Visual prompting | Matches to a reference box | Boxes |
For GUI work, compare this with the agents covered in our guide to CUA-Master 27B and open computer-use agents. For wider tooling choices, see best AI agents for coding in 2026.
Will it fit on a consumer GPU?
Possibly, but nobody has confirmed it. Our arithmetic, not a vendor claim: 4.84B parameters at two bytes each come to about 9.7 GB of weights before KV cache and image tokens, within reach of a 12 to 16 GB card in principle.
| Item | Detail |
|---|---|
| Weights (our estimate) | About 9.7 GB in BF16 |
| Quantised checkpoint | None shipped |
| Tested hardware (model card) | NVIDIA B300, B200, H200, H800 and PPU |
| Consumer GPUs | Untested; expect to experiment |
Can you use GroundingPI commercially?
The model card says original contributions are Apache 2.0 with no extra restrictions, but third-party material keeps its own licences, including the Kimi K3 License for Kimi-derived material such as the vision encoder lineage (model card). Check the conditions applying to the combined package before you commit. For how derived weights are handled, see Heretic 2.0, ARA and abliteration.
For regional context on other open model releases, see our coverage of IndiaAI's Eka-1, Dhir-2B and Dhir-5B.
When did GroundingPI actually launch?
The 4 November announcement is XPeng's official spotlight of a release that was already public: paper on arXiv on 30 September 2026, GitHub code on 3 October, project page on 4 October, ModelScope weights on 14 October, and first among Hugging Face Daily Papers on 13 October, per the model card's release section.
Verdict
Use GroundingPI as a dedicated localisation layer: boxes, points, text regions and interface elements. It leads GPT-6 Astra on the team's own evaluation, is small enough to serve yourself and ships with working code. Hold off on production commitments until independent evaluations appear and your legal team has read the layered licence. Keep your general-purpose vision-language model for reasoning and conversation.
FAQ
Q: What is GroundingPI?
A: GroundingPI is an open-source visual perception model from researchers affiliated with XPeng and several universities, outputting points and bounding boxes for objects, text and interface elements.
Q: Does GroundingPI really beat GPT-6 Astra?
A: On the research team's evaluation: 73.68 across 34 benchmarks versus 71.54 for GPT-6 Astra (XPeng). Independent tests of the released weights have not yet been published.
Q: Is GroundingPI free for commercial use?
A: Original contributions are Apache 2.0, but third-party material keeps its licences, including the Kimi K3 License - check those terms for the combined package before shipping.
Q: What hardware does GroundingPI need?
A: The card lists NVIDIA B300, B200, H200, H800 and PPU as tested (model card). Our estimate is about 9.7 GB of BF16 weights, so smaller GPUs may work untested.
Q: Will XPeng put GroundingPI in its cars?
A: XPeng calls the work exploratory: downstream results are simulation and offline evaluations, and nothing commits it to product integration.
Q: Is GroundingPI a replacement for a general vision-language model?
A: No. It specialises in localisation and perception; keep your existing vision-language model for general reasoning.
Top comments (0)