GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Apollo-ULTA
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Apollo-ULTA
Model Tree
Qwen/Qwen2.5-VL-3B-Instruct
Input Modalities
Output Modalities
Supported Functionality
DeepGrounder is a Qwen2.5-VL-3B-based model for remote sensing visual grounding: localizing the object referred to by a natural language description. The method first establishes domain knowledge through supervised fine-tuning, then applies GRPO with hybrid verifiable rewards for reasoning format, geometric accuracy, and vision-language semantic consistency.
This release provides checkpoint-6090 as a complete Transformers model, with two Safetensors weight shards, the weight index, config, and tokenizer/processor files. The weights use BF16. It is not an adapter-only release; a separate LoRA merge is not required.
The source checkpoint config identifies the Qwen2.5-VL architecture. The precise training-data/run provenance and correspondence between this checkpoint and individual paper table rows have not been independently re-benchmarked as part of this release. The paper's reported metrics below describe the paper experiments.
Install a CUDA-compatible PyTorch build, then:
bash
git clone https://github.com/ULTA-Web/DeepGrounder.gitcd DeepGrounderpip install -r requirements-inference.txthuggingface-cli download Apollo-ULTA/DeepGrounder --local-dir checkpoints/DeepGrounder python tools/prepare_dior_rsvg.py \ --data_root /path/to/DIOR_RSVG \ --output_dir data/processed/DIOR-RSVG python src/eval/infer_allbox.py \ --reasoning_model_path checkpoints/DeepGrounder \ --json_path data/processed/DIOR-RSVG/test_eval.jsonl \ --image_base_path /path/to/DIOR_RSVG/JPEGImages \ --output_dir outputs/DIOR-RSVGSee the inference guide for the JSONL schema, pixel-coordinate convention, and metrics. See the training guide for SFT and GRPO.
The repository walkthrough uses DIOR-RSVG, downloadable from the official project and its Google Drive. The paper also evaluates RRSIS-D and zero-shot transfer to VRSBench.
| Paper evaluation | Pr@0.5 (%) | Pr@0.7 (%) | mIoU (%) |
|---|---|---|---|
| DIOR-RSVG | 64.03 | 48.27 | 54.85 |
| RRSIS-D | 64.21 | 47.66 | 57.11 |
| VRSBench, zero-shot | 59.1 | 38.8 | — |
These are paper-reported values, not a new evaluation of this upload.
Intended for research on remote sensing visual grounding and multimodal spatial reasoning. Performance depends on image resolution, prompt, coordinate convention, and distribution. Generated explanations and boxes can be incorrect and should be verified. The repository's full hybrid-reward training additionally requires an external Git-RSCLIP semantic scoring service.
This repository retains its existing Apache-2.0 metadata. Consult the base model and dataset authors for their applicable terms; DIOR-RSVG has separate research/noncommercial restrictions.
bibtex
@article{zhang2026deepgrounder, title={DeepGrounder: Generalized Reasoning for MLLM-Based Remote Sensing Visual Grounding}, author={Zhang, Jiafan and Chen, Keyan and Liu, Chenyang and Lin, Baihong and Zou, Zhengxia and Shi, Zhenwei}, journal={IEEE Transactions on Geoscience and Remote Sensing}, year={2026}, volume={64}, doi={10.1109/TGRS.2026.3697769}, url={https://ieeexplore.ieee.org/document/11538158}}git clone https://github.com/ULTA-Web/DeepGrounder.gitcd DeepGrounderpip install -r requirements-inference.txthuggingface-cli download Apollo-ULTA/DeepGrounder --local-dir checkpoints/DeepGrounder python tools/prepare_dior_rsvg.py \ --data_root /path/to/DIOR_RSVG \ --output_dir data/processed/DIOR-RSVG python src/eval/infer_allbox.py \ --reasoning_model_path checkpoints/DeepGrounder \ --json_path data/processed/DIOR-RSVG/test_eval.jsonl \ --image_base_path /path/to/DIOR_RSVG/JPEGImages \ --output_dir outputs/DIOR-RSVG@article{zhang2026deepgrounder, title={DeepGrounder: Generalized Reasoning for MLLM-Based Remote Sensing Visual Grounding}, author={Zhang, Jiafan and Chen, Keyan and Liu, Chenyang and Lin, Baihong and Zou, Zhengxia and Shi, Zhenwei}, journal={IEEE Transactions on Geoscience and Remote Sensing}, year={2026}, volume={64}, doi={10.1109/TGRS.2026.3697769}, url={https://ieeexplore.ieee.org/document/11538158}}