nvidia/EGM-4B
nvidia/EGM-4B is an 8 billion parameter efficient visual grounding language model developed by NVIDIA, built upon Qwen3-VL-4B-Thinking. It is specifically designed for visual grounding tasks, demonstrating superior performance by outperforming much larger models in average IoU on benchmarks like RefCOCO. This model achieves high accuracy in identifying regions described by complex prompts while maintaining faster inference speeds.
Loading preview...
EGM-Qwen3-VL-4B: Efficient Visual Grounding
EGM-Qwen3-VL-4B is an 8 billion parameter visual grounding model from NVIDIA's EGM (Efficient Visual Grounding Language Models) family. Built on the Qwen3-VL-4B-Thinking architecture, this model is engineered to achieve high performance in visual grounding tasks, where it identifies specific regions in images based on textual descriptions.
Key Capabilities & Differentiators
- Superior Visual Grounding Performance: Achieves an impressive 90.9 average IoU on the RefCOCO benchmark, a significant +3.7 IoU improvement over its base model (Qwen3-VL-4B-Thinking).
- Efficiency: Outperforms much larger models, including Qwen3-VL-235B-A22B-Instruct (88.2 avg IoU), while offering dramatically faster inference speeds.
- Mitigates Text Understanding Gap: Addresses the primary limitation of small VLMs in handling complex prompts by generating numerous mid-quality tokens to match the performance of larger models.
- Advanced Training Pipeline: Utilizes a two-stage training process involving supervised fine-tuning (SFT) with proprietary VLM-generated chain-of-thought reasoning, followed by reinforcement learning (RL) using GRPO (Group Relative Policy Optimization) for enhanced accuracy.
When to Use This Model
- Visual Grounding Applications: Ideal for tasks requiring precise localization of objects or regions within images based on natural language queries.
- Resource-Constrained Environments: Suitable for scenarios where high accuracy in visual grounding is needed but computational resources or inference speed are critical considerations.
- Complex Query Handling: Excels at interpreting and grounding complex textual descriptions, including those with multiple relational elements.