OpenGVLab's InternVL2-1B Beats Larger 4B Vision Models at Under 2GB
OpenGVLab released InternVL2-1B , a 0.9B parameter vision language model under MIT license. Combines InternViT-300M-448px vision encoder with Qwen2-0.5B-Instruct language backbone via MLP projector. Scores 754 on OCRBench, beating larger PaliGemma-3B (614) and Mini-InternVL-2B (654).
Takeaways
- OpenGVLab released InternVL2-1B , a 0.9B parameter vision language model under MIT license.
- Combines InternViT-300M-448px vision encoder with Qwen2-0.5B-Instruct language backbone via MLP projector.
- Scores 754 on OCRBench, beating larger PaliGemma-3B (614) and Mini-InternVL-2B (654).
- Trained with 8k context on long texts, multi-image, and video inputs for richer multimodal reasoning.
- Supports vLLM, SGLang, LMDeploy serving, plus BNB 8-bit quantization for edge deployment.
- Successor InternVL2_5-1B is now available for new projects.
OpenGVLab has released InternVL2-1B , the smallest checkpoint in its InternVL 2.0 multimodal family. At roughly 0.9 billion parameters, the model handles OCR, document and chart analysis, image grounding, and video question answering while fitting into a substantially smaller memory budget than most general-purpose vision-language models.
The full InternVL 2.0 family ranges from 1 billion to 108 billion parameters. Its 1B checkpoint targets local applications, latency-sensitive services, and domain-specific fine-tuning, with reported results that match or exceed some 2B to 4B models on selected visual tasks.
Small language core, tiled vision
InternVL2-1B combines a 300M-parameter vision encoder with a compact language model and a learned projector that translates visual features into the language model’s embedding space.
| Component | Role |
|---|---|
| InternViT-300M-448px | Encodes 448 by 448 image tiles |
| MLP projector | Maps visual features into language-model tokens |
| Qwen2-0.5B-Instruct | Processes prompts and generates responses |
| 8K context window | Holds visual tokens, text prompts, and generated output |
Its preprocessor selects a supported aspect ratio and divides each image into as many as 12 tiles, with an optional thumbnail representing the full image. Keeping each tile at 448 by 448 helps preserve small text and chart labels that would blur when a large document is compressed into one square image.
InternVL 2.0’s instruction tuning includes long text, multiple images, and video sequences. That mixed training format gives the 1B checkpoint support for multi-image conversations and frame-based video analysis despite the limited capacity of its language backbone.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
