Models in the 7B parameter class have long been a favorite for edge deployment and individual developers.

Their lightweight profile lets them adapt flexibly to a range of endpoint scenarios, while strong performance can cover high-frequency needs such as image information extraction, document understanding, video analysis, and object localization.

Huawei has just introduced a major new open-source entrant, openPangu-VL-7B, aimed squarely at this core use case.

With an Ascend-native model architecture, openPangu-VL-7B delivers highly cost-effective inference performance:

For 720P images, first-token inference latency on a single Ascend Atlas 800T A2 card, combining ViT and LLM model latency, is just 160 milliseconds, enabling real-time inference at 5 FPS; during training, MFU reaches 42.5%.

More notably, the model completed stable, spike-free long-duration cluster training on more than 3 trillion tokens during pretraining, offering a valuable practical reference for developers using Ascend clusters.

openPangu-VL-7B stands out across core tasks including general visual question answering, document and chart understanding and OCR, visual grounding, and short-video understanding. It outperforms similarly sized models on open-source leaderboards, showing strong all-around capability.

The official cookbook also demonstrates the model’s strong performance in these areas.

For example, when given an image of a dish and asked to find how many cherry tomatoes there are, the model can identify all their locations and count them correctly.

When given a screenshot from an annual report, the model can also convert it into Markdown format, sparing users the pain of manual extraction.

Beyond its strong benchmark results and Ascend-oriented training and inference optimizations, the technical report discloses several core technical details, revealing the design choices behind the model’s high performance:

High-Performance Visual Encoder Adapted for Ascend

Traditional visual encoders in the industry are mostly designed for GPU architectures and do not fully exploit Ascend hardware advantages.

Through extensive pilot experiments and performance analysis, the team found an optimal balance in the model structure: at the same parameter count, this visual encoder delivers 15% higher throughput on Ascend chips than ViT-H series encoders using window attention.

At the same time, it adopts a multi-label contrastive learning framework, giving the model stronger fine-grained understanding and laying a solid foundation for learning visual grounding data in later VLM training.

Sample-Balanced Loss Design

To address learning balance across training samples of different lengths, openPangu-VL-7B introduces a hybrid training scheme combining “weighted per-sample loss + per-token loss,” with weighting coefficients dynamically determined by token position and sample importance.

This design allows the model to fully learn from long-response data during training without neglecting short-response information, avoiding a trade-off between the two. Ablation experiments have fully validated its effectiveness.

Padded Grounding Data Format

Unlike the industry’s mainstream 0-999 grounding scheme, openPangu-VL-7B uses padded relative coordinates in the 000-999 permille format for visual grounding.

Neatly using three tokens for position regression not only reduces the model’s learning difficulty, but also significantly improves format compliance, raising both accuracy and efficiency in grounding tasks.

In addition, the technical report takes a deeper look at key strategies including pretraining data mix, positional encoding, and model fusion, providing developers with a comprehensive technical reference.

For Ascend users, the open-sourcing of openPangu-VL-7B is clearly good news.

This multimodal model, combining a lightweight footprint, high performance, and broad general-purpose capability, gives edge developers and individual users a new option while further expanding application scenarios in the Ascend ecosystem and adding fresh momentum for innovation.

、