Liquid AI's LFM2.5-VL-450M: Innovations in Vision-Language Models

Discover what's new in Liquid AI's LFM2.5-VL-450M: bounding box prediction, multilingual support and edge inference under 250ms.

Liquid AI's LFM2.5-VL-450M: Innovations in Vision-Language Models
Liquid AI logo with a depiction of LFM2.5-VL-450M in action on edge devices

Liquid AI has recently released LFM2.5-VL-450M, an advanced vision-language model that brings significant improvements over its predecessor, LFM2-VL-450M. This 450-million-parameter model is designed to run real-time inference on edge hardware, such as embedded AI modules and mini-PCs, with latency below 250 milliseconds.

Key Innovations in LFM2.5-VL-450M

Bounding Box Prediction

One of the main new features of LFM2.5-VL-450M is its ability to predict bounding boxes, allowing the model to identify and locate objects within an image. This advance improves visual understanding and accuracy in object detection applications.

Expanded Multilingual Support

The model now handles prompts in eight languages (Arabic, Chinese, French, German, Japanese, Korean, Portuguese and Spanish) with superior accuracy, extending visual understanding to a global level without the need for separate localization models.

Edge Inference with Reduced Latency

LFM2.5-VL-450M is optimized to run on edge devices, ensuring fast inference. For example, on NVIDIA Jetson Orin, the model processes a 512×512 image in less than 250 milliseconds, making it suitable for real-time applications such as industrial monitoring and surveillance.

Practical Applications

Industrial Automation

In resource-constrained environments, such as motor vehicles, agricultural machinery and warehouses, LFM2.5-VL-450M offers advanced semantic scene understanding, improving operational efficiency and safety.

Wearables and Continuous Monitoring

For devices such as smart glasses and dashcams, the model provides local visual understanding, reducing the need for cloud processing and improving user privacy.

Retail and E-Commerce

In retail, LFM2.5-VL-450M facilitates high-speed image processing, supporting tasks such as catalog ingestion, visual search and shelf compliance, improving operational efficiency.

Conclusion

LFM2.5-VL-450M represents a significant advance in vision-language models, offering:

  • Bounding Box Prediction: Improves accuracy in object detection and localization.
  • Expanded Multilingual Support: Broadens visual understanding on a global scale.
  • Low-Latency Edge Inference: Optimizes real-time applications on edge devices.

These features make LFM2.5-VL-450M a versatile and powerful solution for a wide range of applications in artificial intelligence and computer vision.