Descriptions:
Fahd Mirza puts Liquid AI’s LFM2.5-VL-450M through its paces in a hands-on local deployment, walking through exactly what this compact 450-million-parameter vision-language model can and cannot do when run entirely on CPU. Built on Liquid AI’s proprietary recurrent architecture (not a standard transformer), the model pairs an LFM 2.5 350M language backbone with a SigLIP 86M vision encoder, handling images up to 512×512 natively and supporting bounding box prediction, function calling, and nine languages.
The video covers a full installation using vllm, serving the model via Open WebUI with minimal CPU and memory overhead. Test cases include image captioning (a flag recognition task that the model gets wrong — documented transparently), multilingual OCR across all nine supported languages (English, French, German, Portuguese, and Spanish transcribed accurately; Arabic, Japanese, Korean, and Chinese all failed), and bounding box object detection using normalized JSON coordinate output.
Mirza’s honest assessment is that the model’s sweet spot is basic-to-medium vision tasks on edge devices where full GPU infrastructure isn’t available. Non-Latin script support is a clear weak point, and the flag misidentification is a notable failure. For developers evaluating tiny vision-language models for on-device or CPU-constrained deployments, this video provides concrete reproduction steps and realistic performance expectations rather than benchmark-sheet optimism.
📺 Source: Fahd Mirza · Published April 09, 2026
🏷️ Format: Hands On Build
