Multilingual powerhouse — testing for mobile deployment
This model covers German, Urdu, Arabic, Polish, Turkish — exactly the kind of multilingual capability we need for global mobile AI.
At Dispatch AI (FZE, UAE), we're building mobile AI that works for everyone. Models like this are the foundation.
We benchmark multilingual models on our 40-phone farm (Snapdragon 865) to see which maintain quality across languages when quantized to 4-bit. Results vary wildly — some lose 30% quality in non-English after quantization.
Would love to see multilingual eval at different quantization levels.
- Dispatch AI (FZE), Sharjah UAE
Thanks for your interest!
Just to clarify, this model is fine-tuned on English data and is based on PEGASUS-large (huggingface.co/google/pegasus-large), which was pre-trained on large-scale English text corpora (C4 and HugeNews). As a result, it is not intended for multilingual use cases.
For a quantized version of this model, we have published a hands-on notebook here: https://github.com/medoidai/model-quantization-blog-notebooks/blob/main/notebooks/A_Hands_On_Walkthrough_on_Model_Quantization.ipynb
More details about our quantization approach in this blog post: https://www.medoid.ai/blog/a-hands-on-walkthrough-on-model-quantization/