>
**tl;dr** \- Fine-tuned Qwen3 1.7B – called HyprLLM – which outperforms llama 3.2 3B in summarization for user experience because “vanilla” models suck at summarization.
**Context** \- I am building an [open-source](https://github.com/fastrepl/hyprnote) privacy-first AI notetaker for people in compliance-sensitive environments. It uses on-device AI models to process everything locally. Used to use llama 3.2 3B q8 which sucks at summarizing so had to post-train a new model.
**Selection** \- Juggled between Gemma and Qwen. But found Qwen to show more promising results.
**Preparing** \- Since I can’t get user data, I had to create a pipeline for synthetic data generation.
**Training** \- Just boring stuff. Used Modal.
Planning to fine-tune whisper as well. Also trying to create next version for HyprLLM for multi-lingual support; our user base is global.
Would love to get any tips on synthetic dataset generation or suggestions on models!
Trend Engine: AI-Powered News & Trends 