Budapest Data Ml Forum 2025

On June 18, I gave an online talk at the Budapest Data+ML Forum. The event is a solid, mid-sized conference that draws practitioners from Central Europe; it is a good place to exchange technical expertise with other engineers. My session focused on Phi Silica, a small language model designed to run entirely on Copilot+ PC laptops, and on Rewrite, a publicly available paraphrasing tool built on top of the small language model.

I covered a diverse set of approaches our team has applied for the model’s optimization, including:

• QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs. One of the best existing approaches toward outlier removal and subsequent quantization.
• EAGLE: Speculative Sampling Requires Rethinking Feature Uncertainty. Speculative decoding that gave a throughput boost during Phi Silica generation.

More importantly, I focused on my personal contributions as well:

• Rewrite: a publicly available paraphrasing tool shipped with multiple tones.
• Soft prompts and LoRA adapters for quality improvements in Rewrite.

I’m pleased with the steady progress toward practical, small language models. If you’re experimenting with on-device LMs on Copilot+ PCs, I’d love to compare benchmarks – feel free to reach out.