Omar Sanseviero leads developer experience at Google DeepMind. He previously led platform and community work at Hugging Face, and writes tutorials explaining how to choose and run machine-learning models.[1]
Start with LLM Evals and Benchmarking →
X 
GitHub Career
Sanseviero describes his work as connecting engineering, product, research and community. At Hugging Face he led platform and community teams and launched Spaces, where developers publish interactive machine-learning demos. He also worked on integrations with libraries such as spaCy and Sentence Transformers. At Google DeepMind his developer-experience remit covers AI Studio, the Gemini API and Gemma. These are team and platform contributions, not claims that he authored the underlying models.[1]
Contributions
His quantization introduction explains that a model’s parameters are numbers stored at a chosen precision. Using fewer bits per number can reduce the memory required to load them. The guide works through estimates and warns that loading weights is only part of runtime memory use. It also separates smaller storage from faster execution: converting compressed values during computation can add overhead. This gives readers a useful first check before choosing a local model or buying hardware.[3]
Ideas
His evaluation guide distinguishes a pretrained base model from a model adapted for conversation. A high score on a test of knowledge does not necessarily predict helpful chat responses. He walks through benchmark design, implementation differences and the limits of using another language model as a judge. That turns model selection into a question about the task, hardware and evaluation method rather than the highest number on one chart.[2]
Tools, writing & talks
Writing, research & talks
Start here
LLM Evals and Benchmarkingosanseviero.github.ioExplains base versus chat models, common benchmark tasks and why evaluation methods can produce misleading comparisons.
Start with the evaluation guide before comparing model rankings. It introduces the main terms without requiring training experience; read the quantization guide next if you plan to run a model locally.
Author[2]A DeepMind engineer introduces Gemma 4 and its per-layer embedding design, where lookup tables can be offloaded to CPU or disk so a larger model runs with a small GPU footprint.
Catalog workOriginal source- A minimal Introduction to Quantizationosanseviero.github.io
Uses simple memory estimates to explain lower-precision model weights and their tradeoffs.
Author[3]