Omar Sanseviero

Omar Sanseviero leads Developer Experience at Google DeepMind, working across product, engineering, research, and community to help developers build with AI Studio, the Gemini API, and Gemma. His responsibilities include influencing product roadmaps, leading education and documentation, and integrating models with developers’ tools. His Hugging Face profile uses the handle osanseviero.[2][3]

Sanseviero’s team participates from early technical designs and feature ideas through launches and continuing iteration. He leads developer experience and launches for models including Gemini 3, Gemma, and Nano Banana. He also participates in open-model efforts including MedGemma and EmbeddingGemma, explaining his personal commitment on his website: “Open ML has a big place in my heart”.[2]

With Ian Ballantyne, Sanseviero coauthored the June 26, 2025 Gemma 3n developer guide. They explained how the mobile-first model gives developers choices under hardware constraints: MatFormer nests smaller models within a larger one, training a 2B effective-parameter sub-model alongside the 4B effective-parameter model. Developers could use either released model or construct intermediate sizes through Mix-n-Match, adjusting feed-forward dimensions and selectively skipping layers.[4]

The guide also clarified the difference between total parameters and accelerator-memory requirements. Gemma 3n’s E2B and E4B models contain 5B and 8B parameters respectively; Per-Layer Embeddings allow some parameters to be loaded and computed on the CPU, leaving approximately 2B or 4B core transformer parameters in accelerator memory. Dynamic switching between E2B and E4B inference paths was presented as a future capability, rather than part of the launched implementations.[4]

The coauthors stated that “Making Gemma 3n accessible from day one has been a priority.” They credited partner and open-source teams for support across tools including Hugging Face Transformers, llama.cpp, MLX, Ollama, and vLLM, connecting the launch to the software developers already used.[4]

Before his Google DeepMind role, Sanseviero worked on model sharing, library integrations, and developer communities at Hugging Face. Earlier Google work included machine-learning models and infrastructure for Google Assistant and a 20% Product Manager role on TensorFlow Graphics.[2]

At Hugging Face, Sanseviero held the title Chief Llama Officer / Head of Platform and Community. He reports launching Spaces and scaling it from 500 to more than 400,000 community-built demos, as well as creating and growing the Developer Advocacy Engineering and Moonshot organizations to 25 people when he left. He built partnerships with labs and open-source communities and established integrations with libraries including spaCy and Sentence Transformers.[2]

A concrete example was the Sentence Transformers Hub integration, which Sanseviero described with Nils Reimers in their June 28, 2021 article. Sentence Transformers produces sentence, paragraph, and image embeddings for applications such as semantic search. The integration connected model sharing to automatically generated model cards describing architecture and usage, widgets for embeddings and sentence similarity, and programmatic inference endpoints. A shared model could therefore be documented, tried in the browser, and called through an API.[5][2]

Sanseviero also coauthored Mixture of Experts Explained, published December 11, 2023, with Lewis Tunstall, Philipp Schmid, Sourab Mangrulkar, Younes Belkada, and Pedro Cuenca. Their explanation separates active computation from total model memory: a router sends tokens to selected experts, reducing computation, while the described inference setup still requires all model parameters in memory. Their deployment guidance favors sparse experts for high-throughput settings with many machines and dense models for low-throughput settings with little VRAM.[6]