Gemma, DeepMind's Family of Open Models — Omar Sanseviero, Google DeepMind

Omar Sanseviero (Google DeepMind) presents Gemma 4, the most capable open model family Google has released, switching to Apache 2 and hitting 10M downloads in one day.

• Gemma 4 reached 10 million downloads within one day of release; over 1,000 community fine-tunes appeared immediately.
• New E2B architecture (per-layer embeddings) lets a 5B-parameter model load only 2B params into GPU; remainder runs on CPU or disk.
• Gemma 4 switched from Google’s custom license to Apache 2, addressing the top community complaint about previous versions.
• Smallest Gemma 4 models run fully offline on Android, iPhone, Raspberry Pi, and even a Nintendo Switch via llama.cpp.
• DeepMind researchers used Gemma 3 to propose cancer therapy pathways that were validated in an actual lab.
• MedGemma (multimodal, Gemma 3-based) handles radiology and chest X-ray understanding as a freely downloadable open model.
• Gemma 4 trained on 140+ languages using Gemini’s tokenizer; enables fine-tuning on low-resource languages like Quechua out of the box.
• Sarvam and AI Singapore are building government-backed sovereign AI models for India and Southeast Asia on top of Gemma.

2026-04-20 · Watch on YouTube