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

by Omar Sanseviero

Day 3 · Keynote · 15 min · 3 min read

Foundation Models Applied AI

TL;DR

Omar Sanseviero presented Gemma 4, Google DeepMind's open model family released seven days before the conference. The lineup runs from 2B to 32B parameters, all fitting on consumer hardware, now under an Apache 2 license. A novel per-layer embeddings architecture lets the E2B model offload 3 billion of its 5 billion parameters to CPU or disk, cutting GPU memory pressure in half. One week post-launch: 10 million downloads, over 1,000 community fine-tunes, and a cancer-therapy proposal from a Gemma 3-based model validated in a physical lab.

What Gemma 4 Ships

▶ Watch (01:26)

Gemma 4 runs from 2 billion to 32 billion parameters. The two smallest models fit on an Android phone, an iPhone, or a Raspberry Pi. The 27B mixture-of-experts model targets low-latency inference. The 31B dense model is for maximum capability, and still fits on a consumer GPU. All four are now under an Apache 2 license, replacing the more restricted terms that Gemma 3 shipped with. Sanseviero showed 10 Gemma agents generating SVGs in parallel on a laptop, each hitting 100 tokens per second with llama.cpp, no network connection.

Per-Layer Embeddings: The Architecture Behind the E Models

▶ Watch (05:28)

The “E” in E2B and E4B stands for “effectively.” A Gemma E2B model has 4 billion total parameters, but per-layer embeddings are stored as a lookup table rather than computed through matrix multiplication. That lets the embeddings sit on CPU or disk instead of GPU memory. The GPU loads only the 2 billion active parameters. In llama.cpp, one flag override moves the embeddings off the GPU. The result is a 5B-class model that behaves like a 2B model in terms of memory pressure, which is what makes it viable on a phone.

Multimodal and Multilingual Capabilities

▶ Watch (06:55)

The smallest Gemma 4 models handle images, video, and audio, including speech recognition and speech-to-translated-text (speak Spanish, get French text). The 31B handles fine-grained visual detail: object detection, bounding-box localization, and reading Japanese text inside an image. Gemma 4 was trained on more than 140 languages and shares the Gemini tokenizer, which was built with multilingual use cases in mind. That tokenizer choice matters for fine-tuning: even a low-resource language like Quechua or a regional Indian language tends to work without extra tokenizer surgery.

Community Adoption and Specialist Fine-Tunes

▶ Watch (08:42)

One week after release, Gemma 4 hit 10 million downloads. The broader Gemma family has crossed 500 million downloads and spawned over 100,000 community models. Google ships official variants: Shield Gemma for content moderation, Med-Gemini (Gemma 3-based) for radiology and chest X-ray analysis. Community groups go further: AI Singapore builds Southeast Asian language models on Gemma; Sarvam trains India’s government-backed national models on it. A December paper found that a Gemma 3-based model proposed cancer therapy pathways, then validated those pathways in a physical lab.

Notable Quotes

license to an actual Apache 2 license Omar Sanseviero · ▶ 5:16

that Gemma is not just about always some model Omar Sanseviero · ▶ 9:03

open models now can do very cool, Omar Sanseviero · ▶ 14:10

Key Takeaways

Gemma 4’s E2B/E4B models load only 2B parameters into the GPU even when the full model has 5B parameters.
10 million downloads in one week; over 100,000 community fine-tunes span medicine, content moderation, and sovereign AI.
The Apache 2 license replaces the previous restricted license, removing a barrier cited by prior Gemma users.

Filed under

Foundation Models Applied AI