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

April 20, 202615:2625,494 views

Summary

Omar Sanseviero from Google DeepMind presents the Gemma 4 model family, released just seven days before the conference, marking his first public talk on the new release. The presentation traces the evolution from Gemma 1 through Gemma 3 to Gemma 4, emphasizing consistent capability improvements without increasing model sizes.

Gemma 4 spans a range from 2 billion to 32 billion parameters, designed explicitly for developer-friendly, on-device deployment. The smallest models (E2B and E4B architectures) can run on Android phones, iPhones, and even Raspberry Pi devices. Sanseviero demonstrates 10 parallel instances of Gemma running simultaneously on a laptop via llama.cpp, generating SVG files at roughly 100 tokens per second entirely offline. A live Android demo shows the model performing agentic tasks — including playing piano — fully on-device in airplane mode.

A key architectural innovation highlighted is the E2B (Effectively 2 Billion) design, which uses per-layer embeddings introduced in summer 2024. Rather than computing matrix multiplications for the embedding layer as in standard transformers, these embeddings act as a lookup table and can reside in CPU or disk memory rather than GPU VRAM. This means a 5-billion-parameter model effectively loads only 2 billion parameters into the GPU. This architecture is accessible via a single llama.cpp flag (--override-tensor), enabling practical mobile deployment.

On benchmarks, Gemma 4 models occupy the top-left quadrant of LM Arena score vs. parameter count charts, indicating strong performance relative to model size. The 31B dense model is positioned as the highest-intelligence variant, still runnable on a consumer GPU. The MoE 27B variant prioritizes low-latency inference.

Gemma 4 ships under the Apache 2.0 license — a deliberate change from the prior custom license that developers had criticized. Multimodal capabilities now cover images, video, and audio (including speech-to-translated-text across languages). The model was trained on over 140 languages using the same Gemini-derived tokenizer, which Sanseviero highlights as particularly valuable for low-resource and indigenous languages such as Quechua or regional Indian languages.

Within one week of launch, Gemma 4 base models accumulated 10 million downloads and spawned over 1,000 derivative models (quantizations and fine-tunes) on Hugging Face. The broader Gemma family has surpassed 500 million total downloads and over 100,000 community models. Integrations span Unsloth, MLX, llama.cpp, Hugging Face Transformers, vLLM, and C/Lang, as well as Android Studio's agent/offline coding mode.

Google has released two official Gemma variants: Shield Gemma (a family of reward/safety models for production content moderation) and MedGemma, a multimodal Gemma 3-based model covering radiology and chest X-ray interpretation. Community-driven sovereign AI efforts include AI Singapore building models for Southeast Asian languages and Sarvam building government-backed national models for India's official languages.

A notable real-world application cited is a December 2024 DeepMind paper in which Gemma 3 was used to propose cancer therapy pathways that were subsequently validated in a physical laboratory setting. Sanseviero closes by urging engineers to spend one to two hours over the next two weeks exploring the latest open models to understand where on-device AI capabilities now stand, predicting that in 6–12 months the ecosystem will offer highly capable, fully customizable open models running directly on personal devices.

Key Takeaways

•Gemma 4's E2B architecture uses per-layer embeddings as lookup tables, allowing a 5B-parameter model to occupy only 2B parameters in GPU VRAM — enabling practical mobile inference via a single llama.cpp --override-tensor flag.
•All Gemma 4 models now ship under Apache 2.0, removing the restrictive custom license that previously limited commercial and open-source adoption.
•The 2B and 4B Gemma 4 variants run fully on-device on Android, iOS, and Raspberry Pi — including multimodal and agentic workloads — with no API calls required.
•Running 10 parallel Gemma instances on a laptop via llama.cpp achieves approximately 100 tokens per second per instance, demonstrating viable local multi-agent pipelines.
•The Gemini-derived tokenizer, trained across 140+ languages, provides strong out-of-the-box multilingual performance even when fine-tuning on very low-resource or indigenous languages.
•Shield Gemma provides open, fine-tunable reward/safety classifiers for text and image content moderation in production deployments.
•MedGemma is a specialized open multimodal variant for radiology and clinical tasks, further fine-tunable for niche medical use cases.
•10 million downloads and 1,000+ derivative models were generated within one week of the Gemma 4 launch, indicating rapid community adoption and ecosystem momentum.
•Gemma 3 was used in published DeepMind research (Dec 2024) to propose cancer therapy pathways subsequently validated in a wet lab, illustrating that small open models can contribute to high-stakes scientific discovery.
•For maximum raw intelligence, developers should use Gemini or a frontier API model; Gemma's value proposition is privacy-preserving, customizable, offline-capable inference at consumer hardware scale.

Transcript

[music] >> All right. Hi everyone. It's cool here. So I'm super excited to give this talk because just 7 days ago we released Gemma 4. So before this conference, who here has heard about Gemma already? Okay, so most of you. Great. So Gemma is Google DeepMind's family of open models. Open models means that these are models that you can take, you can download, you can run in your own infrastructure, your own devices. You can fine-tune for your own use cases. So about a year ago we released Gemma 3...