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

Key Concepts

• Gemma 4: Google DeepMind’s latest family of open-weight models, ranging from 2B to 31B parameters.
• Open Models: Models that can be downloaded, run on local infrastructure/devices, and fine-tuned for specific use cases.
• E2B/E4B Architecture: A novel "per-layer embedding" architecture designed for high-speed, on-device performance by offloading memory-intensive tasks.
• Agentic AI: Systems capable of autonomous decision-making and task execution (e.g., coding, controlling device functions).
• Multimodality: The ability of models to process and understand text, images, video, and audio.
• Apache 2.0 License: The new, permissive licensing model adopted for Gemma 4 to ensure developer flexibility.

---

1. Overview of Gemma 4

Gemma 4 is Google DeepMind’s most capable open-model family to date. It spans a parameter range of 2 billion (2B) to 31 billion (31B). The models are designed to be developer-friendly, fitting into various hardware environments from mobile devices (Android/iOS) and Raspberry Pis to consumer-grade GPUs.

• Performance: The 31B model offers the highest raw intelligence, while the smaller models (2B/4B) are optimized for low-latency, on-device agentic tasks.
• Licensing: Responding to community feedback, Gemma 4 is released under the Apache 2.0 license, providing users with greater control and commercial flexibility.

2. Technical Innovations: The "E" Architecture

A significant technical breakthrough in Gemma 4 is the E2B (Effectively 2 Billion) and E4B architecture.

• Per-Layer Embeddings: Instead of traditional dense matrix multiplications for every layer, this architecture uses a lookup-table approach.
• Memory Optimization: By offloading these embeddings to the CPU or disk, the model requires significantly less VRAM on the GPU. This allows a 5B parameter model to function as if it were a 2B model in terms of GPU memory footprint, enabling high-performance execution on mobile hardware.
• Implementation: This can be leveraged via llama.cpp using the override_tensor flag.

3. Multimodal and Multilingual Capabilities

• Multimodality: Gemma 4 supports image, video, and audio processing. It can perform tasks like object detection, fine-grained video analysis, and speech-to-text translation across different languages.
• Multilingualism: Trained on over 140 languages, the model utilizes a tokenizer derived from the Gemini architecture. This design ensures high performance even for low-resource languages (e.g., Quechua or regional Indian languages) without requiring extensive retraining.

4. Ecosystem and Community Integration

Google emphasizes collaboration with the open-source community to ensure Gemma 4 is accessible via standard tools:

• Tooling Support: Native support for llama.cpp, Unsloth, MLX, Hugging Face Transformers, vLLM, and C Lang.
• Adoption: Within one week of release, the base models reached 10 million downloads, with over 1,000 community-derived variants (fine-tunes/quantizations) already available.
• Product Integration: Android Studio now features an offline mode powered by Gemma, allowing developers to use the model for code generation directly within their IDE.

5. Real-World Applications and Case Studies

• Agentic Workflows: Demonstrations showed Gemma running locally on phones to perform tasks like playing the piano or generating SVG code without API calls.
• Medical Research: Researchers used previous Gemma iterations to propose cancer therapy pathways, which were subsequently validated in laboratory settings.
• Sovereign AI: Organizations like AI Singapore and initiatives in India (e.g., Sarvam) are using these models to build national-level AI solutions for local languages and sovereign data requirements.
• Safety: Shield Gemma was introduced as a specialized model family for production environments, designed to filter toxic content and ensure compliance with safety policies.

6. Notable Quotes

• "Gemma is not just about a model that you can use, but it's more about enabling the ecosystem to build on top of it."
• "I do think we'll have extremely capable models running directly in our own devices, in our own pockets."

7. Synthesis and Conclusion

Gemma 4 represents a shift toward "on-device intelligence," where the focus is not merely on parameter count, but on architectural efficiency and accessibility. By combining a permissive license, a novel per-layer embedding architecture, and deep integration with the open-source ecosystem, Google is enabling developers to run sophisticated, agentic, and multimodal AI entirely offline. The trajectory suggests that in the coming year, highly customized, private, and capable AI will become a standard feature of personal computing and mobile devices.