Gemma 4: Open Models for Every Device

▶ Watch (15:33)

Seven days before the conference, Google DeepMind released Gemma 4. The family spans 2B to 32B parameters. The smallest two models run on iPhones, Android phones, and even a Raspberry Pi. A 27B mixture-of-experts model delivers low latency. The 31B dense model provides the most raw intelligence. All fit in a single consumer GPU. Gemma 4 uses an Apache 2 license, replacing the more restrictive license of earlier versions.

Per-Layer Embeddings for On-Device Speed

▶ Watch (20:48)

The E2B and E4B models use a novel architecture called per-layer embeddings. Instead of storing all parameters in GPU memory, the embedding tables live on CPU or disk. An E2B model has 4B total parameters but only 2B need GPU compute. This design is optimized for mobile use cases. Llama.cpp supports it with a simple flag override. The model runs entirely offline with no API calls.

Multimodal and Multilingual by Design

▶ Watch (22:11)

Gemma 4 handles images, video, and audio. It performs speech recognition and speech-to-text translation. The model was trained on over 140 languages. The tokenizer, shared with Gemini, was designed for multilingual use cases. This makes fine-tuning for low-resource languages like Quechua or Indian regional languages effective out of the box.

Community Adoption and Ecosystem

▶ Watch (23:58)

In its first week, Gemma 4 reached 10 million downloads. The community created over 1,000 fine-tuned or quantized variants. Total Gemma family downloads exceed 500 million. Google collaborated with Ollama, MLX, llama.cpp, Hugging Face, and Silang to ensure compatibility. Android Studio now includes an offline agent mode powered by Gemma for Android development.

Notable Quotes

Gemma 4 is the family of most capable of open models that Google has released ever. Cassidy Hardin · ▶ Watch (16:39)

if I think where we’ll stand in a year from now or in two years from now I do think we’ll have extremely capable models running directly in our own devices in our own pockets. Cassidy Hardin · ▶ Watch (19:44)

with the M4 we change our license to an actual Apache 2 license Cassidy Hardin · ▶ Watch (20:27)

Key Takeaways

  • Gemma 4 runs on devices from Raspberry Pi to consumer GPUs.
  • Per-layer embeddings cut GPU memory requirements by half.
  • The model supports 140+ languages and multimodal inputs.
  • Apache 2 license enables broad commercial use.