Welcome Gemma 4: Frontier multimodal intelligence on device

Summary

Google DeepMind’s Gemma 4 family is available on Hugging Face as Apache 2-licensed, multimodal models intended for use across agents, inference engines, fine-tuning libraries, and on-device deployments. The family accepts image and text inputs across five sizes, while E2B, E4B, and 12B Unified also support audio; context windows range from 128k to 256k tokens. Its design combines alternating local and global attention, dual RoPE configurations, Per-Layer Embeddings, shared KV caches, and model-specific vision and audio handling, including an encoder-free unified 12B variant. The post documents integrations for Transformers, llama.cpp, MLX, WebGPU, Rust, and other tools, alongside fine-tuning examples and demos. Reported instruction-tuned results include a 1452 estimated text-only LMArena score for 31B and benchmark results spanning reasoning, coding, vision, audio, and long-context tasks.

Takeaways

1. 1

   Gemma 4 has five model sizes: E2B, E4B, 12B Unified, 31B, and 26B A4B MoE, with 128k or 256k context windows.

2. 2

   Shared KV caches let selected final layers reuse key-value states from earlier layers, reducing inference memory and compute with minimal reported quality impact.

3. 3

   The 12B Unified model projects raw image patches and audio waveforms directly into the LLM embedding space, avoiding separate vision and audio encoders and enabling one-pass fine-tuning.