Welcome Gemma 4: Frontier multimodal intelligence on device

2026-04-02 · Source: Hugging Face - Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Advanced, extended

Summary

Google DeepMind has released the Gemma 4 family of multimodal models on Hugging Face, offering Apache 2 licensed, high-quality models with pareto frontier arena scores. These models support image, text, and audio inputs, generating text responses, and are designed for deployment across various environments, including on-device. Gemma 4 comes in four sizes: E2B (2.3B effective parameters), E4B (4.5B effective), 31B (dense), and 26B A4B (mixture-of-experts with 4B active parameters), all available as base and instruction-tuned versions. Key architectural features include Per-Layer Embeddings (PLE) for smaller models and a Shared KV Cache for efficiency, enabling long context windows up to 256K tokens. The models demonstrate strong performance in benchmarks for reasoning, coding, vision, and long-context tasks, with the 31B dense model achieving an estimated LMArena score of 1452.

Key takeaway

For AI Engineers and ML Engineers seeking versatile, efficient multimodal models, Gemma 4 presents a compelling option due to its open licensing, on-device capability, and broad integration with popular tools like Hugging Face Transformers, llama.cpp, and MLX. You should explore its various sizes and fine-tuning options to match your specific application's performance and resource constraints, especially for agentic or edge deployments requiring robust multimodal understanding and function calling.

Key insights

Gemma 4 offers open, multimodal AI models optimized for on-device deployment and broad ecosystem integration.

Principles

• Efficiency through architectural innovation
• Broad compatibility across inference engines
• Multimodal input for diverse applications

Method

Gemma 4 models integrate Per-Layer Embeddings and a Shared KV Cache to enhance efficiency and long-context handling, supporting multimodal inputs (image, text, audio) and generating text outputs.

In practice

• Deploy on-device using llama.cpp or MLX
• Fine-tune with TRL or Unsloth Studio
• Integrate with local agents like Hermes or Open Code

Topics

• Gemma 4 Multimodal Models
• On-Device AI
• Per-Layer Embeddings
• Shared KV Cache
• Open-Source AI

Code references

• huggingface/blog
• huggingface/huggingface-gemma-recipes
• Blaizzy/mlx-vlm
• EricLBuehler/mistral.rs
• huggingface/trl

Best for: AI Engineer, Machine Learning Engineer, AI Scientist

Related on AIssential

• A Visual Guide to Gemma 4 — Gemma 4 models offer multimodal capabilities and efficiency through architectural innovations like interleaved attention and specialized encoders.
• [AINews] Gemma 4: The best small Multimodal Open Models, dramatically better than Gemma 3 in every way — Gemma 4 advances open-source AI with multimodal capabilities, efficient architecture, and a permissive Apache 2.0 license.
• Introducing Gemma 4 12B: a unified, encoder-free multimodal model — Gemma 4 12B unifies multimodal processing with an encoder-free architecture for efficient local agentic AI.
• Google Gemma 4 is here! #ai #gemini #google — Gemma 4 is a multimodal open model optimized for local execution on mobile devices, offering efficient on-device AI capabilities.
• What’s new in Gemma 4? — Gemma 4 offers a family of open-source, locally runnable models for diverse hardware, from mobile to desktop.
• The Sequence AI of the Week #839: Gemma 4 and the Compression of Intelligence — Gemma 4 transforms frontier AI capabilities into practical, deployable infrastructure for diverse applications.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Hugging Face - Blog.