Small but Mighty Gemma4

Google DeepMind's Gemma 4 has been released on Hugging Face.

• It comes in 4 sizes: E2B (2.3B effective), E4B (4.5B effective), 31B dense, and 26B A4B MoE (4B active / 26B total).
• All sizes have base and instruction-tuned (IT) checkpoints, with context lengths of 128K or 256K.
• It supports image, text, and audio inputs, with E2B/E4B also handling audio.
• The image encoder supports variable aspect ratios and token budget control, widening the options among speed, memory, and quality.
• The architecture centers on local sliding-window + global attention, dual RoPE, PLE (Per-Layer Embeddings), and shared KV cache.
• Vision can be encoded with various token budgets (70-1120), and audio is based on a USM-style conformer.
• Based on private preliminary checkpoints, it states that the 31B dense scored an estimated LMArena 1452 and the 26B MoE scored 1441 on text.

Deployment coverage is also broad. It can be integrated right away with transformers, llama.cpp, MLX, WebGPU, Rust, transformers.js, Mistral.rs, and more, with a design built with local agents, on-device inference, and quantization use in mind.

The multimodal demos show examples spanning OCR, speech recognition, object detection, pointing, GUI detection, function calling, code generation/editing, and video understanding. Use cases such as directly outputting bounding boxes from images in JSON format are also presented.