Google and Hugging Face release Gemma 4
google and hugging face released gemma 4, a family of frontier multimodal models available in four sizes ranging from 2b to 31b parameters. featuring apache 2.0 licensing, the models support text, image, and audio inputs with architectural upgrades like per-layer embeddings and shared kv cache. the 31b dense model achieves an estimated lmarena score of 1452, while the 26b mixture-of-experts variant reaches 1441 with only 4b active parameters.
Google and Hugging Face released Gemma 4, a suite of open-weight multimodal models capable of processing text, images, video, and audio. The family consists of four sizes (E2B, E4B, E9B, E31B), all released with Apache 2.0 licenses and optimized for on-device deployment and quantization. Architectural improvements include Per-Layer Embeddings (PLE), Shared KV Cache, and alternating local-global attention to support long context windows and efficient agentic workflows. Benchmarks indicate the 31B dense model achieves an estimated LMArena score of 1452, rivaling proprietary frontier models, while the 26B Mixture-of-Experts (MoE) model achieves a score of 1441 using only 4B active parameters. Integration is immediately available across major frameworks including Transformers, llama.cpp, MLX, and WebGPU.