Particle.news
Get it on Google Play
Download on the App Store
33 ARTICLES
·
last mo.

Google Releases Gemma 4 as Fully Open Models Under Apache 2.0

The switch to Apache 2.0 enables unrestricted commercial use with local deployment from phones to data‑centre GPUs.

Image
Image
Google unveils gemma 4 as its most advanced open ai model for reasoning and agentic tasks
Image

Overview

  • Google DeepMind released Gemma 4 on Thursday under the Apache 2.0 license, opening the models for free use, modification, and redistribution.
  • The family ships in four sizes—E2B, E4B, 26B Mixture‑of‑Experts, and 31B Dense—spanning phones, edge boards, laptops, and workstation‑class GPUs.
  • Google says the E2B and E4B edge models run fully offline with near‑zero latency on Android phones, Raspberry Pi, and Jetson Nano following work with the Pixel team, Qualcomm, and MediaTek.
  • All models support images and video, the edge variants take audio input, agent features enable function calling and structured JSON, context windows reach 128K on edge and up to 256K on larger models, and training covers 140+ languages.
  • Google reports strong benchmark showings with the 31B at No. 3 and the 26B at No. 6 on Arena AI, and the weights are available now via Google AI Studio, AI Edge Gallery, Hugging Face, Kaggle, and Ollama.