Bring state-of-the-art agentic skills to the edge with Gemma 4

Google DeepMind launched Gemma 4, a family of open models under Apache 2.0 license designed for on-device agentic AI experiences across a wide range of hardware.

• •Gemma 4 supports multi-step planning, autonomous action, offline code generation, and audio-visual processing without specialized fine-tuning, with support for 140+ languages.
• •Agent Skills, a new feature in Google AI Edge Gallery (iOS/Android), enables fully on-device multi-step agentic workflows such as querying Wikipedia, generating visualizations, and integrating with other models.
• •LiteRT-LM runs Gemma 4 E2B in under 1.5GB memory using 2-bit/4-bit weights, supports constrained decoding, dynamic context up to 128K tokens, and processes 4,000 tokens in under 3 seconds.
• •Performance benchmarks include 133 prefill/7.6 decode tokens/s on Raspberry Pi 5 CPU, and 3,700 prefill/31 decode tokens/s on Qualcomm Dragonwing IQ8 NPU.