Google DeepMind released Gemma 4 on February 26, 2026, under the Apache 2.0 license, marking a significant advancement in on-device agentic AI capabilities. The open-source model family runs completely offline on devices ranging from smartphones to Raspberry Pi units, enabling autonomous multi-step planning and tool calling without cloud infrastructure.
Four Model Variants Optimize for Different Use Cases
Gemma 4 includes four distinct models: E2B (2B parameters) and E4B (4B parameters) for edge devices, a 26B Mixture of Experts model, and a 31B dense model. The edge models activate an effective 2B and 4B parameter footprint during inference to preserve RAM and battery life. The E2B model can run in under 1.5GB memory on some devices using 2-bit and 4-bit weights with memory-mapped per-layer embeddings.
Native Tool Calling Built Into On-Device Runtime
All Gemma 4 models feature native tool calling built into LiteRT-LM, Google's on-device runtime. This enables autonomous action and multi-step planning without requiring specialized fine-tuning. The models support offline code generation, audio-visual processing, and handle 140+ languages. Context windows reach 128K tokens for edge models and up to 256K for larger variants.
Performance Competes With Much Larger Cloud Models
The 31B model ranks #3 among open models globally on the Arena AI text leaderboard, while the 26B model holds #6 position. Google claims these models "outcompete models 20x their size." On Apple's M4 Max chip, the Gemma 4 26B model runs at approximately 50 tokens per second locally.
Multimodal Capabilities Run Fully On-Device
All Gemma 4 models handle text and image input natively and can analyze video as frame sequences. The E2B and E4B edge models additionally support audio input. This multimodal processing happens entirely on-device without cloud connectivity, enabling near-zero latency applications.
Developer Ecosystem Rapidly Adopting Edge-First Architecture
Google released an AI Edge Gallery app for iOS and Android to showcase Gemma 4 capabilities. Early developer projects include Parlor (1,286 GitHub stars) for on-device voice assistants, PokeClaw (409 stars) for autonomous phone control, and Gemma Gem (613 stars) for browser-based AI via WebGPU. Developers have also deployed Gemma 4 on NVIDIA Jetson platforms for robotics applications.
Benchmark testing showed Gemma 4 E2B achieving 94% accuracy (31/33 queries) on on-device legal RAG tasks across 7 domains with mixed English and Hinglish input, compared to 64% (21/33) for Llama 3.2 1B. The model handled code-mixed queries without specialized training.
Key Takeaways
- Google DeepMind released Gemma 4 on February 26, 2026, under Apache 2.0 license with four model variants (E2B, E4B, 26B MoE, 31B dense) optimized for on-device deployment
- The 31B model ranks #3 globally among open models on Arena AI leaderboard, while 26B ranks #6, with performance that "outcompetes models 20x its size"
- Native tool calling built into LiteRT-LM runtime enables autonomous multi-step planning and action execution completely offline on devices like smartphones, Raspberry Pi, and NVIDIA Jetson
- E2B edge model runs in under 1.5GB memory and achieved 94% accuracy on on-device legal RAG benchmarks, outperforming Llama 3.2 1B by 30 percentage points
- All models support multimodal input (text, images, video-as-frames) with E2B/E4B adding audio support, plus 140+ languages and context windows up to 256K tokens