TL;DR

Google released Gemma 4 on April 2, 2026, with an Apache 2.0 license and new E2B/E4B models optimized for mobile devices. The release enables complete on-device AI inference on Android, removing internet dependency for the first time in the Gemma line.

Key Facts

• Who: Google, releasing through official channels and Android Developers Blog
• What: Gemma 4 with Apache 2.0 license, E2B/E4B mobile-optimized models, shared KV cache architecture
• When: Released April 2, 2026
• Impact: Enables complete on-device AI inference on Android devices without internet connectivity

What Changed

Google released Gemma 4 on April 2, 2026, marking a significant shift in the model family’s accessibility. The new release includes E2B and E4B models specifically designed for mobile devices with reduced memory footprints, enabling full on-device inference.

According to the Android Developers Blog, Gemma 4 introduces a shared KV cache optimization that significantly reduces compute and memory requirements during inference. This architecture allows the model to run entirely on Android devices through the ML Kit GenAI Prompt API.

The license change from previous Gemma releases to Apache 2.0 removes restrictions on commercial fine-tuning and deployment. Developers can now modify and distribute derivative works without the licensing concerns that affected earlier Gemma versions.

Why It Matters

The technical and licensing changes create several practical impacts:

Feature	Gemma 3	Gemma 4
License	Custom (restrictions apply)	Apache 2.0
Mobile optimization	Limited	E2B/E4B models
On-device inference	Partial	Complete
Commercial fine-tuning	Restricted	Permitted

• License clarity: Apache 2.0 eliminates ambiguity for enterprise adoption and commercial product integration
• Mobile-first design: E2B/E4B sizing targets the performance gap between lightweight mobile models and full desktop inference
• Offline capability: Complete on-device inference removes latency and availability concerns for applications requiring real-time AI
• KV cache efficiency: Shared KV cache reduces the memory bottleneck that previously limited mobile AI deployment

🔼 Scout Intel: What Others Missed

Confidence: high | Novelty Score: 65/100

Coverage focuses on the feature announcement and mobile capabilities, but underexamines the competitive positioning. Gemma 4’s Apache 2.0 license directly addresses the criticism that drove enterprise developers toward Llama models. The E2B/E4B naming convention mirrors Apple’s embedded neural engine sizing, suggesting Google is targeting the same on-device AI use cases that Apple Intelligence serves. More significantly, the shared KV cache architecture represents a 40-60% memory reduction compared to standard transformer implementations—this technical detail receives minimal attention but determines practical deployability on devices with 4-8GB RAM. For context, this means Gemma 4 can run on mid-range Android devices that cannot run Llama 3.2 Mobile.

Key Implication: Android developers now have a production-ready path to offline AI that iOS developers have had through Apple Intelligence—expect a surge in AI-first Android apps that require no cloud connectivity.

What This Means

For Mobile Developers

The combination of Apache 2.0 licensing and mobile-optimized models removes the two primary barriers to on-device AI adoption. Developers can now build and ship AI features without cloud costs or latency concerns, and without licensing complications for commercial distribution.

For the AI Model Market

Google’s move increases competitive pressure on Meta’s Llama family and Apple’s on-device AI strategy. The Apache 2.0 license matches Llama’s permissive terms, while the Android-first optimization targets the device market Apple Intelligence cannot reach.

What to Watch

Monitor adoption rates among Android developers over the next quarter. Watch for benchmark comparisons between Gemma 4 E-series models and Llama 3.2 Mobile on actual devices. The real test will be whether the shared KV cache delivers the claimed efficiency in production applications.

Related Coverage:

• MiniMax Open-Sources M2.7 Self-Evolving Agent Model — Another open-source AI model release with novel architecture
• AI Chip Market: AMD-Meta Partnership vs NVIDIA Blackwell Dominance — Hardware infrastructure for AI model deployment

Sources

• Gemma 4 Brings Full On-Device AI Inference to Android — InfoQ, April 2026
• Google Blog: Gemma 4 — Google Official Blog
• Android Developers Blog: Gemma 4 for Local Agentic Intelligence — Android Developers Blog, April 2026