Google DeepMind Releases Gemma 4, Most Capable Open Models to Date Under Apache 2.0

TLDR

Google DeepMind launched Gemma 4 on April 2, 2026, describing it as the most capable open models in the Gemma series to date. The family brings frontier-level reasoning and agentic capabilities to open weights under an Apache 2.0 license for the first time in the line. Models span edge devices (phones, Raspberry Pi-class) to larger server deployments, support multimodal and multilingual use (140+ languages), and target multi-step planning, autonomous action, code generation, and audio-visual tasks without heavy fine-tuning.

Gemma 4 Family

The release includes variants optimized across hardware tiers:

• Smaller efficient models (E2B, E4B) for on-device and edge agentic use.
• Larger models (including 26B MoE and 31B-class) for higher capability.

Key characteristics highlighted:

• Built from the same research lineage as Gemini models.
• Strong performance on reasoning, agentic benchmarks, and Arena-style evaluations (high scores reported as of early April 2026).
• Multimodal capabilities and long context suitable for complex workflows.
• Apache 2.0 license removes prior custom restrictions, enabling broad modification, commercial use, and redistribution.

Google and partners (including Hugging Face day-one support) emphasized accessibility for developers building agents and autonomous systems locally or at scale.

Agentic and Edge Focus

A companion post from the Google AI Edge team positioned Gemma 4 for bringing state-of-the-art agentic skills to the edge. Use cases include multi-step planning, offline code generation, and audio-visual processing directly on hardware without constant cloud round-trips. The models are presented as a practical bridge between research advances and real deployment constraints for phones, laptops, and embedded devices.

Why this story matters

Open models have been a key vector for distributing AI capability beyond the largest labs. Gemma 4's combination of claimed capability leap, broad licensing, and explicit agentic/edge orientation accelerates experimentation with autonomous systems outside hyperscaler clouds. The move to clean Apache 2.0 terms also addresses developer feedback on previous custom licenses. As agentic AI becomes a central focus, open high-quality bases at multiple scales provide measurable infrastructure for research, startups, and on-device innovation.

Sources

• Google DeepMind / The Keyword: “Gemma 4: Byte for byte, the most capable open models” (April 2, 2026)
• Google AI Edge: “Bring state-of-the-art agentic skills to the edge with Gemma 4” (April 2, 2026).
• Model cards and Hugging Face availability notes from release day.
• Cross-referenced coverage confirming Apache 2.0 license and performance positioning.