Google Releases Gemma 4 Open Models Under Apache 2.0 — Frontier Agentic AI With 86% Tool-Use Score Now Fully Open

On April 2, 2026, Google DeepMind released Gemma 4, its most capable open-weight model family to date, built directly from the same research and technology underpinning the Gemini 3 frontier model. In a significant strategic shift, Google moved from its previous restrictive Gemma license to the fully permissive Apache 2.0 license, removing all commercial restrictions.

The Gemma 4 family consists of four models spanning from smartphones to workstations: Gemma 4 31B IT Thinking (flagship), Gemma 4 26B A4B IT Thinking (Mixture-of-Experts with 4B active parameters), Gemma 4 E4B IT, and Gemma 4 E2B IT (edge/mobile).

Benchmark results for the flagship 31B model are exceptional: 86.4% on tau2-bench (agentic tool use in retail scenarios), 89.2% on AIME 2026 (mathematics), 80.0% on LiveCodeBench v6 (competitive coding), 84.3% on GPQA Diamond (scientific knowledge), 85.2% on MMLU (multilingual QA), and 76.9% on MMMU Pro (multimodal reasoning).

The agentic capabilities represent the most dramatic improvement. Gemma 4 31B achieves 86.4% on tau2-bench compared to just 6.6% for Gemma 3 27B — a 13x improvement in autonomous tool use. Models support native function calling, enabling agents to plan, navigate applications, and complete multi-step tasks autonomously.

Google simultaneously announced Gemma 4 integration with Android AICore Developer Preview for on-device agentic AI on mobile. The E2B variant is specifically optimized for IoT and mobile deployment with minimal memory footprint.

The Apache 2.0 license change is the most strategically significant aspect. Previous Gemma models had commercial restrictions — Apache 2.0 enables unrestricted commercial deployment, fine-tuning, and redistribution, directly competing with Meta Llama and Mistral in the open ecosystem.

Developer response was immediate: the HackerNews post reached #1 with 722+ points and 204 comments within 3 hours. The model is available on HuggingFace and Kaggle.