Google DeepMind released Gemma 4 on April 2, 2026, comprising four vision-capable reasoning models under a fully permissive Apache 2.0 license. The release includes 2B, 4B, and 31B parameter models plus a 26B-A4B mixture-of-experts variant, marking a significant shift in Google's open model licensing strategy.
Model Capabilities and Architecture
Gemma 4 models deliver frontier-level intelligence per parameter, built from the same research and technology powering Gemini 3. The models combine four major capability domains: advanced reasoning and math, professional-grade code generation, long-document intelligence, and autonomous agentic workflows.
All sizes support multimodal inputs (text and images), with edge models adding audio capability. The models feature context windows up to 256,000 tokens and are optimized for deployment across devices ranging from smartphones and Raspberry Pi to high-end GPUs.
Thinking Variants and Performance
Gemma 4 includes "thinking variants" trained to reason step-by-step before producing final answers, trading minimal latency for substantially better performance on complex tasks. This approach mirrors chain-of-thought reasoning methodologies.
The 31B model leads open models on multiple benchmarks:
- MMLU-Pro: 85.2%
- MMMU-Pro: 76.9%
- LiveCodeBench: 80%
- AIME 2026: 89.2%
- Agentic τ2-bench: 86.4%
These scores position Gemma 4 31B as the third-ranked open model globally on the Arena AI text leaderboard.
Apache 2.0 Licensing Breakthrough
Unlike earlier Gemma versions released under custom licenses, Gemma 4 ships under the fully permissive Apache 2.0 license. This enables unrestricted commercial use, modification, and redistribution—representing a major shift in Google's open model strategy and competitive positioning against other frontier labs.
Strategic Context
The release comes as executives at Google, OpenAI, and Anthropic increasingly describe the frontier AI race as "effectively neck-and-neck, with companies making different tradeoffs around cost, speed and computing resources." Google's strategy includes both proprietary models like Gemini 3 and open releases like Gemma 4.
Additionally, Google partnered Boston Dynamics with Google Cloud and DeepMind to integrate Gemini Robotics-ER 1.6 into Spot robots and Orbit AI visual inspection platforms, demonstrating broader applications of the underlying technology.
Key Takeaways
- Google DeepMind released Gemma 4 on April 2, 2026 with four model sizes (2B, 4B, 31B, and 26B-A4B MoE) under Apache 2.0 license
- Models support up to 256K token context windows and multimodal inputs (text, images, and audio on edge models)
- The 31B model ranks third globally among open models on Arena AI text leaderboard with 85.2% on MMLU-Pro
- Apache 2.0 licensing enables unrestricted commercial use, marking a strategic shift from Google's previous custom licenses
- Thinking variants trade minimal latency for improved performance on reasoning, math, and coding benchmarks