Google has released Gemma 4, its latest open AI models, shifting to the Apache 2.0 license to address developer concerns about restrictive terms. The new suite includes four models optimized for local use: two large variants (26B Mixture of Experts and 31B Dense) designed for high-performance hardware like Nvidia's $20,000 H100 GPU, and two smaller models (Effective 2B and Effective 4B) built for mobile devices such as smartphones and Raspberry Pi. Google claims the 26B model activates only 3.8 billion parameters during inference, boosting speed, while the 31B model prioritizes quality for fine-tuning. The mobile-focused models promise near-zero latency and reduced battery drain, with Google collaborating with Qualcomm and MediaTek for optimization. The company positions Gemma 4 as more capable than its predecessor, asserting that even the largest variant is cheaper to run than competitors like GLM-5 and Kimi 2.5.
When Google says Gemma 4's 26B model "activates only 3.8 billion parameters in inference mode," that means it's not just faster—it's smarter about resource use. That matters because Nigerian developers working with limited budgets can now run high-quality AI locally without relying on cloud services, cutting costs and dependency on foreign tech giants. The shift to Apache 2.0 also signals Google's response to developer frustration, but whether this openness extends beyond licensing to real collaboration remains to be seen.
