Google released Gemma 4 with four model variants and switched to Apache 2.0 licensing, abandoning the restrictive custom Gemma license that frustrated developers for over a year. The lineup includes 26B Mixture of Experts, 31B Dense, and two mobile-optimized models (E2B and E4B) designed for smartphones and edge devices. Google claims the 31B variant will rank third on Arena's open model leaderboard, though it's significantly smaller than the top contenders GLM-5 and Kimi 2.5.
The licensing change matters more than the models themselves. As I've covered before, Google's custom Gemma license created legal uncertainty that kept many developers away from otherwise capable models. Apache 2.0 removes that friction, potentially making Gemma 4 the go-to choice for commercial applications that need truly open licensing. The timing aligns with growing enterprise demand for models they can modify and redistribute without licensing headaches.
Google's "local" marketing remains misleading despite the technical improvements. Yes, the 26B and 31B models run on a single H100 GPU, but that's a $20,000 piece of hardware most developers will never touch. The mobile models are more realistic for actual local deployment, but Google's "near-zero latency" claims need real-world testing. Function calling and structured JSON output are table stakes now, not breakthrough features.
For developers, the Apache 2.0 switch is the real news here. If you've been avoiding Gemma due to licensing concerns, that barrier is gone. The mobile variants could be interesting for edge applications, but wait for independent benchmarks before believing Google's performance claims. The hardware requirements for the larger models remain a reality check on what "local AI" actually means in 2024.