Google announces Gemma 4 open AI models, switches to Apache 2.0 license
Four new models optimized for local hardware, with a 256k context window and a major licensing shift.
Google has released Gemma 4, a significant upgrade to its family of open-weight AI models, now available under the permissive Apache 2.0 license. This move directly addresses developer frustrations with the previous custom license, which had restrictive prohibited-use policies. The new suite consists of four models optimized for local hardware: the 26B Mixture of Experts (MoE) and 31B Dense models for powerful workstations, and the Effective 2B (E2B) and Effective 4B (E4B) models for mobile and edge devices like smartphones and Raspberry Pi.
The 26B MoE model is engineered for speed, activating only 3.8B of its parameters during inference for high token throughput, while the 31B Dense model prioritizes quality for fine-tuning. Both large models support a 256k token context window. The smaller E2B and E4B models, co-developed with Qualcomm and MediaTek, promise "near-zero latency" and low memory usage. All models are built on the same tech as Gemini 3, offering improved reasoning, math, code generation, and native support for agentic workflows with function calling and structured JSON output.
Google claims the 31B Dense model will debut at #3 on the open model Arena leaderboard, a notable achievement given its smaller size compared to competitors. The shift to Apache 2.0 is a landmark change, removing legal uncertainty and making Gemma 4 a more viable, commercially-friendly option for developers and enterprises looking to build and deploy powerful AI locally without cloud dependency or restrictive licensing terms.
- Licensing Shift: Drops custom 'Gemma License' for standard Apache 2.0, removing restrictive use clauses and unilateral update powers.
- Four Model Sizes: Includes 26B Mixture of Experts (for speed), 31B Dense (for quality), and mobile-optimized 2B/4B 'Effective' models.
- Local-First Design: Optimized for on-device inference with 256k context windows, native function calling, and improved code generation.
Why It Matters
Enables powerful, commercially-safe local AI deployment with permissive licensing, reducing cloud dependency and vendor lock-in for developers.