Google DeepMind Launches Gemma 4 Under Apache 2.0, Opening Commercial Use

For two years, enterprise legal teams found Google's custom Gemma license too restrictive to deploy confidently. That barrier fell on April 2 when Google DeepMind released Gemma 4, a family of four open-weight models under the Apache 2.0 license, stripping away the custom clauses, commercial-use carve-outs, and redistribution limits that had kept organizations on the sidelines.

The architectural scope is wide by design. Google released four variants to cover the full deployment spectrum: an Effective 2B (E2B) and Effective 4B (E4B) tuned for mobile and edge hardware, a 26B Mixture-of-Experts model for mid-range server workloads, and a 31B dense model for high-performance inference on developer workstations and cloud infrastructure. The E2B runs on hardware as modest as a Raspberry Pi; the 31B targets production pipelines that previously required closed-API arrangements with providers like OpenAI or Anthropic.

Performance backs the ambition. The 31B model ranks as the number three open model globally on the Arena AI text leaderboard, with the 26B MoE variant at number six. Those rankings translate into measurable gains over the prior generation: AIME 2026 math benchmark scores jumped from 20.8 percent to 89.2 percent compared to Gemma 3, while the GPQA science benchmark climbed from 42.4 percent to 84.3 percent. Google described the family as "our most capable open models to date," noting it draws on the same research that powers the proprietary Gemini 3 architecture.

The license is where the competitive calculus shifts most sharply. Previous Gemma versions used a custom Google license with restrictions that enterprise legal teams found problematic; Apache 2.0 allows unrestricted commercial use, modification, and redistribution with no custom clauses. That means a hospital system building an on-premises diagnostic assistant, a startup embedding a document intelligence layer, or a university running fine-tuning experiments can all do so without signing special commercial agreements with Google. The Google Open Source team wrote that the change "broadens the horizon for Gemma 4's applicability and usefulness." All four models also ship with day-one integrations for Hugging Face Transformers, vLLM, llama.cpp, Ollama, and NVIDIA NIM, reducing the friction of standing up deployments. The prior Gemma lineage had accumulated over 400 million downloads and spawned more than 100,000 community variants; the licensing shift is likely to accelerate both numbers.

The opening raises legitimate concerns alongside the opportunity. Releasing a 31B multimodal model, capable of processing text, images, and a 256,000-token context window, with no usage restrictions means safety filters and content policies become the responsibility of whoever deploys the weights. Enterprises embedding Gemma 4 in customer-facing products inherit the obligation to manage misuse, data handling, and regulatory compliance without Google's guardrails dictating terms.

The timing is notable: as some Chinese AI labs, including Alibaba with its latest Qwen models, have begun pulling back from fully open releases, Google is moving in the opposite direction. That divergence puts pressure on the rest of the market. Closed-API providers now compete with an Apache-licensed model ranked among the world's top three open systems, and the question they face is no longer just about benchmark scores but about what licensing terms developers will tolerate when a capable alternative comes with no strings attached.