Google released Gemma 4 last week, and the thing that matters most about it is not the benchmarks. It is the license.

Gemma 4 ships under Apache 2.0 — the permissive open-source license that places no meaningful restrictions on commercial deployment, modification, or redistribution. [1] Any developer, any company, any government can take these models, run them locally, fine-tune them for proprietary use, and ship products without asking Google's permission or paying Google a fee. That is a structural choice, not a marketing one.

The technical specifications are formidable. The family includes four model sizes — 2 billion, 4 billion, 26 billion, and 31 billion parameters — in both pre-trained and instruction-tuned variants. [2] The larger models feature a 256K-token context window, which means they can process roughly the equivalent of a 500-page book in a single prompt. The smaller models support 128K tokens. All are multimodal, handling text, images, and audio inputs. [3]

Google says Gemma 4 is built from the same research that powers Gemini 3, the company's closed commercial model. [4] The benchmarks support this claim. On standard reasoning tasks, the 31-billion-parameter Gemma 4 approaches the performance of models that cost money to use — a convergence that the open-source community has been predicting and the closed-model companies have been dreading.

The practical implications are significant. A 4-billion-parameter model that fits on a phone changes what "AI deployment" means. It is no longer a cloud API call. It is a local computation. Hugging Face provided day-zero support for all Gemma 4 variants, and the models are already running on vLLM, Ollama, and other inference engines that developers actually use. [5]

The competitive context is the week's other AI headline: OpenAI just closed a $122 billion funding round at an $852 billion valuation. That valuation assumes continued dominance in a market where the best models are proprietary and expensive. Google is now releasing comparable models for free. The tension between these two strategies — capitalize versus commoditize — will define the next phase of the industry.

For the developer sitting at a desk in Lagos or Bangalore or São Paulo, Gemma 4 means that state-of-the-art AI is no longer behind a paywall. Whether that is good for Google's business model is one question. Whether it is good for the world is a different one, and arguably more interesting.

-- KENJI NAKAMURA, Tokyo