Google DeepMind released Gemma 4 this week, a family of four open-weight AI models that represent the company's most aggressive move yet in the open-source AI race. Released under the commercially permissive Apache 2.0 license — a notable shift from previous, more restrictive terms — the models are designed to run on everything from smartphones to single-GPU workstations.
The family ships in four sizes: two edge-optimized models called Effective 2B and Effective 4B, built for phones and IoT devices; a 26-billion-parameter Mixture of Experts model focused on inference speed; and a 31-billion-parameter dense model aimed at maximum quality. Google says the architecture draws directly from the same research foundation as its proprietary Gemini 3, effectively giving developers a distilled version of its most capable commercial technology.
The benchmarks tell a compelling story. The flagship 31B model scores 89.2 percent on AIME 2026 mathematics tests, 84.3 percent on GPQA Diamond for scientific reasoning, and 80.0 percent on LiveCodeBench v6 for competitive coding. On the Arena AI text leaderboard, it currently ranks as the third most capable open model in the world, with the 26B variant at sixth — both outperforming models with vastly larger parameter counts.
That said, not everyone is celebrating without caveat. Independent analysis suggests Gemma 4 still trails the most powerful Chinese open-source models. Alibaba's Qwen 3.5, Zhipu AI's GLM-5, and Moonshot AI's Kimi K2.5 all edge ahead in head-to-head comparisons, though the margins are slim. The gap to OpenAI's own open model, GPT-OSS-120B, is more favorable for Google — Gemma 4 outperforms it handily despite being a fraction of the size.
What makes the release particularly significant is the combination of capability and accessibility. The larger models fit on a single 80GB NVIDIA H100 GPU in unquantized form, while quantized versions run natively on consumer hardware. The edge models, developed in collaboration with Qualcomm and MediaTek, operate entirely offline with near-zero latency. Android developers can already prototype agentic workflows through Google's AICore Developer Preview, with forward compatibility planned for Gemini Nano 4.
The technical feature set is ambitious. All four models natively process images and video with variable resolution support, while the smaller E2B and E4B models add native audio input for speech recognition. Context windows range from 128,000 tokens on edge models to 256,000 tokens on the larger variants. Native function calling, structured JSON output, and system instruction support make the models immediately suitable for agentic applications — the category every major lab is now racing to dominate.
The Apache 2.0 licensing decision deserves its own attention. Previous Gemma releases carried custom licenses with various restrictions. By switching to a standard open-source license, Google is making a clear bid for developer adoption and ecosystem lock-in through quality rather than legal constraints. It is also a strategic concession to the reality that Chinese competitors have been releasing increasingly capable models under permissive terms, building massive developer communities in the process. Gemma models have already been downloaded over 400 million times, spawning more than 100,000 community variants in what Google calls the "Gemmaverse."
Model weights are available immediately on Hugging Face, Kaggle, and Ollama, with day-one support across major frameworks including vLLM, llama.cpp, MLX, and LM Studio. Cloud deployment options span Vertex AI, Cloud Run, and Google Kubernetes Engine.
The release arrives at a pivotal moment for open-source AI. The gap between open and proprietary models continues to narrow, and Google appears to have decided that leading the open-weight category is worth more than protecting its proprietary advantage. For the thousands of companies and developers building AI products on their own infrastructure, Gemma 4 may be the most consequential open model release of the year so far.
