Google Releases Gemma 4 Under Apache 2.0 — The Most Capable Open Models It Has Ever Shipped

Google DeepMind shipped Gemma 4 on April 2 and changed what “open” means for its model family. Four sizes — 2B, 4B, 26B MoE, and 31B Dense — all released under Apache 2.0 with zero monthly active user caps and full commercial rights. The days of Gemma’s awkward license carve-outs are over.

The architecture jump is significant. Every model in the Gemma 4 family carries a 256K context window, native multimodal support for vision and audio inputs, and coverage across 140+ languages. The prior Gemma 3 release maxed out at 128K context and was text-only except for the experimental multimodal variants. Gemma 4 ships multimodal by default.

The two smaller models — 2B and 4B — are tagged “Effective” and are designed to run on phones and edge hardware without quantization tricks. Google’s internal benchmarks put them above comparably sized models from Mistral and Meta’s Llama series on instruction-following and coding tasks. Take those numbers with appropriate skepticism, but independent evaluations on Hugging Face’s Open LLM Leaderboard are tracking similarly.

The 31B Dense is the headline model. It places third on the AI Arena leaderboard among all open-weight models, behind only the 405B Llama 3.1 variant and DeepSeek V3-0324. Getting to third with a 31B parameter count, against models five to ten times larger, is a genuine result — not a cherry-picked benchmark.

Day-one support landed for every major inference stack: Hugging Face Transformers, Ollama, vLLM, llama.cpp, MLX for Apple Silicon, and NVIDIA NIM for data centers. You can pull the 4B model with ollama pull gemma4:4b and run it locally today.

The 26B MoE model is worth watching separately. Mixture-of-experts architectures activate only a fraction of total parameters per forward pass, which means the 26B MoE runs closer to a 6-7B dense model in practice while matching or exceeding 26B dense performance on most benchmarks. For teams with constrained GPU budgets, this is the practical choice.

Apache 2.0 licensing is the strategic move here. Gemma 3 shipped under a custom Google license that restricted commercial use above certain traffic thresholds. That restriction caused enterprises to default to Llama and Mistral models for production deployments. Apache 2.0 removes that friction entirely — Gemma 4 can be fine-tuned, modified, redistributed, and shipped in commercial products with no royalties and no usage caps.

This is a direct shot at Meta’s Llama franchise. Meta has dominated the permissive open-model space since Llama 2, partly because Google kept hedging on licensing. Gemma 4 is Google finally playing to win in that market.

The move also pressures the paid API-only providers. When a 31B open model sits three positions from the top of the global leaderboard, the cost calculus for choosing a closed API starts getting harder to justify — especially for non-consumer workloads where data privacy matters.

Weights are available now on the Google Gemma Hugging Face page, Kaggle, and Vertex AI Model Garden.