Meet Gemma 4

Gemma 4 is Google DeepMind's latest family of lightweight, state-of-the-art open models. It's built from the same research and technology used to create Gemini, but released with open weights so you can download, inspect, fine-tune, and deploy the models on your own infrastructure.

Gemini is Google's closed, hosted flagship model, available through Google's APIs. Gemma 4 shares much of the underlying research but is released as open weights under a permissive license, so you can run it locally, fine-tune it for your own data, and deploy it without sending requests to Google.

Yes. Gemma 4 is released under the Gemma license, which permits both research and commercial use. You are free to build products, services, and businesses on top of Gemma 4 — including fine-tuned derivatives — subject to the license's responsible use policy.

It depends on the size. Gemma 4 Nano (2B) runs comfortably on a modern laptop or phone. The 9B and 27B models run on a single high-end GPU such as an NVIDIA RTX 4090 or H100. The 70B Ultra model is best suited for multi-GPU servers or TPU pods. Quantized variants (GGUF, AWQ) reduce requirements further for on-device use.

Thinking variants are Gemma 4 models trained to reason step-by-step before producing a final answer. They trade a small amount of latency for substantially better performance on math, science, coding, and multi-step reasoning benchmarks — as seen in our AIME 2026 and GPQA Diamond results.

Gemma 4 is available through Google AI Studio, Vertex AI, Kaggle, Hugging Face, and Ollama. You can also run it directly with popular inference frameworks including llama.cpp, vLLM, and MLX for Apple Silicon.

Absolutely. Gemma 4 supports the full range of fine-tuning techniques: full supervised fine-tuning, LoRA, QLoRA, DPO, and RLHF. We provide reference training recipes and notebooks for all four model sizes to help you get started.

Every Gemma 4 release goes through extensive safety evaluation including red-teaming, bias testing, and responsible AI reviews. We publish detailed model cards for each variant and ship with a built-in responsible use policy. Because the weights are open, the broader research community can audit and improve safety as well.

Gemma 4 introduces Thinking variants for step-by-step reasoning, a new sparse Mixture-of-Experts architecture (26B A4B), native multimodal input across the entire family, a 128K context window, and substantial gains on math, coding, and agentic benchmarks. Gemma 4 27B IT beats Gemma 3 27B IT by over 25 points on AIME 2026 and nearly triples its score on LiveCodeBench v6.

Yes. All instruction-tuned variants were trained with structured tool-use data and support function calling via a standard JSON schema. They can plan multi-step workflows, invoke external APIs, and recover from tool errors — as reflected in our τ2-bench retail scores.

Gemma 4 was trained on more than 140 languages with balanced representation across European, Asian, African, and Indic language families. Instruction tuning covers the top 40 languages with human-verified evaluations; the remaining languages benefit from strong transfer learning.

The Gemma community lives on GitHub, the Google Developer forums, and the Hugging Face discussions board. Security-sensitive reports can be sent privately to the DeepMind responsible disclosure address listed in every model card.