Google DeepMind Releases Gemma 4, Open Models Built on Gemini 3 Technology

Google DeepMind released the Gemma 4 family of open models on April 2, 2026, built on the same technology as the flagship Gemini 3. It is the first Gemma generation released under the fully permissive Apache 2.0 license, allowing unrestricted commercial use, modification and redistribution without the limits imposed by Google's earlier custom licensing terms.

Five sizes in one family

Gemma 4 splits into two groups. The E2B and E4B variants are designed for maximum memory and compute efficiency, built to run offline on smartphones, Raspberry Pi boards and other IoT devices. Alongside them are three larger models, 12B, 26B and 31B, aimed at consumer graphics cards and offering, according to Google, a strong quality-to-parameter ratio.

All variants draw on the research and technology behind Gemini 3, which Google says translates into better reasoning, multi-step planning, and the ability to operate autonomously in agentic mode without requiring additional fine-tuning for specific tasks.

Benchmark results

The largest model, Gemma 4 31B IT Thinking, scored 1452 Elo points on the Arena AI leaderboard for text tasks and 85.2 percent on the MMLU Multilingual knowledge test. On the multimodal MMMU Pro benchmark it scored 76.9 percent, on the AIME 2026 math benchmark 89.2 percent, on the LiveCodeBench coding benchmark 80.0 percent, and on the GPQA Diamond science benchmark 84.3 percent.

Google notes that every Gemma 4 size handles text and image inputs, while the on-device models additionally understand audio. The family supports more than 140 languages, and the company stresses this covers not just translation but understanding of each language's cultural context.

On-device performance without the cloud

A key part of the launch is performance on lower-end hardware. According to data Google published for its LiteRT-LM engine, the E2B model can process 4,000 input tokens across two different tasks in under 3 seconds, using less than 1.5 GB of memory on some devices.

On a Raspberry Pi 5 running CPU-only, Google recorded 133 tokens per second during input processing and 7.6 tokens per second during response generation. On a Qualcomm Dragonwing IQ8 NPU chip the results are far higher, 3,700 tokens per second in prefill and 31 tokens per second in decoding, showing how heavily performance depends on dedicated neural hardware.

Where to find the models

Gemma 4 models are available on Hugging Face, Kaggle, Ollama, LM Studio and as Docker images, and on Android they can be accessed through the public AICore beta as well as the Google AI Edge Gallery app for iOS and Android. Technical documentation and model cards were published alongside the launch.

For developers, this means the ability to run fully local AI agents, generate code offline, and build multimodal applications without sending data to the cloud, which could matter especially for companies in privacy-sensitive sectors such as finance or healthcare.

The switch from a custom license to Apache 2.0 removes some of the legal uncertainty that came with earlier Gemma deployments in commercial products. The previous terms required accepting an additional Google license agreement and included restrictions on certain uses, while Apache 2.0 is a standard widely accepted across the open-source industry.

The scale of Gemma's popularity, over 400 million downloads and more than 100,000 community-built derivative variants, makes it one of the most widely used open model families alongside Meta's LLaMA and Chinese models such as Qwen and GLM. The Gemma 4 launch is part of an ongoing race between providers of large closed models and increasingly capable open-source alternatives that can be run without API access fees.