<p>Google has released Gemma 4, a family of four open-weight models under an Apache 2.0 license. That last part matters: unlike previous Gemma releases, commercial use, fine-tuning, and redistribution are now unrestricted.</p>
<h2>The model lineup</h2>
<p>The four models cover a range of hardware targets:</p>
<ul>
<li><strong>E2B</strong> (5.1B params) — small, fast, designed for on-device use</li>
<li><strong>E4B</strong> (8B params) — more capable edge model</li>
<li><strong>26B MoE</strong> (A4B active parameters) — Mixture-of-Experts architecture for workstations</li>
<li><strong>31B</strong> — dense flagship model</li>
</ul>
<p>All four support text and vision inputs. The larger models also handle audio. Context windows are 128K for the edge models and 256K (up to 512K in some configurations) for the workstation variants. The audio encoder has been compressed by 50% compared to Gemma 3N.</p>
<p>All models fit on a single H100 GPU. NVFP4 quantized checkpoints for Blackwell hardware are coming soon.</p>
<h2>On-device and Android deployment</h2>
<p>The E2B and E4B models are the architectural foundation for the upcoming Gemini Nano and are deployed on Android via AI Core. Developers can opt into the core developer preview, download the model, and use the MLKit Prompt API to prototype features. Code written now is forward compatible with future MPU optimizations.</p>
<p>Gemma 4 is also available for local AI code assistance in Android Studio, which is useful for developers who need offline access or strict data privacy. It supports agent mode with no quota limits.</p>
<h2>Edge and desktop deployment via NVIDIA</h2>
<p>NVIDIA supports Gemma 4 across its hardware stack:</p>
<ul>
<li><strong>DGX Spark</strong> — local prototyping and agentic workflows</li>
<li><strong>Jetson Orin Nano</strong> — edge robotics and embedded systems</li>
<li><strong>RTX GPUs</strong> — desktop development</li>
</ul>
<p>Deployment is supported through vLLM, Ollama, llama.cpp, and Unsloth. Fine-tuning is available via NeMo Automodel using SFT and LoRA. Enterprise users can access a hosted NIM API for free prototyping or self-hosted production deployment under an NVIDIA Enterprise License.</p>
<h2>Docker Hub</h2>
<p>Gemma 4 is also available on Docker Hub as OCI artifacts. You can pull any variant with a single command:</p>
<pre><code>docker <span class="hljs-keyword">model</span> pull gemma4
</code></pre>
<p>No custom toolchains required. Integration with Docker Model Runner for Docker Desktop is coming soon.</p>
<h2>Where to get it</h2>
<p>Models are available on Hugging Face under Apache 2.0 and on Google Cloud, with serverless deployment via Cloud Run using Nvidia RTX Pro 6000 GPUs.</p>


Collections

Google has released Gemma 4, a family of four open-weight models (E2B, E4B, 26B MoE, 31B) under Apache 2.0, enabling unrestricted commercial use, fine-tuning, and redistribution. All models support text and vision inputs, with larger variants adding audio support and context windows up to 512K tokens. Deployment options span Android via AI Core and MLKit, NVIDIA hardware (DGX Spark, Jetson Orin Nano, RTX GPUs) through tools like vLLM, Ollama, and llama.cpp, Docker Hub as OCI artifacts, and Google Cloud via Cloud Run. Fine-tuning is supported through NeMo Automodel with SFT and LoRA.

Gemma 4: what's new across the full model family