Google has released Gemma 4 — its strongest open model family to date. It shares lineage with the closed-source Gemini 3 but ships under Apache 2.0: you can download the weights and run them on your own machines, in your own cloud accounts, with data never leaving your premises. For developers who need data sovereignty, offline inference, or just lower API bills, this generation finally brings "frontier capability" and "runs out of the box locally" together.

Four Sizes — Pick by Hardware

Gemma 4 isn't one model but four versions optimized for different hardware:

The 26B MoE takes the "low latency" route — 26B total parameters with only 3.8B active at inference, so tokens-per-second is fast. The 31B Dense takes the "highest quality + fine-tuning base" route. Google says the 31B currently ranks 3rd among open models worldwide on the Arena AI text leaderboard, with the 26B at 6th.

Core Capabilities

• Long context: 128K on the edge models, 256K on the 26B/31B — enough to fit an entire code repository or a long document into a single prompt
• Native multimodality: all four versions natively process video and images (variable resolution, OCR, chart understanding); the E2B/E4B also support native audio input
• Agent workflows: native function calling, structured JSON output, and system instructions — well suited to building automation agents
• Code generation: works as a local coding assistant, writing code offline
• 140+ languages trained natively
• Reasoning: multi-step planning and deep logic, with marked gains on math and instruction-following tasks

How to Get It Running

Google has spread availability wide, covering nearly every mainstream local inference stack:

Try in the Cloud (Zero Setup)

• Google AI Studio: run the 31B and 26B MoE directly
• Google AI Edge Gallery: run the E4B and E2B

Local Deployment

Weights are published on Hugging Face, Kaggle, and Ollama, with day-one support across:

• Hugging Face (Transformers, TRL, Transformers.js, Candle)
• Ollama (ollama run gemma4:31b-it-q4_K_M)
• LM Studio
• vLLM, llama.cpp, MLX
• NVIDIA NIM / NeMo
• Unsloth, SGLang, Cactus, Baseten, Docker, MaxText, Tunix, Keras

Android Development

The edge models are callable via the AICore Developer Preview, forward-compatible with Gemini Nano 4. You can also prototype agent flows with Android Studio's Agent Mode.

💡 Tip: for a quick try, the fastest path is a single Ollama command: ollama run gemma4:31b-it-q4_K_M. For an in-IDE coding assistant, LM Studio plus the Continue plugin is the smoothest combination.

Fine-Tuning

Gemma 4 is designed to be fine-tunable on consumer hardware (a gaming GPU, say). Available platforms include:

• Google Colab (free/paid)
• Vertex AI
• Your own GPUs

The official examples cite two fine-tuning cases: INSAIT trained the Bulgarian-first BgGPT with it, and Yale University used it for Cell2Sentence-Scale to discover new cancer treatment pathways.

Ecosystem and Hardware Support

• NVIDIA: optimized across the line, from Jetson Orin Nano to Blackwell GPUs
• AMD: integrated via the open-source ROCm stack
• TPU: deployable at scale on Trillium and Ironwood TPUs
• Google Cloud: Vertex AI, Cloud Run, GKE, Sovereign Cloud, and TPU-accelerated inference

License

Apache 2.0 — commercially friendly, with no usage gates. You can deploy the model locally or in any cloud environment, keeping full control over your data. That's a notably bigger opening from Google than in previous generations.

Who It's For

• Teams that need data sovereignty: finance, healthcare, government — anywhere data can't be sent to a third-party API
• Developers looking to cut API costs: deploy once, call without limit, zero marginal cost
• Offline scenarios: mobile, IoT, and environments with unreliable networks
• Researchers: Apache 2.0 plus multiple sizes makes experimentation and fine-tuning easy

References

• Gemma 4 official blog post (Google)
• google/gemma-4-31B-it (Hugging Face)
• Gemma 4 model documentation (Google AI for Developers)
• gemma4:31b-it-q4_K_M (Ollama)
• google/gemma-4-31b (LM Studio)