Chat with Gemma 4

Google's most capable open models — now one click away.
Pick a Gemma 4 model, type your prompt, start chatting.

1 credit per message · Pay as you go or subscribe · Cancel anytime

Gemma 4 Game Master

Game Master

Let Gemma 4 run your text adventure. It remembers your entire world across 256K tokens.

Gemma 4 Code Buddy

Code Buddy

Debug, refactor, and explain code. Gemma 4 scores 80% on LiveCodeBench.

Gemma 4 Book Devourer

Book Devourer

Feed Gemma 4 an entire book. It digests 256K tokens for breakfast.

Gemma 4 Polyglot Pal

Polyglot Pal

140+ languages. Not just translation — Gemma 4 understands cultural context.

Gemma 4 Science Nerd

Science Nerd

84.3% on GPQA Diamond. Ask Gemma 4 the hard questions.

Gemma 4 Math Wizard

Math Wizard

89.2% on AIME 2026. Gemma 4 solves competition-level math step by step.

Gemma 4 Agent Builder

Agent Builder

Gemma 4 has native function calling. Build agents that plan and execute.

Gemma 4 Creative Writer

Creative Writer

Stories, lyrics, ad copy. Give Gemma 4 a spark, get a fire.

What Is Gemma 4?

Gemma 4 is Google DeepMind's most intelligent open model family, built from Gemini 3 research to maximize intelligence per parameter. Every Gemma 4 model is released under the Apache 2.0 license with full commercial freedom — no MAU caps, no usage restrictions. The Gemma 4 model family delivers frontier-level reasoning, code generation, multimodal understanding, and agentic capabilities in compact packages that run on consumer hardware. Gemma 4 is purpose-built for developers, researchers, and builders who want the power of Google's best models without vendor lock-in. On Gemma4.io, you can chat with the two most powerful Gemma 4 models instantly — no GPUs, no setup, no infrastructure.

AIME 2026: 89.2%

The Gemma 4 model achieves 89.2% on the American Invitational Mathematics Examination 2026, demonstrating competition-level mathematical reasoning that rivals models with far more parameters.

LiveCodeBench: 80%

With an 80% score on LiveCodeBench v6, the Gemma 4 model writes, debugs, and refactors production code across dozens of programming languages with professional-grade accuracy.

GPQA Diamond: 84.3%

The Gemma 4 model scores 84.3% on the Graduate-Level Google-Proof Q&A Diamond benchmark, proving expert-level scientific reasoning in physics, chemistry, and biology.

256K Token Context

Every Gemma 4 model on Gemma4.io supports a 256,000 token context window — enough to process entire codebases, full research papers, or book-length documents in a single conversation with the Gemma 4 model.

140+ Languages

The Gemma 4 model goes beyond simple translation. Trained on 140+ languages, the Gemma 4 model understands cultural context, idiomatic expressions, and register-appropriate tone across languages from Arabic to Zulu.

Apache 2.0 License

Every Gemma 4 model is fully open-weight under Apache 2.0. Use Gemma 4 model outputs commercially without restriction. No MAU caps, no acceptable-use clauses — complete freedom to build whatever you want with the Gemma 4 model.

What the Gemma 4 Model Can Do

The Gemma 4 model is not just another chatbot — it is a multimodal reasoning engine with agentic capabilities. Here are four advanced features that set the Gemma 4 model apart from other open models.

Configurable Thinking Mode

The Gemma 4 model supports configurable thinking modes that let you control how deeply it reasons before responding. Enable extended thinking for complex problems where the Gemma 4 model needs to plan, decompose, and verify multi-step solutions — like competition math, multi-file code refactoring, or scientific analysis. Or use fast mode for quick conversational responses. The Gemma 4 model adapts its reasoning depth to match your task, giving you the right balance of speed and accuracy on every message.

Multimodal Vision

The Gemma 4 model natively understands images alongside text. Upload a screenshot, diagram, chart, or photo and the Gemma 4 model will analyze it with the same reasoning power it applies to text. Ask the Gemma 4 model to extract data from charts, describe complex diagrams, debug UI screenshots, read handwritten notes, or compare visual layouts. The Gemma 4 model supports variable aspect ratios and resolutions, so you can send images exactly as they are without cropping or resizing.

Agentic Workflows

The Gemma 4 model is built for autonomous action. With native function calling and structured JSON output, the Gemma 4 model can plan multi-step tasks, call external APIs, navigate applications, and complete complex workflows independently. Build agents that monitor systems, triage issues, draft responses, and execute decisions — all powered by the Gemma 4 model. The Gemma 4 model scores 86.4% on the t2-bench agentic tool use benchmark, proving its ability to orchestrate real-world tool chains reliably.

Fine-Tuning Ready

Every Gemma 4 model is fully open-weight, meaning you can fine-tune the Gemma 4 model on your own data using your preferred frameworks — LoRA, QLoRA, full fine-tuning with JAX, PyTorch, or Keras. Customize the Gemma 4 model for your specific domain: medical, legal, financial, creative, or technical. The Gemma 4 model supports quantized deployment (8-bit and 4-bit) so your fine-tuned Gemma 4 model runs efficiently even on consumer GPUs. Start with Gemma4.io to test, then fine-tune the Gemma 4 model locally when ready.

Gemma 4 Models on Gemma4.io

We offer the two most powerful Gemma 4 models available through Google AI Studio. Both Gemma 4 models support 256K context, multimodal input, function calling, and 140+ languages. Choose the Gemma 4 model that fits your task.

Gemma 4 31B Dense — Maximum Power

The flagship Gemma 4 model. 31 billion parameters in a dense architecture delivering maximum reasoning power. This Gemma 4 model outperforms much larger models on key benchmarks while using significantly less compute. When you need the absolute best Gemma 4 model output for complex reasoning, long-document analysis, or agentic workflows — this is your model.

Gemma 4 26B MoE — Maximum Efficiency

The efficiency champion. This Gemma 4 model uses a Mixture-of-Experts architecture with 128 small experts, activating only 8 plus 1 shared expert per token. The result: this Gemma 4 model achieves roughly 97% of the 31B Dense quality at a fraction of the compute cost. Ideal for high-volume conversational use where you want near-flagship Gemma 4 model intelligence without the overhead.

Frequently Asked Questions About Gemma 4

Have more questions about the Gemma 4 model or Gemma4.io? Reach out at hi@gemma4.io.









Start Chatting with the Gemma 4 Model Now

Pick a Gemma 4 model. Type your prompt. Get instant responses from Google's most capable open Gemma 4 model. No setup required — just the Gemma 4 model, ready to go.