Overview
Gemma 4 is Google DeepMind's fourth-generation open-weight model family, released under the Apache 2.0 license on March 31, 2026. It introduces native multimodal support, audio input on edge models, and significantly improved reasoning capabilities compared to Gemma 3, which was limited to text and image input.
This release spans four official model sizes — E2B, E4B, 26B A4B MoE, and 31B Dense — covering deployment scenarios from smartphones to server clusters. As of April 1, 2026, the 31B model ranked #3 among open models on the Arena AI text leaderboard, while the edge variants are designed for fully offline deployment on resource-constrained devices.
What's New
Native Multimodal Input Across All Models
Gemma 4 adds native support for text, image, and video input (up to 60 seconds at 1 fps) across the entire model family. The E2B and E4B edge variants also support audio input for speech recognition and translation, with a maximum length of 30 seconds. This is a major upgrade from Gemma 3, where only the 4B+ models supported image input and no models handled video or audio natively. For teams evaluating AI text generator options, Gemma 4's multimodal capabilities expand what's possible with local inference.
Four New Model Sizes with Efficiency Architecture
The lineup brings Gemma 4-specific edge-optimized E2B and E4B variants alongside redesigned larger models:
- E2B — Activates ~2B parameters during inference, optimized for maximum speed (3x faster than E4B) and lowest latency on phones and IoT devices
- E4B — Activates ~4B parameters for higher reasoning power while maintaining on-device efficiency
- 26B MoE — Mixture-of-Experts architecture activating only 3.8B of its total parameters, balancing quality with latency for desktop and small server deployment
- 31B Dense — Full dense architecture maximizing raw quality, positioned as the primary foundation for fine-tuning and server deployment
Expanded Context Windows
Context capacity has doubled for larger models: the 26B and 31B variants now support 256K tokens (up from 128K in Gemma 3), while edge models maintain 128K — still a massive improvement over Gemma 2's 8K limit.
Apache 2.0 License (New)
Gemma 4 drops the previous Gemma license in favor of Apache 2.0, removing monthly active user caps and acceptable-use restrictions. This gives developers full commercial freedom with no registration, royalties, or usage limits — a significant licensing change from all previous Gemma versions.
Enhanced Agentic Capabilities
Purpose-built for AI agent workflows, Gemma 4 adds native function calling, structured JSON output, and system instruction support. The models demonstrate strong multi-step planning and tool-use performance, designed for building autonomous agents that navigate apps and complete tasks.
Performance Benchmarks
Gemma 4 delivers substantial improvements over Gemma 3 across all major benchmarks:
| Benchmark | Gemma 4 31B | Gemma 3 27B | Improvement |
|---|---|---|---|
| Arena AI (text ELO) | ~1452 (#3 open) | Lower ranking | Significant |
| BigBench Extra Hard | 74.4% | 19.3% | +55.1 pts |
| MMMLU (multilingual) | 88.4% | 70.7% | +17.7 pts |
| AIME 2026 (math) | 89.2% | 20.8% | +68.4 pts |
| LiveCodeBench (coding) | 80.0% | 29.1% | +50.9 pts |
| GPQA Diamond (science) | 84.3% | 42.4% | +41.9 pts |
The edge models also show strong performance relative to their size, with the 26B MoE ranking #6 among open models on Arena AI despite activating only 3.8B parameters during inference.
Pricing & Plans
Gemma 4 is completely free and open source.
Open Source (Free)
- All model weights downloadable without restrictions under Apache 2.0
- No registration, MAU caps, or usage limits
- Full commercial use permitted with no royalties
- Self-host on your own infrastructure at zero software cost
Third-Party Hosted API
- Available through Google Cloud Vertex AI, OpenRouter, and other providers
- Pricing varies by provider and model size; check each provider's rate card
- Kaggle model access and Colab-based experimentation are available, subject to each platform's current runtime availability and limits
Hardware Costs (Self-Hosting)
- E2B/E4B: Smartphones and devices with 4-8 GB RAM
- 26B MoE: Positioned for desktop computers and small servers
- 31B Dense: Positioned for large servers or GPU clusters
Best For
- Developers building privacy-first mobile apps that need multimodal AI running entirely on-device without cloud dependencies
- Enterprises migrating from proprietary AI APIs to self-hosted open models for cost control and data sovereignty
- Researchers who need a commercially permissive foundation model for fine-tuning on domain-specific multimodal tasks
- IoT and robotics teams deploying AI on edge hardware like Raspberry Pi or NVIDIA Jetson with strict power and memory constraints
- Startups looking to integrate frontier-level AI into products without per-token API costs or licensing overhead
FAQ
How does Gemma 4 compare to Gemma 3?
Gemma 4 adds video support across the family, native audio on E2B and E4B, and raises larger models to 256K tokens, and delivers major benchmark improvements — including a 55-point gain on BigBench Extra Hard. It also switches to Apache 2.0 licensing, removing the usage restrictions present in previous Gemma versions.
Can I upgrade from Gemma 3 fine-tunes to Gemma 4?
Gemma 4 uses a different architecture from Gemma 3, so existing LoRA adapters and fine-tuned weights are not directly compatible. You will need to re-run fine-tuning on the Gemma 4 base models using your training data. Google provides official fine-tuning guides for JAX, Keras, and Hugging Face Transformers.
Which Gemma 4 model should I choose?
Choose E2B for maximum speed on mobile devices, E4B for better reasoning on phones and tablets, 26B MoE for desktop applications balancing quality and latency, and 31B Dense for maximum quality on servers or as a fine-tuning base. The AI app builder category can help you find tools to integrate these models.
Is the Apache 2.0 license really unrestricted?
Yes. Apache 2.0 permits commercial use, modification, and redistribution subject to the license terms, replacing Gemma's earlier custom licensing approach. This is a significant change from the previous Gemma license, which included acceptable-use restrictions and required acknowledgment of Google's usage policies.
What platforms support Gemma 4 on day one?
Gemma 4 launched with official support across Hugging Face, LiteRT-LM, vLLM, llama.cpp, MLX, Ollama, NVIDIA NIM/NeMo, LM Studio, Docker, and related tooling, Google Cloud Vertex AI, Google Kubernetes Engine, and NVIDIA's RTX AI tools. Mobile deployment is supported through Google AI Edge Gallery and the MediaPipe LLM Inference API.
