Google DeepMind released Gemma 4 on April 2, 2026 — a family of open-weights models built on the same research foundation as Gemini 3, available under a fully permissive Apache 2.0 license with no usage restrictions. The 31B Dense model currently ranks third among all open models globally on the Arena AI text leaderboard. The 26B Mixture-of-Experts variant ranks sixth. Both run on a single 80GB NVIDIA H100 GPU — hardware that’s now accessible to individual developers and small teams, not just enterprise data centers.
Four Model Sizes, One Clear Strategy
Gemma 4 ships in four sizes, each targeting a different hardware tier:
- E2B and E4B (edge models): Designed for smartphones, IoT devices, and Raspberry Pi — multimodal by default, supporting text, images, and audio natively. Built for on-device, offline AI with zero cloud dependency and sub-second latency.
- 26B Mixture-of-Experts (MoE): Activates only 3.8 billion of its total parameters during inference — delivering fast tokens-per-second while maintaining high quality. Best fit for high-throughput enterprise workloads.
- 31B Dense: Maximum raw quality. Designed for developers who want fine-tuning headroom and the strongest possible output on a single GPU. Currently ranked #3 open model globally.
Natively Multimodal From the Ground Up
Previous Gemma versions treated vision and audio as bolt-on features. Gemma 4 treats multimodality as a first-class capability from the architecture level. Supported inputs across model sizes include:
- Vision: Object detection, OCR (multilingual), document and PDF parsing, chart comprehension, UI understanding, handwriting recognition, and variable-resolution image handling
- Video: Up to 60 seconds of video input (at 1 fps) for temporal reasoning tasks
- Audio (E2B/E4B): Automatic speech recognition and speech-to-text translation, natively processed without an external transcription layer
The vision encoder handles variable aspect ratios with a configurable visual token budget (70 to 1,120 tokens) — a practical improvement over square-cropping approaches that lose information from non-square images.
The Apache 2.0 License Is the Biggest Change
Previous Gemma releases shipped with custom Google licenses that imposed restrictions on commercial use, fine-tuning, and redistribution — restrictions that kept enterprise legal teams cautious. Gemma 4 ships under Apache 2.0: fully open, commercially usable, modifiable, and redistributable without restriction. This removes the primary adoption barrier that had kept Gemma behind Llama and Qwen in enterprise deployment.
Google already reports over 100,000 variants of previous Gemma models in use — the Apache 2.0 switch is likely to dramatically accelerate that number.
The Practical Cost Case for Enterprise
The competitive positioning against frontier APIs is direct. Gemma 4 trails Claude Opus 4.6 and GPT-5.4 by roughly seven to ten benchmark points — a real gap for complex reasoning tasks. But it runs locally for free on hardware that costs a few hundred dollars to purchase outright, compared to frontier API costs of $2.50 to $15+ per million tokens.
Analysis suggests that approximately 70% of typical enterprise workloads fall within the capability range where open models perform well — summarization, translation, document Q&A, classification, and routine data extraction. For those tasks, routing to a local Gemma 4 instance instead of a frontier API can cut cloud AI spend by 60% or more. The smart enterprise stack in 2026 is increasingly a split architecture: open models locally for routine tasks, frontier APIs for complex reasoning and agentic work.
Availability
Gemma 4 is available now via Google AI Studio (31B and 26B), Google AI Edge Gallery (E2B and E4B), Hugging Face, NVIDIA NIM and NeMo, Ollama, Docker, and multiple other platforms. Day-zero support from NVIDIA (optimized for Blackwell GPUs) and AMD (optimized for Ryzen AI and Instinct GPUs) means the models are ready for immediate deployment without waiting for ecosystem support to catch up.
Conclusion
Gemma 4 is the most capable open-weights model Google has released, and it arrives at a moment when the gap between open and proprietary AI is narrowing faster than most expected. For developers building on local infrastructure or enterprises looking to reduce frontier API costs, it’s the most compelling option Google has offered to date. Browse our directory to explore Gemma 4 alongside every other open and proprietary model available for your workflows.
