- Google released Gemma 4 on April 2, 2026, offering four model sizes: 31B Dense, 26B MoE, Effective 4B, and Effective 2B, all under the Apache 2.0 license.
- The 31B Dense model ranks third on the Arena AI text leaderboard and outperforms models with 20 times more parameters, including GPT-OSS-120B and Qwen3.5-122B.
- Gemma 4 is built from the same research and architecture as Google’s Gemini 3 and natively processes video, images, and text across more than 140 languages.
- All models are available on Hugging Face, Kaggle, Ollama, Google AI Studio, and Google AI Edge Gallery.
What Happened
Google released Gemma 4 on April 2, 2026, a family of open-weight models that the company describes as its most capable open models to date. Clement Farabet, writing on the Google AI blog, characterized Gemma 4 as “byte for byte, the most capable open models,” purpose-built for advanced reasoning and agentic workflows. The release includes four model sizes: a 31B parameter Dense model, a 26B parameter Mixture of Experts (MoE) model, and two smaller variants designated Effective 4B (E4B) and Effective 2B (E2B) for on-device deployment.
All models are released under the commercially permissive Apache 2.0 license, a significant departure from Google’s previous Gemma license that carried certain usage restrictions. The Google DeepMind model page confirms availability through Hugging Face, Kaggle, Ollama, Google AI Studio, and Google AI Edge Gallery.
Why It Matters
Gemma 4 enters a competitive open model landscape that includes Meta’s Llama 4, Mistral Large 3, and Alibaba’s Qwen 3.5 series. Google’s decision to adopt the Apache 2.0 license removes a friction point that limited enterprise adoption of earlier Gemma releases and puts Google’s open models on equal licensing footing with Meta’s Llama family. The performance claims, if independently verified, position Gemma 4 as the strongest open model relative to its parameter count currently available.
The release also reflects Google’s strategy of deriving open models from its proprietary Gemini research. Gemma 4 is built from the same technology and research as Gemini 3, which means improvements to Google’s frontier closed model now flow more directly into its open-weight offerings. This closed-to-open pipeline gives Google a structural advantage over competitors whose open and closed model lines are developed independently.
Technical Details
The 31B Dense model ranks third on the Arena AI text leaderboard among open models, while the 26B MoE variant ranks sixth. Google claims both models outperform competitors with 20 times their parameter count, specifically naming GPT-OSS-120B, Qwen3.5-122B, and Mistral-Large-3 as benchmarks the Gemma 4 models exceed. The MoE architecture in the 26B model activates only a subset of its parameters per inference pass, reducing computational cost while maintaining model capacity. This design makes the 26B MoE particularly efficient for production deployment where cost per query is a key consideration.
All Gemma 4 models natively process video and images alongside text and are trained on more than 140 languages, making them among the most multilingual open models available. The Effective 4B and Effective 2B variants are designed to run on consumer hardware, including smartphones and edge devices, and are accessible through the Google AI Edge Gallery and the AICore Developer Preview for Android. Nathan Lambert, writing on Interconnects AI, analyzed what makes an open model succeed and highlighted Gemma 4’s combination of permissive licensing, strong benchmarks, and on-device capability as the three factors that most strongly drive developer adoption and community engagement.
Who’s Affected
Developers and enterprises evaluating open-weight models for production use now have a competitive option from Google that combines strong benchmark performance with unrestricted commercial licensing. On-device AI developers benefit from the E2B and E4B variants, which can run locally on Android devices through the AICore Developer Preview. Cloud providers including Google Cloud have already integrated Gemma 4 into their managed AI services. Researchers working on multilingual NLP applications gain access to a model with native support for over 140 languages at a scale that was previously limited to closed commercial models.
What’s Next
Google indicated that Gemma 4 availability on Google Cloud is immediate, with additional platform integrations planned. Independent benchmark evaluations from organizations like Hugging Face and the LMSYS Chatbot Arena will determine whether Gemma 4’s claimed performance advantages hold under standardized testing conditions. The Apache 2.0 license is expected to accelerate fine-tuning and derivative model development across the open-source AI community, and early fine-tuned variants are likely to appear on Hugging Face within days of the release.
