In a significant move to bolster its open-source AI initiatives, Google today unveiled Gemma 4—a set of four models built on research identical to that of Gemini 3 and released under the Apache 2.0 license, marking a shift from earlier versions’ restrictive terms.
The release follows over 400 million downloads of previous Gemma iterations, which have resulted in more than 100,000 community-driven adaptations. This version stands out as Google’s most ambitious open-source endeavor yet.
Announced on Twitter by Google (@Google) on April 2, 2026, Gemma 4 represents the company’s most advanced models to date, leveraging Gemini 3’s groundbreaking research for enhanced reasoning and workflow capabilities directly accessible on personal hardware.
The past year has seen China dominate the open-source AI leaderboard with models such as DeepSeek, Minimax, GLM, and Qwen. By late 2025, Chinese models accounted for approximately 30% of global usage, overtaking Meta’s Llama in terms of worldwide self-hosted model popularity.
Llama once served as a top choice for developers seeking capable local models; however, its reputation has waned due to licensing issues and performance lag. Attempts by the Allen Institute’s OLMo family and OpenAI’s release of gpt-oss models in August 2025 could not regain significant traction.
A notable development came from Arcee AI, a U.S.-based startup, which introduced Trinity—a 400 billion parameter model—indicating that American innovation remains viable. Now, Google DeepMind’s backing of Gemma 4 positions it as a formidable contender in the open-source AI landscape.
The models are derived from Gemini 3 and come in four sizes: Effective 2B and 4B for mobile devices, a 26B Mixture of Experts model designed for speed, and a 31B Dense model aimed at superior quality. On Arena AI’s text leaderboard, the 31B Dense ranks third while the 26B MoE is sixth, both outperforming models over twenty times their size.
During testing, Gemma 4 demonstrated competence with certain caveats: it often applies reasoning to straightforward tasks, making responses appear overly complex. Its creative writing capabilities are adequate but improve with precise guidance. The model excelled in coding tasks, generating a functional game on its first attempt without errors—a testament to its zero-shot reliability.
The four variants cater to diverse hardware needs: E2B and E4B for mobile and edge devices; 26B and 31B for workstations and cloud deployments. All models support image and video processing natively, with the larger versions fitting on a single NVIDIA H100 GPU in full-precision or running quantized on consumer hardware.
The transition to an Apache 2.0 license eliminates legal ambiguities present in previous Gemma releases, allowing developers to modify and commercialize freely without concern for future changes by Google. Clement Delangue of Hugging Face commended this move as pivotal for local AI’s future, while Demis Hassabis, CEO of Google DeepMind, hailed Gemma 4 as the best open models globally in their respective categories.
Hassabis emphasized on Twitter that Gemma 4 is available in four sizes optimized for specific tasks: 31B dense for performance, 26B MoE for low latency, and effective 2B & 4B for edge applications. While proprietary systems still outperform open-weight models in rigorous benchmarks, the competition has thinned considerably. Gemma 4 can be accessed via Google AI Studio (for 31B and 26B) or Google AI Edge Gallery (for E2B and E4B), with model weights available on platforms like Hugging Face, Kaggle, and Ollama.