Google has released Gemma 4, a family of four open-weight artificial intelligence models built on the same research underpinning Gemini 3. The release, announced on April 2, 2026, marks a notable shift in licensing: Gemma 4 is distributed under the Apache 2.0 license, a more permissive framework than the custom terms that governed earlier Gemma versions. Developers have downloaded previous Gemma generations more than 400 million times, producing over 100,000 community variants. Google describes this release as its most ambitious to date.
The open-source AI landscape has been dominated by Chinese models for roughly the past year. DeepSeek, Minimax, GLM, and Qwen have held the top positions on open-model leaderboards, with Chinese open models growing from approximately 1.2 percent of global open-model usage in late 2024 to around 30 percent by the end of 2025. Alibaba‘s Qwen even surpassed Meta‘s Llama as the most widely used self-hosted model worldwide during that period. American alternatives have struggled to keep pace in this environment.
Meta’s Llama had long been the default option for developers seeking a capable, locally runnable model, but its standing has weakened. Questions arose over the true open-source nature of its Meta-controlled license, and its performance fell behind Chinese competitors. The Allen Institute‘s OLMo family attempted to fill the gap without gaining significant traction, and OpenAI‘s gpt-oss models, released in August 2025, were not designed as frontier competitors. Just one day before Gemma 4’s launch, a 30-person American startup called Arcee AI released Trinity, a 400 billion parameter open model that signaled renewed momentum in the domestic open-source scene.
Gemma 4 ships in four sizes designed to cover a broad range of hardware. The Effective 2B and 4B variants target phones and edge devices such as Android phones and Raspberry Pi, running fully offline with near-zero latency, native audio input, and a 128,000-token context window. The 26B Mixture of Experts model prioritizes speed, while the 31B Dense model is optimized for output quality. The two larger models extend context to 256,000 tokens and support native function-calling and structured JSON output for building autonomous agents. All four models process images and video natively, and the larger models’ full-precision weights fit on a single 80GB NVIDIA H100 GPU, with quantized versions available for consumer hardware.
On Arena AI‘s text leaderboard, the 31B Dense model currently ranks third among all open models, while the 26B MoE sits sixth. Google claims both models outperform models up to 20 times their size, a claim that is broadly consistent with the Arena AI rankings, where Chinese models still occupy the top two positions. Google DeepMind CEO Demis Hassabis described Gemma 4 as “the best open models in the world for their respective sizes.”
The Apache 2.0 license is considered a significant development for the developer community. Google’s previous Gemma releases used a custom license that introduced legal uncertainty for commercial applications. Apache 2.0 allows developers to modify, redistribute, and commercialize the models without concern over future changes to the terms. Hugging Face co-founder Clement Delangue praised the move, stating that “local AI is having its moment” and calling it the future of the AI industry.
In testing, Gemma 4 demonstrated reliable code generation, producing functional output on the first attempt without errors. The model applies reasoning processes broadly, which can result in responses that feel over-engineered for straightforward prompts. Creative writing output was described as serviceable. Gemma 4 is available now through Google AI Studio for the 31B and 26B variants, and through Google AI Edge Gallery for the E2B and E4B models. Model weights are also accessible on Hugging Face, Kaggle, and Ollama.
Originally reported by Decrypt.
