Google recently announced its next-generation open-source large language model series, Gemma 4, which has garnered significant attention for a major shift in its licensing strategy. The series not only achieves a significant performance boost but also, for the first time, fully adopts the Apache 2.0 license, offering developers and enterprises unrestricted commercial use and redistribution rights.
Core Model Release: The Gemma 4 Series and Its Performance
The Gemma 4 series includes four models of different sizes, designed to cover a wide range of applications from high-intensity cloud computing to offline execution on mobile devices. The core model of the series is the Gemma 4 31B, with 31 billion parameters, which Google claims achieves extremely high “intelligence-per-parameter”.
According to public benchmark data, the Gemma 4 31B excels in multiple domains:
- Math Ability: Achieved an 89.2% accuracy rate on the AIME math competition benchmark.
- Scientific Reasoning: Reached 84.3% on the GPQA Diamond benchmark.
- Coding Ability: Scored 80.0% on the LiveCodeBench v6 test.
Thanks to its outstanding performance, the model ranks third among global open-source models on the Arena AI leaderboard, run by the LMSYS Organization, with an Elo score of 1452. In terms of hardware, the Gemma 4 31B can run on a single NVIDIA H100 GPU with 80GB of VRAM, while the quantized version is compatible with consumer-grade graphics cards.
The series also includes a 26-billion-parameter Gemma 4 26B model. This model uses a Sparse Mixture-of-Experts (MoE) architecture, activating only 3.8 billion parameters during inference, thus achieving high performance at a lower computational cost. Its scores on AIME, GPQA Diamond, and LiveCodeBench are 88.3%, 82.3%, and 77.1% respectively, placing it sixth on the open-source model leaderboard.
A Key Shift: Fully Embracing the Apache 2.0 Open Source License
The most significant change in this release is the license. Previously, Google’s Gemma 3 series, despite its impressive capabilities, used a custom, restrictive license. This license not only limited many commercial use cases but also allowed Google to unilaterally modify the terms, leading many development teams to turn to models like Alibaba’s Qwen and DeepSeek, which use Apache 2.0 or MIT licenses, due to commercial risks. Tech outlet VentureBeat once commented that this kind of “open source with strings attached” was not “truly open source”.
The release of Gemma 4 marks a fundamental adjustment in Google’s open-source strategy. By adopting the standard Apache 2.0 license, Google has removed all custom terms and usage restrictions, ensuring that the community can freely engage in commercial deployment, derivative development, and redistribution, clearing the way for the Gemma ecosystem to flourish.
Technical Architecture and On-Device Deployment
The Gemma 4 series models are built on the same core technology as Google’s Gemini 3, supporting over 140 languages. The 31B and 26B large models feature a context window of 256K tokens.
For edge computing scenarios, Google has introduced two lightweight models: Gemma E4B and E2B. These models are the result of collaborative optimization with Qualcomm and MediaTek, enabling them to run completely offline on devices like smartphones, Raspberry Pi, and NVIDIA Jetson Nano. They feature a 128K context length and natively support images, videos, and audio as inputs, forming the basis for the next-generation Gemini Nano 4 technology.
Developer Ecosystem and Community Response
Following its release, Gemma 4 quickly received a positive response from the developer community. Hugging Face CEO Clément Delangue called it a “milestone achievement”. The model weights are already available on major platforms like Hugging Face, Kaggle, and Ollama. Support for popular frameworks and tools such as Transformers, llama.cpp, vLLM, Ollama, and NVIDIA NIM is also ready, allowing developers to quickly start and experiment with a simple command like ollama run gemma4.
According to official data, the Gemma series has been downloaded over 400 million times since its initial release, with the community creating over 100,000 fine-tuned variants based on it. The full openness of Gemma 4 is expected to further accelerate its adoption and innovation in various AI applications.
