Technical Architecture and Model Specifications

Gemma 4 is built on the same world-class research and technology as Gemini 3, and represents the most powerful model series that developers can run on their own hardware. Google has released four model specifications this time, precisely optimized for different hardware environments:

Workstation-class Large Models:

  • 31B Dense Model: 31 billion parameter dense architecture, pursuing the highest output quality, ranked 3rd globally on the open-source model Arena
  • 26B Mixture of Experts Model: 26 billion parameter MoE architecture, containing 128 small experts, activating only 3.8 billion parameters per inference, achieving extremely fast token generation speed

Edge Device Small Models:

  • E2B Model: 5.1 billion total parameters, only 2.3 billion effective parameters at runtime, memory footprint can be compressed to under 1.5GB
  • E4B Model: 8 billion total parameters, delivering 4.5 billion parameter effectiveness at runtime, balancing performance and power consumption

Core Capability Breakthroughs

Advanced Reasoning Capabilities: Gemma 4具备 multi-step planning and deep logical reasoning abilities, showing significant improvements in math and instruction-following benchmarks requiring complex reasoning. The 31B dense model scores 89.2% on the AIME 2026 math test and 80.0% on the LiveCodeBench programming test.

Intelligent Agent Workflow Support: All models natively support function calling, structured JSON output, and system instructions, allowing developers to build autonomous agents that can interact with various tools and APIs and stably execute workflows.

Multimodal Capabilities: The full model lineup natively supports video and image processing, supporting variable resolution (70 to 1120 image patches), and performs excellently in visual tasks such as OCR and chart understanding. The E2B and E4B models also feature native audio input capabilities for speech recognition and understanding.

Long Context Processing: Edge models support a 128K context window, and large models support up to 256K, enabling processing of entire codebases or long documents in a single prompt.

Major Changes in Open Source Licensing

One of Gemma 4's most important decisions is the change in licensing. Google has abandoned the previously controversial restrictive terms and fully adopted the Apache 2.0 license. This means enterprises can freely deploy, modify, and commercialize without worrying about Google unilaterally changing the rules.

Hugging Face co-founder and CEO Clément Delangue commented: "The release of Gemma 4 under the Apache 2.0 license is a significant milestone." This change allows many large companies that were previously stuck in legal review to safely use and fine-tune the models.

Hardware Adaptation and Deployment

Workstation Deployment: The 26B and 31B models are optimized for researchers to achieve top-tier reasoning capabilities on general-purpose hardware. Non-quantized bfloat16 weights can efficiently fit on a single 80GB NVIDIA H100 GPU. Quantized versions can run directly on consumer GPUs, supporting IDEs, programming assistants, and agent workflows.

On-device Deployment: The E2B and E4B models are designed from the ground up for extreme computational efficiency and memory efficiency. Google has deeply collaborated with mobile hardware manufacturers such as the Pixel team, Qualcomm, and MediaTek, enabling these multimodal models to run completely offline with near-zero latency on edge devices such as phones, Raspberry Pi, and NVIDIA Jetson Orin Nano.

Android developers can now build agent workflows in the AICore developer preview, achieving forward compatibility with Gemini Nano 4.

Performance and Industry Impact

On the industry-standard Arena AI text leaderboard, the 31B model ranks 3rd among global open-source models, and the 26B model ranks 6th. Gemma 4's performance even surpasses models 20 times its size.

For developers, this new height in intelligence per parameter means obtaining frontier model-level capabilities with minimal hardware overhead. One netizen gave the评价: "So small, so powerful."

Multilingual Support and Ecosystem Building

Gemma 4 is trained on over 140 languages,具备 multi-step planning and complex logical reasoning capabilities, can be used for building agents, and supports code generation. This multilingual capability helps developers build inclusive, high-performance applications for global users.

Looking back at the development of the Gemma series, from Gemma 1 to Gemma 3, developers have downloaded over 400 million times, and the community has produced over 100,000 variants. Gemma 4's goal is clear: to enable developers to achieve near-frontier closed-source model intelligence levels on their own hardware, completely returning the choice to developers.

Conclusion

The release of Gemma 4 marks a new stage in the development of open-source AI models. It is no longer simply pursuing parameter scale through "brute-force aesthetics," but rather maximizing "performance per dollar" through architectural optimization. The combination of Apache 2.0 licensing, multi-specification coverage, and stunning performance demonstrates Google's strategic determination in the open-source market. With deeper community fine-tuning, Gemma 4 is expected to give rise to more innovative applications, promoting AI technology to serve all industries more inclusively.