Google has officially launched its next-generation open model series Gemma 4, calling it the smartest open model to date. The series inherits research from the flagship Gemini 3 and adopts a fully commercial-friendly Apache 2.0 license, allowing developers to build and deploy with full control over data and infrastructure.
Four Versions Covering Edge to Flagship
Gemma 4 comes in four sizes for different hardware and use cases. The lightest E2B (2B parameters) targets mobile devices and browsers; E4B (4B parameters) balances performance and efficiency, natively supporting vision and audio input. On the high-performance side, the 26B A4B uses a Mixture-of-Experts (MoE) architecture activating around 4B parameters during inference, slashing memory requirements so it runs smoothly on consumer hardware like a Mac Mini with 24GB RAM. The top-tier 31B dense model is the series' performance leader.
256K Context Window, Native Multimodal and Function Calling
In terms of specs, the large versions support up to 256K tokens of context, enabling processing of entire codebases or large documents. Besides native text and image support (E2B and E4B also handle audio), Gemma 4 boasts robust native function calling with structured JSON output, ideal for building autonomous agents. Training data covers over 140 languages, ensuring global applicability.
Benchmarks Impress, Matching Much Larger Models
Gemma 4 emphasizes high performance per parameter. According to the AI Arena open model leaderboard, Gemma-4-31B ranks third among open models, performing on par with the much larger Qwen3.5-397B despite being only one-tenth its size. On the graduate-level reasoning benchmark GPQA Diamond, the 31B model achieved 84.3%.
Developers can try Gemma 4 on Google AI Studio or download weights from Hugging Face, Ollama, and other platforms. The community has already released quantization versions for GPU optimization. Some developers noted that Gemma 4 still has room for improvement in real-world complex debugging scenarios. Nevertheless, this open-source release injects fresh momentum into local AI applications.