Google released Gemma 4, an open model family for reasoning and agentic workflows. 🚀

Key features:

- Four sizes: E2B/E4B for edge devices (phones, Raspberry Pi, Jetson Nano); 26B MoE and 31B Dense for single cloud GPUs or workstations.
- Enhanced multi-step reasoning, math, and instruction-following.
- Native function-calling, JSON output, and system instructions for agentic workflows.
- Integrated vision/video (OCR, charts, variable resolution) across all models.
- Audio input for speech recognition on E2B/E4B. 🎤
- High-quality offline code generation. 💻
- 140+ languages. 🌍
- 128K context (