Google Unveils Gemma 3n: Sub-10B Multimodal AI Model Tailored for Edge Devices

Details

• Google AI Developers has launched Gemma 3n, an open-weights multimodal language model specifically designed for phones, single-board computers, and other edge devices.
• The model understands 140 written languages and supports vision and audio capabilities in 35 languages, essentially doubling the linguistic range of February’s Gemma 2.
• Offered in 2B and 4B parameter versions, Gemma 3n is the first model under 10B parameters to achieve a 1,300+ score on the LM Arena benchmark.
• Featuring the MatFormer backbone and new Per-Layer Embeddings (PLEs), it reduces VRAM requirements by about 30 percent compared to models like Llama-3-8B and enhances reasoning, math, and code accuracy.
• Model checkpoints are accessible on Kaggle and HuggingFace, and Gemma 3n runs natively on popular platforms such as llama.cpp, Ollama, LM Studio, and Unsloth.
• Developers can fine-tune or prototype with Gemma 3n immediately in Google AI Studio, with Vertex AI API access and supplementary guides rolling out as well.

Impact

Google's Gemma 3n delivers cloud-level reasoning to edge devices, challenging the cost-effectiveness of server-bound GPT-3.5 models. With built-in vision support, it intensifies competition against rivals like Apple's upcoming "Apple Intelligence" and Meta’s Llama-3-8B. The model’s efficient architecture and open licensing could speed up multilingual AI adoption, pressuring incumbents and influencing regulatory debates around on-device AI in Europe.