Gemma 3n
2025-06-26 · Publisher : Google · Domain : Artificial Intelligence
Google releases the full version of Gemma 3n, a multimodal open model for mobile devices in effective sizes E2B and E4B, which accepts text, image, audio and video as input and exceeds 1,300 Elo on LMArena.
What changes
- Native inputs: image, audio, video and text, text output.
- Two effective sizes, E2B and E4B (5 and 8 billion raw parameters), for only 2 GB and 3 GB of memory.
- E4B: first model under 10 billion parameters to exceed 1,300 on LMArena, according to Google.
- MatFormer architecture (nested E2B sub-model), Per-Layer Embeddings, LAuReL, AltUp, MobileNet-v5 audio and vision encoders.
- 140 languages for text, multimodal understanding in 35 languages.
### Specifications - Availability: weights on Hugging Face and Kaggle, compatibility with Transformers, llama.cpp, Google AI Edge, Ollama, MLX. ### Pricing Open model, with no usage pricing of its own. ### Benchmarks | Model | Benchmark | Score | |---|---|---| | Gemma 3n E4B | LMArena (Elo) | over 1,300 | Score published by the vendor. ### Key takeaways A multimodal model able to run on a phone, useful for offline use cases or those subject to privacy constraints.
Links
Tags : Open Source, LLM, Google, Gemma, Multimodal