Gemma 3n Model Series

3 posts · Jun 2025 – Oct 2026

What this trail covers

Gemma 4 is released, expanding the Gemma model series with new sizes (E2B, E4B, 26B MoE, 31B Dense) optimized for advanced reasoning, agentic workflows, and on-device AI. These models offer improved math and instruction-following benchmarks, native support for function-calling and structured JSON output, enhanced code generation, multimodal processing (vision and audio), and longer context windows (up to 256K). The E2B and E4B models are designed for mobile and IoT devices, running offline with EmbeddingGemma 2 providing multimodal embedding capabilities.

All DeepMind trails

Oct 2026EmbeddingGemma 2: an open, lightweight multimodal embedding modelIntroduces EmbeddingGemma 2, a lightweight, multimodal embedding model built on Gemma 4 architecture. It expands beyond text to unify code, images, video, and audio in a shared embedding space, optimized for on-device inference. Key features include best-in-class performance for its size, modular design, storage efficiency via Matryoshka Representation Learning (MRL), optimized on-device performance with low RAM usage, and an 8K token context window. It achieves significant improvements in code performance and enables on-device RAG pipelines when paired with generative models like Gemma 4.
Jun 2026Introducing Gemma 4 12B: a unified, encoder-free multimodal modelIntroduces Gemma 4 12B, a new mid-sized multimodal model. Key technical contributions include an encoder-free architecture for vision and audio inputs, where inputs are projected directly into the LLM backbone. Vision processing is handled by a lightweight embedding module (single matrix multiplication, positional embedding, and normalizations), and raw audio signals are projected into the same dimensional space as text tokens. This design aims to reduce latency and memory usage compared to traditional multimodal models with separate encoders. The model achieves benchmark performance close to the 26B MoE model and is optimized for local execution on hardware with 16GB of VRAM.
Apr 2026Gemma 4: Byte for byte, the most capable open modelsThis post introduces Gemma 4, an evolution of the Gemma model series. Key technical contributions include the release of four new model sizes (E2B, E4B, 26B MoE, 31B Dense) with enhanced reasoning capabilities and native support for agentic workflows (function-calling, structured JSON). It highlights multimodal processing (vision and audio) and significantly increased context windows (up to 256K). The E2B and E4B models are specifically engineered for edge devices, emphasizing offline, low-latency performance. The post details optimizations for diverse hardware, including consumer GPUs and mobile platforms, and reiterates the Apache 2.0 license for broad accessibility.
Jul 2025T5Gemma: A new collection of encoder-decoder Gemma models- Google Developers BlogThis post introduces T5Gemma, a new collection of encoder-decoder Gemma models. These models are built on the Gemma architecture and are designed for a wide range of tasks, including text generation, summarization, translation, and question answering. T5Gemma models offer improved performance and efficiency compared to previous encoder-decoder models, making them suitable for both research and production environments. They are available in various sizes to accommodate different computational resources and performance requirements.
Jun 2025Introducing Gemma 3n: The developer guide- Google Developers BlogThis post introduces Gemma 3n, an expansion of the Gemma model series with new sizes (E2B, E4B, 26B MoE, 31B Dense) optimized for advanced reasoning, agentic workflows, and on-device AI. Key improvements include enhanced math and instruction-following benchmarks, native support for function-calling and structured JSON output, improved code generation, multimodal processing (vision and audio), and longer context windows (up to 256K). The E2B and E4B models are specifically designed for mobile and IoT devices, enabling offline operation.
Advertisement