# Gemma 3 4B ![Organization](https://img.shields.io/badge/Org-Google%20DeepMind-4285F4) ![Parameters](https://img.shields.io/badge/Params-4B-1f6feb) ![Context](https://img.shields.io/badge/Context-128K%20tokens-1f6feb) ![License](https://img.shields.io/badge/License-Gemma%20ToU-2da44e) ![Modality](https://img.shields.io/badge/Modality-Text%20%2B%20Vision-8957e5) ![Weights](https://img.shields.io/badge/Weights-Open-2da44e) ![Released](https://img.shields.io/badge/Released-Mar%202025-555) > **Quick answer:** Gemma 3 4B (released March 12, 2025) is Google DeepMind's 4-billion-parameter open-weight model with a 128K-token context window and native vision capabilities. It ranked #1 on the LMSys Chatbot Arena overall and at its size tier at launch, and runs on a single GPU or modern laptop — making it the most capable open model for edge and local deployment at 4B scale. ## At a Glance **Where Gemma 3 4B leads** - #1 single-GPU model on LMSys Chatbot Arena at launch - Native multimodality: understands images alongside text in a single model - 128K context window — the longest available in a 4B open-weight model at release - Runs on a laptop GPU (6–8GB VRAM at INT4), enabling private on-device deployment **Where it lags** - 4B parameters limits peak reasoning depth for hard math, coding, and multi-step agent tasks - Gemma Terms of Use apply (not pure Apache 2.0) — review restrictions for commercial deployment at scale **Best for:** edge deployment, mobile/on-device AI, private local inference, vision-enabled chatbots, and low-resource environments where cloud API costs are prohibitive. ## What Gemma 3 4B Is Gemma 3 is Google DeepMind's third generation of open-weight models, released March 12, 2025. The 4B variant is designed for on-device and edge deployment — capable enough for real tasks but small enough to run on consumer hardware. At launch it supported 35 languages and introduced vision understanding to the Gemma line, with the same architecture as the larger 12B and 27B siblings. The architecture uses grouped query attention, rotary position embeddings, and a sliding window attention pattern — a combination that maintains long-context capability while keeping inference memory requirements practical. ## Specifications | Field | Value | |---|---| | Organization | Google DeepMind | | Parameters | 4 billion | | Architecture | Transformer decoder (grouped query attention) | | Context window | 128,000 tokens | | License | [Gemma Terms of Use](https://ai.google.dev/gemma/terms) | | Release date | March 12, 2025 | | Knowledge cutoff | September 2024 | | Modality | Multimodal (text and vision) | | Min VRAM (INT4) | ~3–4GB | | HuggingFace | [google/gemma-3-4b-it](https://huggingface.co/google/gemma-3-4b-it) | ## Gemma 3 4B vs Alternatives | Model | Params | Context | Multimodal | License | |---|---|---|---|---| | **Gemma 3 4B** | 4B | 128K | Yes | Gemma ToU | | [Gemma 3 12B](/models/google/gemma-3-12b) | 12B | 128K | Yes | Gemma ToU | | Llama 3.2 3B | 3B | 128K | No | Llama 3.2 | | Phi-4 Mini | 3.8B | 128K | Yes | MIT | ## Related Models & Benchmarks - Compare: [Gemma 3 12B](/models/google/gemma-3-12b) · [Gemma 3 27B](/models/google/gemma-3-27b) · [Gemma 3 1B](/models/google/gemma-3-1b) - Benchmarks: [MMLU-Pro](/benchmarks/mmlu-pro) ## Frequently Asked Questions

Can Gemma 3 4B run locally?

Yes — at INT4 quantisation it requires ~3–4GB VRAM, running on most modern laptop GPUs and Apple Silicon Macs via llama.cpp or Ollama.

Does Gemma 3 4B support images?

Yes — Gemma 3 introduced native vision understanding across all size variants (1B, 4B, 12B, 27B).

--- *Specs sourced from [Google — Gemma 3](https://deepmind.google/models/gemma/gemma-3/). Last updated 2026-06-19.* --- # Gemma 3 4B Gemma 3 4B is a large language model developed by Google.

Gemma 3 4B — Files