Features of Gemma 3's 1B, 4B, 12B, and 27B Model Variants 196

Gemma 3 open-sourced four parameter versions: 1B, 4B, 12B, and 27B, all of which support running on a single GPU or TPU. In the LMArena leaderboard, Gemma 3's 27B parameter version achieved a high score of 1339 ELO, beating o1-preview, o3-mini high, and DeepSeek V3, becoming the best open-source model second only to DeepSeek R1. In addition, the 1B, 4B, 12B, and 27B versions of the Gemma 3 series were trained on 2T, 4T, 12T, and 14T tokens of data, respectively.

The following are the features of these 4 models of different sizes:

1. Gemma 3 - 1B

• Parameter scale: 1 billion parameters.

• Features: Lightweight, suitable for running on resource-constrained devices such as mobile phones or laptops, instruction-optimized for rapid deployment and low-power scenarios.

2. Gemma 3 - 4B

• Parameter scale: 4 billion parameters.

• Features: Performs excellently on multimodal tasks (text, images), suitable for scenarios requiring higher accuracy, supports high-resolution image processing and complex task reasoning.

3. Gemma 3 - 12B

• Parameter scale: 12 billion parameters.

• Features: Has stronger multimodal processing capabilities, suitable for complex image and video analysis tasks, instruction fine-tuned (Instruction Fine-Tuning), suitable for applications requiring high accuracy and complex logic.

4. Gemma 3 - 27B

• Parameter scale: 27 billion parameters.

• Features: This is the largest model in the Gemma 3 series, with the strongest performance, suitable for high-performance computing scenarios, performing excellently in multiple benchmarks, approaching the performance of Gemini-1.5-Pro.

The four sized models of Gemma 3 cover the needs from lightweight to high-performance, and developers can choose the appropriate version based on specific hardware conditions and task requirements.