On 12th March 2025, Google unveiled its open-source artificial intelligence model family called Gemma 3, which utilises technology based on the Gemini 2.0 models and delivers exceptional performance whilst running on just a single GPU or TPU device. The model's outstanding capabilities include support for 140 languages, a 128,000 token context window, and the fact that it offers its multimodal functionalities—including text and image analysis—in four different sizes (1B, 4B, 12B, and 27B parameters).
Gemma 3's largest variant, with 27 billion parameters, achieved a remarkable Elo score of 1338 in the user preference measurements on LMArena evaluations, which represents 98% of the competing DeepSeek R1 model's score of 1363. The performance difference is marginal; however, there is an enormous disparity in hardware requirements: whilst DeepSeek necessitates 32 Nvidia H100 graphics cards to operate, Gemma 3 functions fully with merely a single GPU. The efficiency of Google's model is ensured by an artificial intelligence technique called "distillation," and three specialised quality control procedures further enhance its mathematical and programming capabilities.
Gemma 3 represents a significant advancement compared to its predecessor, Gemma 2, which had only an 8,000 token context window and exclusively English language support. The new model has also undergone leak testing, which demonstrated that Gemma 3 models memorise long texts at a much lower rate than previous models, thereby reducing the risk of information leakage. An additional advantage of the Google-developed tool is that it supports Hugging Face Transformers, Ollama, JAX, Keras, PyTorch, and other popular tools, whilst providing immediate access through Google AI Studio, Kaggle, or Hugging Face platforms.
Sources:
1.
Clement Farabet
2.
3.
Tiernan Ray
