A Step Ahead in AI performance and efficiency
Subrat Panda, PhD
CTPO @ AGNEXT | AI Expert | PhD in Computer Science | IIT KGP | Ex - Capillary |Quality Food for Billions
Google's new open-weight model, Gemma 3, is a big step forward in AI, offering efficiency, flexibility, and competitive performance. As an AI professional, I view this launch as a strategic step in the ongoing competition for smaller but efficient models that can run with minimal computational power.
Major improvements in Gemma 3
In contrast to its predecessor, Gemma 2, this latest version provides parameter sizes of 1B, 4B, 12B, and 27B, which allows it to be versatile for a wide range of applications.
Its most notable feature is an increased 128K token context window, which enables it to process and hold more information in a single session. This makes it especially efficient in handling long content, coding work, and sophisticated reasoning.
Moreover, Gemma 3 is multimodal, with the capability to analyze text, images, and short videos. It natively supports 35 languages with pre-trained compatibility support for 140 languages, further enhancing its universal usability.
Performance in Benchmarks
In Chatbot Arena, where AI models are evaluated side-by-side by human testers, Gemma 3 (27B) outperformed OpenAI’s o3-mini, DeepSeek-V3, and Meta’s Llama 3-405B. It also delivered strong results in standardized AI benchmarks:
MMLU-Pro (67.5%) and GPQA Diamond (42.4%), surpassing Claude 3.5 Haiku (63% and 41%) and closely competing with GPT-4o Mini (65% and 43%).
Meta's Llama 3 70B is still the top performer (71% MMLU-Pro, 50% GPQA Diamond).
Efficiency
The most thrilling discovery may be Gemma 3's efficiency. It attained these levels of performance on one NVIDIA H100 GPU, while other models needed a maximum of 32 GPUs. Google has optimized the use of KV-cache memory, making it more efficient when processing longer contexts.
Access and availability
Gemma 3 is accessible through Google AI Studio, the GenAI SDK, and deployed locally by Hugging Face, Ollama, and Kaggle. Google also introduced ShieldGemma 2, a 4B parameter image safety model to identify unsafe content.
With Gemma 3, Google is propelling AI towards a future where power is matched with efficiency—raising the bar in the industry.