Last Updated:
Google’s new lightweight AI model Gemma 3 is already making the news not only because of its architecture but its ability to beat GPT and DeepSeek.
The new Gemma 3 version is already outperforming GPT and DeepSeek
The Gemma 3 family of artificial intelligence (AI) models was released by Google. With text and visual reasoning capabilities, the new open-source models are the successor to the Gemma 2 series, which debuted in August 2024.
these models may be adjusted to provide further support for 140 languages, and they currently support over 35 languages. Interestingly, the tech behemoth asserted that these models are designed to operate on Google’s Tensor processing unit (TPU) or a single GPU.
What Is The Gemma 3 Series AI Models?
Google explains the same technology was used to create its flagship Gemini 2.0 model. The Gemma series is claimed to be known for its on-device performance and open-source nature. Google disclosed that the Gemma models have been used to develop over 60,000 variants and have been downloaded over 100 million times so far.
According to the tech giant, Gemma performs better on the LMArena leaderboard than OpenAI’s o3-mini AI models, Meta’s Llama-405B, and DeepSeek-V3. There are four sizes of these models: 1B, 4B, 12B, and 27B parameters. As per the company, they can be operated on a single GPU or TPU.
The Gemma 3 series can analyse text, photos, and short videos and has sophisticated language and visual reasoning capabilities. A 1,28,000 token context window is provided by the AI models. In addition, the models permit function calling, enabling developers to incorporate agentic features into the software and apps they design.
Google said that a thorough risk evaluation was used in the development of the AI models. The company asserted that benchmark assessments and fine-tuning were used to enforce internal safety policies. Furthermore, more competent models were used to test the models. As per the company, the model showed a minimal level of risk.
Along with the Gemma 3 series, the company also introduced ShieldGemma 2, a 4B parameter image safety checker that makes sure AI models don’t produce violent, risky, or sexually explicit content. The ability to further modify ShieldGemma to improve the safety parameters has also been made available to developers. You can download the Gemma 3 family of AI models on Kaggle or through Google’s Hugging Face listing.
- Location :
California, USA