Google appears to have opened the attack on DeepSeek by introducing its Gemma 3 portfolio of lightweight LLMs. These new LLMs from the tech giant also run on a single GPU, allowing more devices to use AI functionality.
Google’s Gemma 3 LLMs, the new iteration in its already existing portfolio of lightweight Gemma LLMs, come in four variants: 1 billion parameters, 4 billion parameters, 12 billion parameters and 27 billion parameters. These LLMs, Google says, allow AI developers to choose the best model for their available hardware and performance when developing AI applications.
The Gemma 3 LLMs are designed to run quickly and directly on devices ranging from workstations to laptops and smartphones. For example, to run a Gemma 3 LLM on a device with a GPU or a TensorFlow processor, a larger model can be used than for a smartphone processor. The latter then uses a smaller model.
Google’s new lightweight LLMs, including the Gemini AI chatbot feature, use the same research work as the Google Gemini LLMs.
Performance
According to the tech giant, the new Gemma 3 LLMs’ performance is very high, surpassing that of OpenAI’s Llama-40B, DeepSeek-V3, and o3-mini, according to benchmarks. However, they still have to beat DeepSeek R1.
The new lightweight LLMs offer technical features such as understanding more than 140 languages, creating AI applications using multimodal capabilities such as advanced text and visual reasoning, and a context window of 128,000 tokens, which can handle large amounts of data.
Furthermore, the LLMs have function calling and other tooling for creating workflows and AI agents.
In addition, much work has also been done to make the Gemma 3 LLMs secure through comprehensive data governance and compliance with Google’s security rules through fine-tuning and robust benchmark evaluations.
Addition of ShieldGemma 2
In addition to the lightweight Gemma 3 models, Google now offers ShieldGemma 2. This is an LLM of 4 billion parameters for checking images for security. The model can also flag images that do not meet security rules, allowing developers to build applications that check uploaded images for potentially harmful content.
ShieldGemma 2 can be used with the new Gemma 3 LLMs without problems.
Google also adds that its existing Gemma ecosystem, Gemmaverse, now offers more than 60,000 community-created Gemma variants, including variants similar to OpenAI’s GPTs.
Also read: Google helps developers choose the best Gemini or Gemma model