Google introduces Gemma 3 and claims that the model on just one GPU has 98% of DeepSeek accuracy

Google claims that Gemma 3 matches 98% of the DeepSeek R1's accuracy estimate, with a share of the estimated computing power of 1338 versus 1363 for the R1. To achieve the claimed result, DeepSeek R1 would need 32 Nvidia «H100» GPUs. Gemma 3 uses only one H100 GPU for this.

Citing previous human advantage estimates in the LMArena leaderboard, the company is confident that the new model outperforms the Llama-405B, DeepSeek-V3, and o3-mini.

Other advantages of Gemma 3 include:

• the model can be placed on a single GPU or TPU host;
• Gemma 3 offers out-of-the-box support for over 35 languages ​​and pre-trained support for over 140 languages;
• the model offers a 128,000-token context window that allows user programs to process and understand vast amounts of information;
• Gemma 3 has function calls and structured output to automate tasks and build agent experiences.

The Gemma 3 is available in sizes such as 1B, 4B, 12B and 27B, and that these models are their «most advanced, portable and responsibly designed open models.»

The model can be uploaded through platforms such as Kaggle and Hugging Face or even Google Studio.

Read the country’s main IT news in our Telegram

On the topic

Read the country’s main IT news in our Telegram

Google DeepMind introduces two new AI models specifically for robots

On the topic

Google DeepMind introduces two new AI models specifically for robots

In 2024, Google paid almost $12 million for bugs found

On the topic

In 2024, Google paid almost $12 million for bugs found

Former Google CEO warns against generative AI development. What he sees as the biggest threat, including the arms race

On the topic

Former Google CEO warns against generative AI development. What he sees as the biggest threat, including the arms race