GOOG

<p>Google has launched the newly upgraded multimodal AI model Gemini 3, which uses a sparse Mixture-of-Experts Transformer architecture and supports multimodal inputs such as text, images, and audio. The model is trained from scratch using Google TPU Pod and JAX, and has the capability to handle ultra-long context. Although it still lags behind ChatGPT 5.1 in deep reasoning, it can meet the AI needs of most users</p>

- Google has launched the upgraded multimodal model Gemini 3, which enhances user interface performance and meets most AI needs despite some gaps compared to ChatGPT 5.1.  
- The model utilizes a Sparse Mixture-of-Experts architecture, supports various input modalities, and can handle extremely long contexts, trained entirely on Google’s TPU infrastructure.  
- While Gemini 3 excels in factual accuracy and reasoning benchmarks, it still struggles with hallucinations, often providing confident but incorrect answers when uncertain.

How was Gemini 3 trained without NVIDIA?