Gemini 3 (AI)

Gemini 3 is a family of multimodal large language models developed by Google DeepMind, released on November 18, 2025.^[1][2] It serves as the successor to the Gemini 2.5 series. The initial release comprised two primary variants: Gemini 3 Pro and Gemini 3 Pro DeepThink. Along with LLM models, Google also released a coding platform based on VS code, Antigravity.^[1] Gemini 3 Pro is positioned as Google's flagship model, designed for high-precision multimodal reasoning across text, code, images, audio, and video.^[3] The DeepThink variant is a specialized reasoning model that utilizes "System 2" cognitive architectures, enabling parallel processing and advanced logic for complex problem-solving.^[4]

Gemini 3 logo

The screenshot shows two models under model switcher; Gemini 2.5 Flash and Gemini 3 Pro (Fast and Thinking with 3 Pro respectively).

Development

It is reported that Google have been training Gemini 3 Pro from the start of the year, due to which its knowledge cutoff is January 2025.^[5]

Unlike a dense model where every parameter is active for every calculation, Gemini 3 Pro utilizes a Sparse Mixture-of-Experts (MoE) architecture. The model consists of numerous specialized "expert" neural networks.^[6] This allows the model to have a massive total parameter count (significantly higher than Gemini 1.0 Ultra) while maintaining manageable inference costs and latency, as only a fraction of parameters are activated per token.

The defining architectural feature of the Gemini 3 Pro DeepThink variant is its departure from linear Chain-of-Thought (CoT) processing. Instead of generating a single linear sequence of thoughts, the model constructs Directed Acyclic Graphs (DAGs) of reasoning.