Google releases the Nano Banana 2.1 image model

Google DeepMind released Nano Banana 2.1, an image generation and editing model built on Gemini 3.6 Flash. The model card lists a 1-million-token input context window for text and images and an output budget of up to 4,000 tokens per image, alongside 64,000 tokens of text output.

Google said the “Thinking” variant of Nano Banana 2.1 outperformed its direct predecessor, Nano Banana 2 (built on Gemini 3.1 Flash Image), on its internal preference evaluations for both text-to-image generation and editing, with overall preference scores rising from 990 to 1050. The company highlighted precise image editing, generation of legible text for posters and diagrams, and localised text rendering across languages, along with drawing on Gemini’s general world knowledge for historically or factually accurate scenes. Google’s own documentation also noted persisting limitations, including poor rendering of small text and imperfect consistency of a character’s appearance between an input image and its edited output.

The model is distributed across the Gemini app, Google AI Studio, the Gemini API, Search’s AI Mode, Google Ads, Flow and Stitch — Google’s existing surfaces for the “Nano Banana” line since the nickname, originally applied informally to Gemini 2.5 Flash Image in August 2025, stuck and became the product’s common name.