October 6, 2026: Google DeepMind has published the model card for Nano Banana 2.1, a multimodal image generation and editing model based on Gemini 3.6 Flash. The documentation outlines its capabilities, evaluation results, deployment channels and known limitations.
Nano Banana 2.1 can process both text and image inputs and generate image and text outputs. It has a context window of up to 1 million tokens, while its outputs support up to 4K tokens for images and 64K tokens for text, according to Google DeepMind.
The model is part of the Gemini 3 series and is designed for image creation and editing tasks. Its stated use cases include generating images with detailed control, producing text for posters and diagrams, handling longer-context real-world information and rendering localised text across multiple languages.
Google DeepMind evaluated the model across text-to-image, image editing and multi-turn use cases. The testing included areas such as visual design, factuality, character and object editing, stylisation, multi-image editing and mask-based editing. Human side-by-side evaluations were used for preference and design assessments, while factuality was assessed through an automated evaluation system.
In the reported October 2026 results, the thinking version recorded an overall preference score of 1,050, compared with 990 for Gemini 3.1 Flash Image and 935 for Gemini 3 Pro Image. For editing, Nano Banana 2.1 recorded higher scores across the listed categories, including multi-character consistency, stylisation and multi-reference editing.
The model is available through the Gemini app, Google AI Studio, Gemini API, Google Search AI Mode, Google Ads, Google Flow and Google Stitch.
Google DeepMind also lists several limitations. These include occasional hallucinations, slow responses or timeouts, difficulties with small text and long paragraphs, imperfect character consistency, partial instruction following in some editing tasks and occasional confusion over spatial relationships such as left and right. The model also has limitations in advanced world knowledge, 3D reasoning and factuality.
source:Nano Banana 2.1 – Model Card — Google DeepMind
