Google DeepMind’s model card for Nano Banana 2.1 concedes that its image generator still struggles with small type, which often comes out blurry at the 1K output size, and with long paragraphs or whole pages of text. DeepMind published the card on 6 October 2026. A model card exists to state limitations and safety performance, so the weaknesses are the most useful part of it.
The list is longer than the text problem. The card says character consistency between an input picture and the generated result is not always perfect, and that masked or hand-drawn edits only partly follow instructions, with the ink sometimes persisting. In rare edits the subject keeps the pose of the original image. The model also gets left and right confused at times, and the card calls its 3D reasoning, world knowledge and factuality “still limited.” Hallucinations, slowness and occasional timeouts are noted too.
On identity, the card is specific. Nano Banana 2.1 belongs to the Gemini 3 series and is built on Gemini 3.6 Flash. It accepts text and images and returns images and text. Input can run to 1 million tokens, image output has a 4K token budget, and text output goes to 64K. All of that holds up against the card, with one wrinkle in the safety section covered below.
The card is thin by design. Architecture, training data, hardware, software, acceptable use, evaluation approach and safety policies each point the reader to the Gemini 3.6 Flash card. Knowledge is dated too: the base model’s cutoff is March 2026, and in some domains the card says users may find it knows nothing newer than January 2025.
Safety results come in prose, with no scores. Specialist red teams outside the development group tested the model, and DeepMind says it met the launch thresholds for child safety that Google’s own expert teams set. On content policies generally it reports similar or better results than Gemini 3 Flash, and no egregious concerns when compared with Gemini 3.1 Pro. The card also says jailbreak resistance is still being worked on.
The frontier-risk assessment was not run on this model directly. DeepMind relied on earlier tests of Gemini 3.7 Flash and, separately, Gemini 3.1 Pro, which reached none of its tracked or critical capability levels, and judged Nano Banana 2.1 unlikely to either because it adds no meaningful new capability. The card does not explain why the base model is 3.6 Flash while the reference model is 3.7 Flash.
Every number in the document comes from Google. Google ran the evaluations: human raters compared outputs side by side to produce Elo scores, and an automated rater scored factuality. On overall preference, the thinking mode of Nano Banana 2.1 scores 1050, against 1015 with thinking off, 990 for Nano Banana 2 and 935 for Nano Banana Pro. The biggest gain is in keeping several characters consistent: 1106 against 978 for Nano Banana 2. The card does not say how many raters took part or who they were.
The infographic factuality row deserves the closest look. The thinking mode scores 0.521, up from 0.328 without thinking and 0.179 for Nano Banana 2. The card never defines the scale. If it runs from zero to one, as the format suggests, the best result is roughly half.
Teams planning to ship generated infographics or text-heavy posters should budget for human checking of every output, because the strongest factuality figure the card reports comes from a test Google designed and scored.
Google DeepMind, “Nano Banana 2.1 Model Card,” published 6 October 2026.
