Google DeepMind publishes the Nano Banana 2.1 model card: built on Gemini 3.6 Flash with a 1050 overall preference score

The specification section is fairly concrete. Nano Banana 2.1 is based on Gemini 3.6 Flash, and for architecture, training dataset, training data processing, hardware and software the card points directly to the Gemini 3.6 Flash model card rather than repeating the details. Inputs are text strings, such as a prompt or documents, plus images, with a token context window of up to 1M. Outputs are image with a 4K token output, and text with a 64K token output. Those figures determine how much context it can take in and how large an image it can put out, and they are the first two things engineering teams look at.

Seven distribution channels are listed, which is worth noting compared with models announced only through an API: the Gemini App, Google AI Studio, Gemini API, Google Search AI Mode, Google Ads, Google Flow and Google Stitch. The card also states that models are available to downstream providers via an application programming interface and subject to relevant terms of use, and that there is no required hardware or software to use the model. Taken together, image generation shows up at the same time in consumer apps, developer tooling, search, advertising and design tools, a much wider footprint than a standalone image model.

The evaluation section gives both method and results. Nano Banana 2.1 was assessed with two broad categories. Capability sets cover diverse public and internal benchmarks for text-to-image and editing: on text-to-image, general text-to-image, i18n text rendering, factuality, tool use and visual design; on editing, general image editing, stylization, character editing, object and environment editing, ink-based doodle editing and multi-image work; and multi-turn covering both static and dynamic conversations. Regression sets draw on popular use cases seen on Gemini 2.5 Flash Image and Gemini 3 Pro Image to make sure the new model shows no noticeable regressions. The methodology is side-by-side human evaluation to derive Elo, with a single-sided AutoRater for factuality.

Among the results as of October 2026, the most legible figure is text-to-image overall preference: Nano Banana 2.1 scored 1050 +/- 14 in Thinking mode and 1015 +/- 13 without thinking, against Gemini 3.1 Flash Image (Nano Banana 2) at 990 +/- 7 and Gemini 3 Pro Image (Nano Banana Pro) at 935 +/- 8. The ordering across the four numbers is clear: both 2.1 modes sit above the two earlier generations. The card stresses these are Elo figures from side-by-side human evaluation, and the right way to read that kind of metric is as a relative ranking, not an absolute quality score.

Frontier safety is the other solid block in the card. Nano Banana 2.1 belongs to the Gemini 3 family, and to assess it for frontier safety Google DeepMind relied on its evaluations of Gemini 3.1 Pro and Gemini 3.7 Flash; the finding was that neither model reached any Tracked or Critical Capability Levels, or T/CCLs. Because the assessments showed Nano Banana 2.1 does not have meaningful new capabilities or material increases in performance versus Gemini 3.1 Pro or Gemini 3.7 Flash, DeepMind says it is confident Nano Banana 2.1 is also unlikely to reach any T/CCLs. That should be read exactly as stated: it is an official judgment resting on an existing evaluation chain, with nothing extrapolated here.

The structure of the card itself is worth a note, because it determines how the document should be read. Google DeepMind lists these sections: Model Information, Model Data, Implementation and Sustainability, Distribution, Evaluation, Intended Usage and Limitations, and Ethics and Content Safety. The company defines model cards as providing essential information on Gemini models, including known limitations, mitigation approaches and safety performance. That makes this a technical reference for engineering and compliance teams rather than a launch announcement, and it reads at a very different pace from a press release.

The capability sets are broken down finely, which helps in judging what the model is suited to. On text-to-image the card lists general text-to-image, i18n text rendering, factuality including world knowledge and educational content, tool use and visual design. On editing it lists general image editing, stylization, character editing, object and environment editing, ink-based doodle editing, and multi-image scenarios such as multi-product recontextualization and multi-character work. Multi-turn covers both static and dynamic conversations. Regression sets serve a different purpose: they compare against popular use cases seen on Gemini 2.5 Flash Image and Gemini 3 Pro Image to confirm the new model shows no noticeable regressions.