View original
Nano Banana 2.1 - Model Card — Google DeepMind
Paper

Nano Banana 2.1 - Model Card — Google DeepMind

2026.10.07
·Web·by Mineru
#LLM#Multimodal#Image Generation#AI Model#Gemini

Key Points

  • 1Nano Banana 2.1 is a natively multimodal reasoning model based on Gemini 3.6 Flash, designed to handle both text and image inputs with a context window of up to 1 million tokens.
  • 2Performance evaluations as of October 2026 demonstrate that the "Thinking" variant of Nano Banana 2.1 consistently outperforms previous versions and the Gemini 3 Pro Image model across various text-to-image and editing benchmarks.
  • 3While the model provides advanced capabilities for infographic design and multi-character consistency, it retains specific limitations including potential hallucinations, imperfect text rendering, and challenges with spatial localization.

Nano Banana 2.1 is a specialized, natively multimodal reasoning model introduced in October 2026 as a member of the Gemini 3 series, specifically built upon the Gemini 3.6 Flash architecture. It is designed for high-precision text-to-image (T2I) generation and complex image editing tasks, featuring a 1M token context window for inputs, with output capabilities supporting 4K tokens for image generation and 64K tokens for text.

Core Methodology and Evaluation

The model employs a dual-stream evaluation strategy, comparing "Thinking" versus "No Thinking" variants across two primary categories: capability sets and regression sets. The core methodology relies on Side-by-Side (SxS) human evaluations to calculate Elo ratings for overall preference and design quality, supplemented by a Single-sided AutoRater to assess factuality.

Benchmarks are categorized into:

  • Text-to-Image (T2I): General generation, i18n text rendering, world-knowledge factuality, tool-use, and visual design.
  • Editing: Includes general editing, stylization, character/object/environment editing, mask/ink-based manipulation, and multi-image recontextualization.

Quantitative performance metrics indicate that the "Thinking" variant of Nano Banana 2.1 achieves superior Elo ratings across all domains compared to both its "No Thinking" counterpart and predecessor models. For instance, in Multi-Character Consistency, the model reaches an Elo of 1106±141106 \pm 14, significantly outperforming the Gemini 3 Pro Image baseline (1011±101011 \pm 10).

Technical Limitations and Safety

Despite its advancements, the model retains specific technical limitations characteristic of the Gemini 3 architecture:
  • Rendering: Suboptimal performance with small text, long paragraphs, or page-length text strings.
  • Consistency: Imperfections in character consistency between input/output and occasional failures in maintaining structural alignment, leading to unintended persistence of subject pose during edits.
  • Reasoning: Limited proficiency in 3D reasoning and spatial localization (e.g., left/right orientation).
  • Knowledge: The knowledge cutoff is bifurcated, with some domains updated to March 2026 (via Gemini 3.6 Flash) and others remaining constrained to January 2025.

Ethics and Frontier Safety

Safety is governed by the Frontier Safety Framework. Because Nano Banana 2.1 does not exhibit material performance increases over Gemini 3.1 Pro or Gemini 3.7 Flash, and given that those models did not reach Tracked or Critical Capability Levels (T/CCLs), the developers conclude that Nano Banana 2.1 is unlikely to pose novel frontier risks. The model has undergone rigorous manual red teaming, meeting established launch thresholds for child safety and demonstrating safety performance parity or improvement relative to the Gemini 3 Flash baseline. Distribution is facilitated through Google’s ecosystem, including Gemini App, Google AI Studio, and the Gemini API.