Gemini (AI model)

Google introduced Gemini 1.0, developed collaboratively by Google DeepMind and other Google teams. Ultra targeted complex tasks, Pro general use, and Nano on-device deployment. Initial integrations included Pro in Bard and Nano in the Pixel 8 Pro; Ultra was still undergoing testing.⁠[1]

Google renamed its Bard conversational app Gemini and introduced Gemini Advanced with Ultra 1.0, alongside mobile experiences. Gemini consequently names both a model family and a user-facing application; the app provides access to models rather than being synonymous with the underlying model family.⁠[2]

Gemini 1.5 introduced a mixture-of-experts architecture that selectively activates relevant expert networks. Its first model, 1.5 Pro, offered a one-million-token context window in private preview to selected developers and enterprise customers. A ten-million-token research experiment was not a generally available product specification.⁠[3]

Google acknowledged inaccurate or offensive depictions of people in the Gemini app and paused that feature. It attributed the problem to overcompensating diversity adjustments and excessive refusals. The feature used Imagen 2, distinguishing this application incident from the Gemini foundation models themselves.⁠[4]

Google introduced Gemini 1.5 Flash, distilled from Pro for low-latency, high-volume tasks. Both 1.5 Pro and Flash entered public preview with one-million-token context windows, while a two-million-token window for Pro was offered to developers through a waitlist.⁠[5]

Google announced Imagen 3 for Gemini Apps and a forthcoming early-access rollout of people-image generation to Advanced, Business and Enterprise users, initially in English. The rollout retained restrictions on identifiable real people, minors, and violent or sexual content.⁠[6]

Gemini 2.0 emphasized tool use and agentic tasks. Experimental Flash gave developers multimodal input and text output, with search, code execution and custom function calls. Native image generation and speech synthesis were initially restricted to early-access partners. The accompanying Multimodal Live API supported real-time audio and video input.⁠[7]

Google released experimental Gemini 2.5 Pro, emphasizing reasoning before answering. It initially appeared in Google AI Studio and Gemini Advanced with a one-million-token context window; two million tokens remained a planned extension. Google highlighted improvements in complex problem-solving, coding and multimodal understanding.⁠[8]

Gemini 2.5 Pro and Flash became stable, generally available models, while 2.5 Flash-Lite entered preview. Flash-Lite targeted high-throughput tasks such as translation and classification, with adjustable thinking budgets, multimodal input, tool connections and a one-million-token context window.⁠[9]

Google DeepMind announced that an advanced research version of Gemini Deep Think solved five of six 2025 International Mathematical Olympiad problems, scoring 35 points at gold-medal standard. It answered in natural language within the 4.5-hour limit. IMO verification covered submitted solutions, not the system, process or underlying model.⁠[10]

Google launched Gemini 3 Pro in preview, emphasizing multimodal reasoning, coding and long-horizon planning across the Gemini app, Search AI Mode, AI Studio and Vertex AI. The enhanced Gemini 3 Deep Think reasoning mode initially went to safety testers ahead of a planned release to Ultra subscribers.⁠[11]

Gemini 3.1 Pro launched in preview with improved complex reasoning across developer, enterprise and consumer products. Google published an ARC-AGI-2 score of 77.1%, a result on a particular logical-reasoning benchmark rather than an accuracy rate for all real-world tasks.⁠[12]

Google released Gemini 3.5 Flash with stronger coding, tool use and multistep task execution across the Gemini app, Search and developer platforms. The announcement described 3.5 Pro as still undergoing internal testing, so Flash availability did not establish equivalent access to Pro.⁠[13]

Google introduced Gemini 3.6 Flash for general agentic workloads and 3.5 Flash-Lite for low-latency, high-throughput use. The accompanying 3.5 Flash Cyber was fine-tuned for vulnerability discovery and repair, with a limited CodeMender pilot planned for governments and trusted partners.⁠[14]

Gemini 3.7 Flash improved software engineering, document understanding and agentic workflows. It became available through developer and enterprise platforms and powered Gemini Spark, whose individual-user access was offered to Google AI Pro and Ultra subscribers in supported regions.⁠[15]

Google introduced Gemini 3.8 Flash and 3.8 Flash Cyber. Flash emphasized long-horizon coding and multistep reasoning, potentially using more reasoning steps, tool calls and tokens on difficult tasks. Cyber had more permissive cybersecurity safeguards and was restricted to trusted defenders through Fairwind.⁠[16]

Gemini 4 Argon expanded to paid API access, AI Studio and enterprise platforms, and to Pro and Ultra users of the Gemini app and Search AI Mode. It supported a unified one-million-token input-and-output window for complex coding, knowledge work and long-horizon tasks, with access conditions varying by product.⁠[17]

Google announced Gemini 4 Carbon, emphasizing coding, long-horizon work and cybersecurity. Its announcement described a phased rollout beginning with limited access for trusted cyber defenders and plans for broader access. The model announcement did not constitute availability to all users.⁠[18]

The Gemini 1.0 technical report describes a Transformer-decoder architecture trained on multilingual, multimodal data spanning web documents, books, code, images, audio and video. Its original Nano variants had 1.8 billion and 3.25 billion parameters, used distillation from larger models and were deployed with four-bit quantization; these figures do not describe every later Gemini model.⁠[19]

Google DeepMind’s Gemini model page links separately to the Gemini app, Google AI Studio and Google Antigravity. The model family, consumer application and developer tools serve different roles; supported inputs, outputs, context limits and regions depend on the model and product.⁠[20]

In explaining the 2024 image-generation issue, Google acknowledged that the Gemini app could produce errors, particularly around news, current events and sensitive topics. Its double-check feature does not guarantee correct answers. Model evaluation results likewise require attention to the version, task and testing conditions.⁠[4]

The Gemini Apps Privacy Hub describes collection of prompts, uploaded files, interaction records and connected-app information, and human review of some data to improve services. Processing depends on features, settings and applicable policies. This notice concerns Gemini Apps and is not a universal data policy for every enterprise deployment or API service.⁠[21]

Model website: https://deepmind.google/models/gemini/ (opens in a new window). Gemini app: https://gemini.google.com/ (opens in a new window). Developer access through Google AI Studio: https://aistudio.google.com/ (opens in a new window). Developer organization Google DeepMind on X: https://x.com/googledeepmind (opens in a new window).⁠[20]