Gemini 1.0
Google's first Gemini models: multimodal AI in Ultra, Pro and Nano sizes
Gemini 1.0 is the first generation of Google's Gemini family of multimodal models, announced in December 2023 by Google DeepMind. Built in three sizes (Ultra, Pro and Nano), it was trained to handle text, code, audio, images and video together, and the company says Ultra exceeded prior results on most academic benchmarks it tested.
- First Gemini generation: Ultra, Pro and Nano sizes
- Natively multimodal: text, code, audio, image and video
- Google says Ultra scored 90.0% on MMLU, above human experts
- Reported best on 30 of 32 academic benchmarks tested
- Specialised Gemini powers AlphaCode 2 for competitive programming
What it is
Gemini 1.0 is the opening release of Google's Gemini model family, announced in December 2023 by Google DeepMind. It comes in three sizes: Ultra for very complex work, Pro for a broad range of tasks, and Nano for running on devices such as phones. The models were built to work across text, code, audio, images and video at once, rather than bolting separate systems together. Google positions Gemini 1.0 as its most capable and most general model to that point, aimed at developers, enterprises and everyday users across its products.
How it came about
Google DeepMind was formed earlier in 2023, and Gemini 1.0 is described as the first fruit of that merger, produced by teams across Google including Google Research. Sundar Pichai framed it as the next step in the company's shift to being AI-first. Demis Hassabis, DeepMind's CEO, said the goal was a model that behaves less like software and more like an intuitive expert helper. The technical report was posted to arXiv on 19 December 2023 and later revised.
Inside the model
Rather than training separate components per modality and stitching them together, Gemini was pre-trained from the start on several kinds of input, then fine-tuned with additional multimodal data. Google says this lets it reason about mixed inputs natively. It handles text, images, audio and video simultaneously, and generates code in popular languages such as Python, Java, C++ and Go. The company reports Ultra results on 32 academic benchmarks, and used a new MMLU approach that lets the model think before answering. A specialised version powers AlphaCode 2 for competitive programming.
What it can do
Google says Gemini Ultra beat prior best results on 30 of the 32 academic benchmarks it tested, scoring 90.0% on MMLU (57 subjects including maths, physics, law and ethics) and 59.4% on the multimodal MMMU benchmark. On image tasks it reportedly led without using OCR to lift text first. It is described as strong at explaining reasoning in maths and physics, and at reading and filtering large document sets. AlphaCode 2, built on a specialised Gemini, solved nearly twice as many competition problems as AlphaCode and is estimated to outperform about 85% of contestants.
How to use it
Gemini 1.0 is offered in three sizes for different deployment targets: Ultra for complex tasks, Pro for scaling across many tasks, and Nano for on-device use. Google says it runs efficiently from data centres down to mobile devices, and is aimed at developers and enterprise customers building with AI. The models are proprietary; no open weights are described.
Safety and testing
Google says it approached the release both boldly and responsibly, building in safeguards and working with governments and outside experts on risks as models grow more capable. The work is guided by the company's published AI Principles. Rigorous testing and evaluation across a wide range of tasks is described, though specific red-teaming or incident details are not given here.
Reception and debate
Google's own announcement presents Gemini 1.0 as its largest and most capable model, with leading results across many benchmarks.
Why it matters
Gemini 1.0 opens what Google calls the Gemini era and is presented as one of the company's biggest science and engineering efforts. Its specialised variant underpins AlphaCode 2, which roughly doubled the problems solved by the earlier AlphaCode system in competitive programming. The company frames the release as groundwork for more capable models and for AI features across its products.
Questions people ask about Gemini 1.0
What is Gemini 1.0?
It is the first version of Google's Gemini family of multimodal models, announced in December 2023 by Google DeepMind. It ships in three sizes, Ultra for complex tasks, Pro for general use, and Nano for on-device work, and handles text, images, audio, video and code.
Who made it?
Google DeepMind built it with help from teams across Google, including Google Research. Demis Hassabis, DeepMind's CEO, presented it, with a note from Google and Alphabet CEO Sundar Pichai. It is the first model family produced since Google DeepMind was formed in 2023.
How is it different from earlier multimodal models?
Earlier systems often trained separate parts for each input type and joined them afterwards. Gemini was pre-trained on multiple modalities from the start, then fine-tuned with more multimodal data, which Google says gives it stronger reasoning across mixed inputs.
How well does it perform?
Google reports that Gemini Ultra exceeded prior best results on 30 of the 32 academic benchmarks it tested, including 90.0% on MMLU and 59.4% on MMMU. These are the company's own reported figures from its announcement and technical report.
Can I download it?
No open weights are described. Gemini 1.0 is proprietary and offered through Google's products and services, with the Nano size intended for on-device use and the larger sizes aimed at developers and enterprises.
Good reading and sources
There is no English Wikipedia article on Gemini 1.0 yet, so the facts on this page come from the sources listed above: blog.google/technology/ai/google-gemini-ai/, arxiv.org/abs/2312.11805. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.