Google releases Gemini 2.0 Pro: Multimodal, explosive programming capabilities, Jeff Dean exclaimed, now free and open
Google DeepMind's latest masterpiece, AI programming capabilities shocked the industry, and now it is open for free experience.
Core content:
1. Gemini 2.0 Pro new model released, AI programming capabilities praised by Jeff Dean
2. Gemini 2.0 Flash officially launched, low latency and high performance, suitable for developers to build applications
3. Gemini 2.0 Pro Experimental experimental model, supports 2 million token context windows, powerful reasoning ability
Google DeepMind just announced that Gemini 2.0 is now open to everyone! This release brings a series of updates and new models designed to provide developers and users with a more powerful, faster, and more cost-effective AI experience, with the focus on the launch of the new Gemini 2.0 Pro model
AI master Jeff Dean praised the programming capabilities of Gemini 2.0 Pro
Jeff Dean, chief scientist at Google and god of artificial intelligence, shared his amazement at the programming ability of Gemini 2.0 Pro. He said: "I love Boggle games (a kind of crossword puzzle). This demo shows the coding ability of our Gemini 2.0 Pro model in AI Studio. It is incredible that it can write a complete code with a relatively simple prompt, including all the correct data structures and search algorithms to find all valid words on the Boggle game board. As a computer scientist, I am also very happy that it completed the data structure correctly the first time." He also humorously used "Discombobulating!" (confusing/shocking) to describe
It seems that Gemini 2.0 Pro is powerful in code generation and can easily handle even complex algorithms and data structures.
Key Highlights:
Gemini 2.0 Flash is now generally available: Gemini 2.0 Flash, an experimental version released in December, is now generally available. This efficient model is known for its low latency and enhanced performance, making it an ideal choice for developers. Now, developers can build production-grade applications through the Gemini API in Google AI Studio and Vertex AI
Gemini 2.0 Pro Experimental released: This is an experimental new model optimized for code performance and complex prompts. It excels in code generation, understanding complex problems, and reasoning. It has a context window of up to 2 million tokens and can analyze and understand a large amount of information. Gemini Advanced users can experience this model in the Gemini application
Gemini 2.0 Flash-Lite is launched: This is the most cost-effective model in the Gemini 2.0 series. It provides higher quality while maintaining the same speed and cost as 1.5 Flash, and surpasses 1.5 Flash in most benchmarks. It also supports context windows of 1 million tokens and multimodal input
2.0 Flash Thinking Experimental is now available to Gemini app users: The 2.0 Flash Thinking Experimental model previously available in Google AI Studio is now also available to Gemini app users. Users can select it from the model drop-down menu on desktop and mobile.
Multimodal capabilities: All of these models will support multimodal input (text output) at launch, with full support for more modalities to come in the coming months
Model performance comparison:
Google has provided a detailed table comparing the performance of Gemini 1.5 Flash, 1.5 Pro, 2.0 Flash-Lite, 2.0 Flash, and 2.0 Pro Experimental in multiple benchmarks. Here are some key metrics:
| ability | Benchmarks | Gemini 1.5 Flash | Gemini 1.5 Pro | Gemini 2.0 Flash-Lite | Gemini 2.0 Flash | Gemini 2.0 Pro Experimental |
| General | MMLU-Pro | 67.3% | 75.8% | 71.6% | 77.6% | 79.1% |
| Code | LiveCodeBench (v5) | 30.7% | 34.2% | 28.9% | 34.5% | 36.0% |
| Bird-SQL (Dev) | 45.6% | 54.4% | 57.4% | 58.7% | 59.3% | |
| reasoning | GPQA (diamond) | 51.0% | 59.1% | 51.5% | 60.1% | 64.7% |
| Factual | SimpleQA | 8.6% | 24.9% | 21.7% | 29.9% | 44.3% |
| FACTS Grounding | 82.9% | 80.0% | 83.6% | 84.6% | 82.8% | |
| Multilingual | Global MMLU (Lite) | 73.7% | 80.8% | 78.2% | 83.4% | 86.5% |
| math | MATH | 77.9% | 86.5% | 86.8% | 90.9% | 91.8% |
| HiddenMath | 47.2% | 52.0% | 55.3% | 63.5% | 65.2% | |
| Long context | MRCR (1M) | 71.9% | 82.6% | 58.0% | 70.5% | 74.7% |
| image | MMMU | 62.3% | 65.9% | 68.0% | 71.7% | 72.7% |
| Audio | CoVoST2(21 lang) | 37.4 | 40.1 | 38.4 | 39.0 | 40.6 |
| video | EgoSchema (test) | 66.8% | 71.2% | 67.2% | 71.1% | 71.9% |
Last words
The focus of this release, Gemini 2.0 Pro, has not met my expectations from the existing information. I thought 2.0 Pro would be a killer feature of Google. As for whether the programming ability is stronger than OpenAI o3 mini high, Claude 3.5 sonnet, DeepSeek R1, further testing is needed.
But Google is still very conscientious. All models are free to use. Gemini 2.0 Pro allows 50 questions per day, and the others are all free for 1,500 times.
