“100% certain, people will fall in love with Grok!” Elon Musk said in a live stream of his xAI model release on February 18th, Taiwan time.
After announcing the previous day on social media that the upcoming Grok-3 would be the “most powerful model on Earth,” the official live stream attracted over 2 million viewers.
During the live stream, the xAI team showcased Grok-3 and the smaller parameter version, Grok-3 mini, comparing them to Google’s Gemini, OpenAI’s ChatGPT, Anthropic’s Claude, and the Chinese startup DeepSeek’s DeepSearch. In terms of mathematics, science, and computer engineering, the performance of Grok-3 and Grok-3 mini outperformed the other existing models.
Grok-3 outperformed benchmark models in mathematics, science, and coding.
Image/xAI
On the model blind test platform Chatbot Arena (formerly LMSYS), the early Grok-3 model, code-named Chocolate, received a high rating of 1,400 in comprehensive question answering, surpassing Gemini 2.0 and Chatgpt-4o. Musk emphasized that Grok-3’s computational capabilities are over 10 times that of the previous generation model and will continue to dynamically adjust after deployment. “This model will improve every day!”
The latest three modes of Grok-3: DeepSearch, Think, and Big Brain
One of the goals of xAI is to solve complex problems, including Musk’s primary concern of space travel, as well as scientific problems requiring extensive data analysis and complex calculations. In addition to Google, OpenAI, and Perplexity, which originated as an AI search engine and launched the “Deep Research” feature specializing in scientific research, xAI has added the “DeepSearch” mode to Grok-3, allowing comprehensive research, thinking, and presentation of analysis results for complex problems.
Musk stated that “DeepSearch” can be considered the “next generation search engine” that allows AI to complete research that previously took hours in just 10 minutes. Particularly, by clicking the “Show Thinking” feature, users can see the complete process of how AI understands and processes problems, changing the previous passive reception of AI-generated content, making the content traceable and transparent.
Grok-3 includes the DeepSearch mode for deep research.
Image/xAI
Think mode: Proficient in reasoning and physics problems
In addition to the above, the Grok-3 model also includes the “Think” and “Big Brain” modes. During the live stream, xAI engineers stated that the “Think” mode excels in handling high-level reasoning and physics problems, such as having Grok-3 write programs to calculate celestial body movements.
Grok-3’s Think mode can also adjust thinking speed.
Image/xAI
Big Brain mode: Proficient in abstract creation
As for the third mode, “Big Brain,” Musk referred to it as the “starting point of AI creativity” and it is capable of engaging in relatively abstract creations. For example, during the live stream, the team demonstrated Grok-3 generating a new game that combines Tetris and color block elimination.
Grok-3’s “Big Brain” mode can harness creativity.
Image/xAI
The xAI team also mentioned that they will soon develop a voice-based chatbot for Grok-3. Currently, Grok-3 has been released to X’s Premium+ subscribers, and xAI is also planning to launch a new subscription plan called “SuperGrok” for the application and web versions, which includes Grok-3, “DeepSearch,” and “Think” functionalities.
Image/xAI
Musk stated that Grok-3’s API is expected to be made public in a few weeks, and following the usual practice, xAI will open source the previous generation model, Grok-2, after the latest model stabilizes, which is expected to happen within the next few months.
This article is licensed and reproduced from “Digitimes.”