Musk’s xAI announces Grok 3 open for free use “until the servers crash”

Musk’s AI company xAI has announced that the Grok 3 model is now available for free (until the server crashes). Subscribers to the X Premium+ and SuperGrok services will get deep access to Grok 3, in addition to priority access to premium features such as Voice Mode.

Musk’s smartest Grok 3 debuts, claims to outperform ChatGPT & DeepSeek

At noon on February 18, the new generation of large models from Musk’s xAI company, Grok 3, was officially released. The launch took the form of a live video broadcast with more than 1 million viewers, and Musk himself appeared in the broadcast room.

　　According to Musk, in a very short time Grok3 has proven to be an order of magnitude more powerful than Grok2, and Grok-3 and Grok-3 mini surpass or rival rivals such as Gemini, DeepSeek, and ChatGPT in terms of multifaceted performance.

　　For their part, xAI engineers say that while Grok started late, it caught up with ChatGPT in MMLU scores at a super-fast rate, Grok 3 involves 10 times more training than Grok 2, and Grok 3 has now been running internally at xAI for 2 weeks. MMLU is a benchmark test for evaluating the ability of a language model to understand language in a variety of tasks.

In addition, xAI unveiled the Grok 3 intelligent search engine called Deepsearch. xAI engineers emphasized that Deepsearch is the first generation of xAI’s broad range of agent tools that not only help engineers, researchers, and scientists write code, but also help everyone answer the questions they face every day.

　　Musk demonstrated Grok 3’s search capabilities, such as asking “when is the next spaceship launch?” Grok 3 will be like DeepSeek, showing what the model is doing, including browsing web pages, thinking process, and confirming whether the information on the web page is credible, and then coming to a relevant conclusion. Ultimately, the big model predicts the next spacecraft launch period in list form.

　　Notably, the search engine also displays various subtasks during the deep thinking process, a feature that is said to be more powerful and time-saving than current search engines due to the redesign of the search engine’s algorithms and logic.

As for when users will be able to experience the “smartest AI on the planet,” Musk said it will be available to scheduled users first. As the model is still being updated and improved, the Grok version on the Apple App Store will be relatively outdated, while the web version will be the most up-to-date. According to him, the Grok 3 voice mode could be launched in a week.

　　Musk also revealed for the first time the cost of Grok3 training, saying that the Grok3 training process cumulatively consumes 200,000 NVIDIA GPUs, and that the training is done in xAI’s data center.

　　At the end of the live broadcast, Musk and xAI engineers answered questions from netizens on the spot. When asked if it will be open source, Musk answered in the affirmative, saying that every time he makes the next version, he will open source the previous version, and in a few months he will open source Grok 3. On the sensitive question of whether Grok 3 is a male or female, Musk said that the gender is defined by the model itself, and also warned “Don’t fall in love with Grok”.

　　Talking about the hardest part of developing Grok 3, Musk said it was training the model and the 100% logical reasoning process, the kind of complexity that’s like predicting the universe anytime, anywhere.

　If Grok 3’s reasoning is as powerful as the video makes it out to be, it could push AI from “generating answers” to “logical deduction.”

　　Founded in July 2023, xAI has previously released Grok-1, Grok-1.5, Grok-1.5 Vision, Grok-2, and other large-scale models. Among them, Grok-1 is the largest open-source language large model with the largest number of participants to date; Grok-1.5 has improved reasoning ability and 128k context length; Grok-1.5 Vision is the first multimodal intelligence model; and Grok-2 is ranked 18th in the Chatbot Arena list.

　　However, the current competitive landscape of the large-scale modeling industry is incredibly fierce. On the day of Grok 3’s release, Ottoman said on X, “For demanding testers, trying out GPT-4.5 brought a feel-good AGI experience that was far more profound than I expected!” This seems to indicate that GPT-4.5 has entered the testing phase and is not too far from an official release.

　　And in early February, Google also released a number of new models, including the Gemini 2.0 Flash, Gemini 2.0 Flash-Lite, and the experimental version of the new generation of its flagship large model, the Gemini 2.0 Pro. It remains to be seen whether Grok 3 will be able to dominate its competitors.

　　What is certain is that AI giants, including OpenAI and Google, have recently felt the shockwaves from DeepSeek, a Chinese big model company, and OpenAI, which used to stick to closed source, has shown signs of further open-sourcing. On February 18, Altman launched a poll on X asking people which big model they would like to see OpenAI open-source next? Is it an o3 mini or a big model that can run on a cell phone? Or is it a warm-up for open-sourcing a specific big model?