On Monday, the Grok 3 family of artificial intelligence (AI) models was officially unveiled. This new series, introduced during a live event hosted by xAI engineers and Elon Musk, represents the evolution of the Grok 2 models. The Grok 3 lineup consists of multiple large language models (LLMs), each varying in parameter sizes and designed for advanced reasoning capabilities. Users can expect features like DeepSearch and a voice response mode, alongside the introduction of a new subscription tier called SuperGrok for elevated access and limit enhancements.
Grok 3 AI Models Unveiled by xAI
During a live stream on X (formerly Twitter), Musk demonstrated the advanced functionalities of the Grok 3 models, asserting that they are “an order of magnitude more capable than Grok 2.” He noted that xAI established a new data center specifically for pre-training these LLMs, with the initial 100,000 GPUs coming online in just 122 days. Capacity was subsequently increased within the following 92 days.
The Grok 3 family features a variety of LLMs, though not all are immediately accessible. The event highlighted models including Grok 3 and Grok 3 mini, a streamlined but speedier version, as well as Grok 3 Reasoning and Grok 3 mini Reasoning, which focus on compute-based reasoning. While the rollout for these models is underway, additional variants remain in beta testing.
Musk pointed out that some chain-of-thought (CoT) steps in the Grok 3 reasoning models will be obscured in an effort to avoid distillation, the process in which synthetic data produced by one AI model trains another, less complex model.
Two noteworthy features were announced alongside the Grok 3 series. The first, DeepSearch, is xAI’s equivalent to the Deep Research capability recently introduced by OpenAI and Google. This feature thoroughly searches the Internet and the X platform to generate extensive analyses for complex queries.
Additionally, Grok 3 will feature a voice mode, allowing it to vocally respond to user inquiries, although this functionality will not be available upon initial launch. Musk mentioned that it could be implemented within a week.
Grok 3 benchmark
Photo Credit: X/Elon Musk
In terms of performance metrics, xAI claims that Grok 3 surpasses GPT-4o on several tests including the American Invitational Mathematics Examination (AIME), Graduate-Level Google-Proof Q&A (GPQA), and LiveCodeBench benchmarks. Internal evaluations also suggest that Grok 3 achieves higher scores than Claude 3.5 Sonnet, DeepSeek-V3, and Gemini-2 Pro. Furthermore, the reasoning models of Grok 3 were reported to exceed the performance of OpenAI’s o3 mini model.
Along with the new models, xAI introduced the SuperGrok subscription tier. Although pricing details were not disclosed, this new plan will incorporate features such as DeepSearch and Think (reasoning mode), greater image generation limits, and prioritized access to future features. Subscribers to X Premium Plus will gain access to Grok 3, although some additional features will not be included in this tier.
Musk also stated that xAI plans to open-source the final version of the AI model once the current iteration reaches full deployment. He added that Grok 2 will be made available in open-source form once Grok 3 stabilizes and matures, which could take several months.
