Grok 3 family of artificial intelligence (AI) models were announced on Monday. The successor to the Grok 2 models was announced in a live event hosted by xAI engineers and the company owner Elon Musk. The new series contains several large language models (LLMs) with different parameter sizes and reasoning-based variants. The new models will come with new features such as DeepSearch and a voice mode. Alongside, the company also announced a new subscription tier dubbed SuperGrok to access higher rate limits and certain new features.
Grok 3 AI Models Unveiled by xAI
Musk hosted a live stream on X (formerly known as Twitter) showcasing the capabilities of the new AI models. Calling Grok 3 “an order of magnitude more capable than Grok 2,” the billionaire highlighted that xAI built its new data centre to pre-train the LLMs, and the first 100,000 GPUs were running within 122 days. The capacity was further doubled in the next 92 days.
The Grok 3 family comprises several LLMs, but not all of them are available currently. At the event, the company unveiled Grok 3, Grok 3 mini (a smaller but faster model), alongside the Grok 3 Reasoning and Grok 3 mini Reasoning (test time compute-based reasoning models). The rollout for the announced AI models has begun, but some other models are currently in the beta phase.
Musk highlighted that some of the chain-of-thought (CoT) steps will be obscured in the Grok 3 reasoning models to prevent instances of distillation. Notably, distillation is the process when the synthetic data generated from one AI model is used to train another, smaller model.
Two new features were also unveiled with the Grok 3 family. First is DeepSearch, which is xAI’s version of the Deep Research feature recently launched by OpenAI and Google. It scours the Internet and the X platform to analyse information when a complex query is asked and generates a comprehensive report.
Grok 3 will also get a voice mode that will allow the AI model to respond to queries verbally. However, this feature will not be available at launch. Musk said it could be rolled out “as soon as a week from now.”
Grok 3 benchmark
Photo Credit: X/Elon Musk
Coming to benchmarks, the company claimed that Grok 3 outperforms GPT-4o on the American Invitational Mathematics Examination (AIME), Graduate-Level Google-Proof Q&A (GPQA), and LiveCodeBench benchmarks. It is also claimed to achieve higher scores than the Claude 3.5 Sonnet, DeepSeek-V3, and Gemini-2 Pro based on internal testing. xAI also said the Grok 3 reasoning models outperform OpenAI’s o3 mini model.
Alongside the new models, xAI also introduced the new SuperGrok subscription tier. While pricing details were not revealed, this tier will offer features such as the DeepSearch and Think (reasoning) mode, higher image generation limits, and early access to new features. Those subscribed to X Premium Plus will get access to Grok 3, however, other features are not available at this tier.
Musk also said that the company will adopt the policy of open-sourcing the last version of the AI model once the current version is fully rolled out. He added that once Grok 3 is stable and mature, which could take up to a few months, Grok 2 will be released in open-source.
