Late on Monday, Elon Musk’s AI venture, xAI, unveiled it’s latest advanced AI model, Grok 3, alongside new functionalities for its web and iOS-based Grok applications. Moving head-on with other AI giants like GPT-4o by OpenAI and Google’s Gemini, Grok, developed by xAI, has the capability to interpret images and cater to queries, while also being the force behind various features on Musk’s social platform, X. Grok 3, which has been under development for a significant span, had an optimistic launch expectation in 2024, which was not met.
xAI’s ambitious launch of Grok 3 on Monday can be traced back to an immense data center in Memphis that houses around 200,000 GPUs, used to facilitate the training of Grok 3. According to a statement by Musk on platform X, Grok 3 was harnessed with a computing power surpassing its predecessor, Grok 2, by a tenfold magnitude. This was possible due to an extensively broadened training dataset, which is speculated to encompass court case documents.
The capabilities of Grok 3 were proudly announced by Musk in a live-streamed presentation held on Monday. Musk shared, ‘Grok 3 is an order of magnitude more capable than Grok 2.’ Musk further spoke about the objective of developing an AI that strives towards truth regardless of potential conflicts with prevailing political correctness. Thuos underlining Grok 3 to be a pursuit of an honest AI.
Delving deeper into the Grok 3, it is identified as a family of models. More specifically, a quicker response variant, termed Grok 3 mini, has been introduced, designed to provide swifter answers at the trade-off of some accuracy. As a notice, not all models and linked features under Grok 3 have been made public yet, some are currently in the beta version, and the rollout began on Monday itself.
xAI boasts of Grok 3’s performance, claiming it to surpass even GPT-4o in benchmarks. This includes competitive arenas like AIME, which evaluates an AI model’s performance on a range of mathematical questions, and GPQA for assessing models based on their performance on PhD-level scientific problems in areas like physics, biology, and chemistry. An early variant of Grok 3 also showed competitive skills in Chatbot Arena, a crowdsourced test that ranks different AI models in head-to-head competition, as corroborated by xAI.
The Grok 3 family embraces two new models; namely, Grok 3 Reasoning and Grok 3 mini Reasoning, which display capabilities analogous to ‘reasoning’ models like OpenAI’s o3-mini and the R1 model from China’s AI firm, DeepSeek. Reasoning models are designed to undergo a self-check before furnishing outcomes, thereby avoiding typical model errors. According to xAI, Grok 3 Reasoning outperforms the top-tier o3-mini version, i.e., o3-mini-high, on various widely referenced benchmarks, including a recently developed mathematics benchmark called AIME 2025.
Users can access these reasoning models through the Grok app. When prompted, Grok 3 can ‘Think,’ or for more complex queries, users can turn on ‘Big Brain’ mode which employs additional processing power to reason through the problem at hand. For xAI, these reasoning models are best suited for tasks related to mathematics, sciences, and programming.
In order to shield some of the ‘thoughts’ of the reasoning models present in Grok from potential distillation by other developers, these are obscured within the Grok app. Distillation is a technique used by AI developers to glean knowledge from other models, which has led to accusations such as the recent one leveled against DeepSeek for allegedly distilling information from OpenAI’s models to develop its own.
DeepSearch, a new feature underpinned by Grok’s reasoning models, is xAI’s response to AI-powered research tools like OpenAI’s deep research. DeepSearch has the ability to scan online resources and X, analyze the information and then generate an abstract in response to a user’s query. The roadmap for Grok app also includes the addition of a ‘voice mode,’ as shared by Musk, enabling a synthesized voice for Grok; this is expected to roll out in the near future, approximately within a week’s time.
Post the voice mode integration, in a few weeks, Grok 3 models will also be made accessible through xAI’s enterprise API, complemented by the DeepSearch capability. Musk also shared plans to make Grok 2 open-source in the coming months whilst discussing their general policy of open-sourcing the preceding Grok version once the successor is completely rolled out.
Musk announced, ‘Once Grok 3 is stable and mature, which is probably within a few months, we’ll open-source Grok 2.’ This reiterates his commitment to the open-source community, a sentiment solidified since the launch of the initial Grok model approximately two years ago.
When Musk initially introduced Grok around two years back, it was pitched as an edgy, unfiltered, and anti-‘woke’ AI model, ready to answer controversial questions that other AI systems typically avoid. To some extent, this vision has been realized. For instance, Grok and Grok 2 did not shy away from vulgar language when instructed to do so, which starkly contrasts with models like ChatGPT.
However, it was observed that earlier Grok versions were apprehensive about tackling politically sensitive subjects and maintained certain boundaries. A particular study even portrayed Grok to have a left-leaning political inclination on issues such as transgender rights, diversity programs, and inequality. Musk attributed such behavior to the nature of Grok’s training data sources, which were mostly public web pages, and promised attempts to shift Grok towards a more politically neutral standing.
However, at this time, it remains unclear whether xAI has managed to realize this goal with Grok 3, and what potential consequences could accompany it. Thus, the effectiveness of Grok 3 in achieving political neutrality in its responses, while not compromising its ability to seek truth, will mark a crucial milestone in AI development moving forward.