xAI has launched Grok 3, the latest iteration of its large language model, alongside new capabilities for its iOS and web applications. The model has been trained on approximately 200,000 GPUs in a Memphis data center, representing what CEO Elon Musk claims is a tenfold increase in computing power compared to its predecessor.

The new release introduces two specialized variants: Grok 3 Reasoning and Grok 3 mini Reasoning, designed to methodically analyze problems similar to OpenAI's o3-mini and DeepSeek's R1 models. According to xAI's benchmarks, Grok 3 outperforms GPT-4o on several technical evaluations, including AIME for mathematical reasoning and GPQA for PhD-level science problems.

A notable addition is the DeepSearch feature, which combs through web content and X posts to generate research summaries. The platform will be available through X's Premium+ subscription and a new SuperGrok tier ($30/month or $300/year), with the latter offering enhanced reasoning capabilities and unlimited image generation. To prevent knowledge extraction through model distillation -- a technique recently attributed to DeepSeek's alleged copying of OpenAI's models -- xAI has implemented measures to obscure the reasoning models' thought processes in the Grok app. The company plans to release the Grok 2 model as open source once Grok 3 achieves stability.

[ Read more of this story ]( https://slashdot.org/story/25/02/18/0711226/xai-releases-its-latest-flagship-model-grok-3?utm_source=atom1.0moreanon&utm_medium=feed ) at Slashdot.