Grok 3 Reasoning (Beta) is a high-performance language model developed by xAI, optimized for complex logical deductions, advanced mathematics, and scientific reasoning. It utilizes a specialized "Think" mode that leverages test-time compute and large-scale reinforcement learning (RL) to generate internal chain-of-thought processes. This allows the model to analyze problems for extended periods, performing self-verification and error correction before delivering a final answer.
Trained on the Colossus supercomputer cluster using approximately 200,000 NVIDIA H100 GPUs, Grok 3 represents a significant scale-up in compute power compared to its predecessors. According to xAI, the model achieved state-of-the-art results on challenging benchmarks, including the 2025 American Invitational Mathematics Examination (AIME) and the Graduate-Level Google-Proof Q&A (GPQA), where it demonstrated expert-level scientific proficiency.
The model features a context window of 1 million tokens, enabling the processing of extensive technical documentation and long-form codebases. While Grok 3 provides users with summaries of its thought process, the full chain-of-thought tokens remain obscured to protect proprietary training methodologies. The reasoning capabilities are part of a broader release that includes a standard non-reasoning Grok 3 model and a more efficient "mini" reasoning variant.