Grok 3 Mini

Grok 3 Mini is a compact reasoning-focused model built to handle logic-based problems with computational efficiency. As a lighter member of xAI's Grok 3 family, it applies a "think before responding" approach—the model generates internal reasoning traces that are exposed to users rather than hidden behind a black box, making its problem-solving process interpretable. It demonstrates particular strength across STEM domains, including graduate-level science knowledge, math competition problems, and multi-step reasoning tasks, achieving top-tier results among comparable non-reasoning models on benchmarks like GPQA Diamond and MATH Level 5. The model processes extended inputs using a large context window, enabling analysis of lengthy documents or complex workflows in a single pass.

Grok 3 Mini inherits its foundation from xAI's Grok 3 development lineage, which leveraged large-scale reinforcement learning during training to refine reasoning capabilities over seconds to minutes of test-time compute. This approach was developed using xAI's Colossus supercluster with substantially more compute than previous state-of-the-art systems, marking a generational step in how the family balances reasoning depth with efficiency. The model is offered in two serving modes—standard Grok 3 Mini optimized for reduced cost and Grok 3 Mini Fast on faster infrastructure for latency-sensitive applications—while both deliver identical response quality. Its combination of accessible reasoning traces, strong performance on academic benchmarks, and cost efficiency positions it for workflows where transparent multi-step reasoning adds value without the overhead of larger reasoning models.