Grok-3 Mini

Grok 3 Mini, the lightweight sibling of xAI’s flagship model, proves that “small” doesn’t have to mean “shallow.” Designed with a leaner architecture and optimized for speed, Mini still demonstrates impressive reasoning capabilities—particularly when enhanced with test-time inference via the Think setting. On AIME 2024, a benchmark of symbolic reasoning under pressure, Grok 3 Mini (Think) achieves a striking 95.8%, outperforming larger rivals like Gemini 2.0 Flash Thinking (73.3%) and even Grok 3 Beta (93.3%) itself. It also leads in code generation with 80.4% on LiveCodeBench, edging out Grok 3 Beta (79.4%) and surpassing Claude 3.5 Sonnet (66.3%) and GPT-4o (32.3%) by wide margins. Its 84.0% on GPQA confirms strong performance on graduate-level, adversarial reasoning tasks. This positions Mini not as a fallback, but as a front-line option for STEM-heavy tasks where speed, cost, and accuracy matter—like on-device tutoring, fast coding agents, or high-throughput evaluation pipelines. Critically, Grok 3 Mini (Think) benefits from large-scale RL fine-tuning, allowing it to engage in multi-step reasoning with error correction and backtracking—something previously reserved for frontier models. While it lacks the full contextual breadth of Grok 3 (1M token context, advanced multimodal fluency), Mini still posts competitive scores on MMLU-Pro (78.9%) and MMMU (69.4%), reflecting robust general knowledge and image understanding. In short, Grok 3 Mini isn't a "cut-down" model—it's a recalibrated one. For researchers and builders focused on compute efficiency without compromising reasoning depth, Grok 3 Mini (Think) offers an unusually capable tradeoff.