Grok 3 Released: xAI Advances from Large-Scale Pretraining to Reasoning and Search Agents
xAI released Grok 3, Grok 3 mini, and their Think reasoning versions in February 2025, and previewed API and DeepSearch.
Grok enters the reasoning model competition, with products no longer emphasizing only real-time knowledge and expression style, but combining reinforcement learning, inference-time computation, long context, and search tools into agent capabilities.
Officially disclosed that Grok 3 was trained on the Colossus cluster, supporting 1M context; the Think version learns backtracking, verification, and multi-path solving through reinforcement learning, and can invest longer reasoning time per task.
Competition among frontier models now simultaneously tests pretraining scale, inference-time computation, and tool use, making single static benchmarks less able to explain real task capabilities.
Buyers need to compare Grok 3's success rate, latency, and search evidence quality on real tasks, rather than directly adopting peak scores at release.
Observe API stability, DeepSearch citation quality, tool call success rates, and cost curves under different reasoning budgets.