When Elon Musk‘s xAI released Grok 4 Heavy in July 2025, it became the first AI system to score above 40% on Humanity’s Last Exam, a graduate-level benchmark spanning science, law, and mathematics. The result did not come from a larger model or a longer context window. It came from an architecture that runs up to 16 instances of Grok 4 simultaneously on a single query, has each instance reason independently, and then synthesizes their answers before producing a final response. (Bloomberg, July 10, 2025) Thirteen months later, xAI’s Grok Bot extended that same multi-agent design from a benchmark result to a commercial product for professional workflows. (Bloomberg, August 11, 2026)
Quick context
Musk founded xAI in 2023 after departing OpenAI’s board. The company runs the Grok model family, integrated into the X platform, and operates Colossus, a supercomputing cluster in Memphis, Tennessee. StartupHub.ai data shows xAI employs 5,551 people as of August 2026, one of the larger dedicated AI lab headcounts we track, behind OpenAI (10,540) and Google DeepMind (8,831).
Sixteen agents, one answer: how Grok 4 Heavy works
The standard Grok 4 model, released July 10, 2025, operates as a single-instance reasoning system. The Heavy tier runs a different process: the system spawns up to 16 parallel instances of Grok 4, each working through the same query independently. Once the instances finish reasoning, they compare their outputs and converge on a final answer through a debate-style synthesis. xAI describes the design as analogous to a study group where participants work separately, then compare notes. (The Information)