Grok 5: 6 Trillion Params, 10% AGI Shot, Q2 2026

Grok 5 is xAI's biggest bet yet — a 6 trillion parameter model using a Mixture-of-Experts (MoE) architecture, trained on the Colossus 2 supercluster in Memphis (now scaling from 1GW to 1.5GW). The Q1 2026 launch window came and went; xAI now points to Q2 2026.

Musk pegs Grok 5's probability of hitting AGI at 10% and rising — a claim he's careful enough to hedge, even as he publicly predicted AGI (smarter than any single human) by end of 2026. For context, he said the same about 2025 last year.

The 6T parameter count sounds staggering, but MoE means only a fraction activates per query — it's more about capability breadth than raw compute per call. Whether that translates to AGI or just a very good benchmark score remains the real question.

Why it matters: Grok 5 is the clearest signal yet that the parameter arms race is still alive — and xAI is willing to burn gigawatts to stay in it.