Musk says xAI's 2-trillion-parameter model finished initial training

Musk said xAI's next frontier model — a 2-trillion-parameter system — is finishing its initial training run, with the completion milestone expected in the final days of July and evaluation plus safety testing to follow. Broader availability could come as soon as August 2026. The model represents a 33% parameter jump over the 1.5T 'V9' that currently powers Grok 4.5, and Musk claims it will be 'better than the 1.5T in every way.'

The scale-up mechanics are conventional — more parameters trained on xAI's expanding Colossus compute — but the timing is pointed. It lands the same week Kimi K3 debuted at 3-trillion-parameter open scale and Google's Gemini 3.5 Pro slipped, letting Musk frame xAI as the lab still executing on aggressive cadence. Musk also teased 'Grok Imagine' on X this week, signaling a multimodal push alongside the core model.

Competitively, raw parameter count no longer maps cleanly to capability — Moonshot's own K3 concedes it trails smaller closed models — so the 2T figure is as much marketing as engineering. What matters is whether the new model closes the agentic-coding and reasoning gap with Claude Fable 5 and GPT-5.6 Sol, which currently lead.

The caveat: xAI is simultaneously fighting legal fires — a lawsuit against a Grok user over sexualized deepfakes and a child-exploitation class action that added Stability AI as co-defendant this week — which complicates any 'safety testing complete' narrative. A frontier release into that legal backdrop invites intense scrutiny. Watch the August window and whether xAI publishes benchmarks rather than superlatives.