Musk says Grok V9-Medium has finished training

On 25 May 2026, Elon Musk said xAI's V9-Medium — reportedly ~1.5 trillion parameters, around three times the current production model — had finished training, with internal evaluations described as positive. A mid-June public release was floated in subsequent tech coverage. By mid-June: no confirmed date, no model card, no independent benchmark numbers. The separately hyped Grok 5 — a different, reportedly even larger model — remained in training, and prediction markets were skeptical of a June ship for that one too. xAI's current developer-accessible flagship is Grok 4.3, per the official developer docs.

Three reasons to hold back. First, these are unverified claims: the parameter count, the 'positive' eval results, and the timeline all come from xAI with no third-party confirmation and no published technical report. Second, xAI's release timelines have slipped before — the floated mid-June window passed without a public update, a named pattern, not editorial sniping. Third, parameter count alone tells you nothing about quality: smaller, well-tuned models regularly beat larger ones on real-world tasks, so '1.5 trillion' is a measure of training cost, not of how good the model is. 'Finished training' is a genuine internal milestone, but it is not 'shipped', not benchmarked, and not a product you can evaluate. The signal to upgrade from 'intent' to 'real' is simple: a model card with numbers and a callable API.