xAI's Grok V9-Medium Completes Training — 1.5T-Parameter Coding Model Targets Mid-June Launch
xAI's next flagship model has finished training at 1.5 trillion parameters, triple the current production model. Trained on real Cursor developer workflows, V9-Medium is a direct challenge to Claude Code and GPT on SWE-bench coding benchmarks.
xAI’s Grok V9-Medium completed training on May 25, and a public release is expected in mid-June 2026 — the most direct challenge yet to Claude Code and GitHub Copilot’s dominance in AI-assisted coding.
The model runs 1.5 trillion parameters, three times the 500-billion-parameter V8 currently serving all production Grok traffic. That gap matters: Grok 4 scores 72–75% on SWE-bench Verified, while Claude Opus 4.6 hits 80.8% and GPT-5.5 reaches 88.7%. xAI is explicitly targeting that delta.
Trained on Real Developer Workflows
The distinguishing technical bet is the training data. Instead of relying solely on public GitHub repositories, xAI trained V9-Medium on Cursor data — real-world coding sessions from one of the most widely used AI code editors, with power users at OpenAI, Stripe, and Perplexity in the mix. The premise: if you want a model that works the way developers actually work, train on what developers actually do.
Supervised fine-tuning was already underway at the time of Elon Musk’s announcement, with reinforcement learning scheduled to begin days later. The training ran on xAI’s Colossus 2 supercluster in Memphis, Tennessee — the same facility used to train Grok 3.
Musk described the current production model as “just 0.5T” and acknowledged it lacks important training data — unusually candid framing ahead of a launch.
What Comes Next
V9-Medium will replace V8 in production when it ships. Musk also confirmed plans to open-source the current 500-billion-parameter model by the end of 2026, a move that keeps developer goodwill intact while protecting the frontier model.
The mid-June window puts V9-Medium squarely against Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5. Whether training on Cursor workflows is enough to close a 10–15 percentage point SWE-bench gap is the real question.
Independent benchmark results will be the actual story here. xAI has been known to slip release windows — V8 shipped two weeks late — but training completion is a concrete milestone. The coding AI wars have a new entrant arriving shortly.