2026-06-02 — views

xAI finishes training Grok V9-Medium, a 1.5T-param model tuned on Cursor developer data

Read this because The headline isn't the 1.5T parameter count — it's the corpus. Tuning a frontier model on Cursor's real developer workflows is a direct bid for the coding layer Claude and Codex dominate. Treat the benchmarks and timeline as vendor-sourced until weights or an API ship.

Musk says xAI's 1.5T-param Grok V9-Medium finished training (May 25), ~3x its production model and trained on Cursor dev data — mid-June release expected.

On May 25, 2026, Elon Musk announced that xAI’s new Grok foundation model, V9-Medium, has completed training with what he described as positive evaluation results. The model carries 1.5 trillion parameters — roughly three times larger than the version currently handling all Grok production traffic. Musk said supervised fine-tuning was already underway, reinforcement learning would begin within days, and a public release was expected roughly two to three weeks out, placing it in mid-June 2026.

The detail that matters

The number most people will repeat is “1.5 trillion parameters.” It is not the interesting part. The interesting part is what V9-Medium was trained on: Musk said the model was explicitly trained on Cursor data — real-world developer workflows from one of the most widely used AI code editors.

That is a deliberate choice, not a footnote. Most frontier models learn to code from public repositories, documentation, and synthetic problems. Training directly on the interaction traces of working developers — what they ask for, how they iterate, where they accept or reject suggestions — is a bid to model the act of building software, not just the finished artifact. It signals exactly where xAI wants to compete.

The two-model picture

There are really two Grok models in flight, and the gap between them is the story:

Model	Size	Status	Notes
Grok V9-Medium	1.5T params	Training complete; SFT + RL next	~3x current production model; trained on Cursor data; release ~mid-June
Grok 5	6T params (MoE)	Still actively training	Prediction markets give it ~33% odds of shipping by June 30

The flagship Grok 5 — a 6-trillion-parameter Mixture-of-Experts model — is still in training, and prediction markets give it only about a 33% probability of shipping by June 30, 2026, amid post-training uncertainty and reported staff departures since SpaceX absorbed xAI in February 2026. So xAI’s near-term play isn’t the flagship at all. It’s the smaller, coding-specialized Medium model, shipped now.

Why it matters for builders

If the mid-June date holds, V9-Medium lands squarely on the workload that monetizes fastest in AI today: coding. This is the layer where Anthropic’s Claude and OpenAI’s Codex currently dominate developer mindshare and spend. A frontier-scale model tuned specifically on Cursor’s workflows is an explicit attempt to contest that layer head-on — not with a general assistant, but with a system shaped by how developers actually work.

The sequencing is the pragmatic part. Rather than wait on the 6T flagship, xAI is shipping a smaller, faster, coding-focused tier first — the same fast/cheap-vs-frontier split that rivals already run. A 1.5T coding specialist that arrives in June can start applying pressure long before a 6T generalist with uncertain timing ever ships.

Practitioner note

Treat every number here as vendor-sourced until proven otherwise. The training-completion and benchmark claims come from xAI and Musk, and have not been independently verified — there are no public weights, no API, and no third-party evals yet. “Training complete” is also not “release ready”: supervised fine-tuning and reinforcement learning still sit between today and a usable model, and that gap is exactly where capability and behavior get decided. If you’re planning to evaluate V9-Medium for a coding workflow, the only test that counts is your own — run it against your real repositories, your review gates, and your acceptance criteria once an API exists. Until then, “trained on Cursor data” is a positioning claim about intent, not a measured result about quality. Watch for weights or an endpoint; ignore the parameter count.

The under-considered angle

Training a frontier model on Cursor’s developer data raises a question that has nothing to do with parameters: whose workflows become the teacher? When a model learns to code from the interaction traces of an editor’s users, the resulting product is shaped by — and competes with — the very tool that supplied the signal. That entangles xAI, Cursor, and the developers whose sessions became training data in a way that is rarely spelled out at announcement time. The capability story is straightforward. The data-provenance and incentive story underneath it is the one worth watching as the coding-agent layer consolidates.