xAI has completed the initial training run for its Grok V9 foundation model, which incorporates 1.5 trillion parameters.
This represents a threefold increase from the 0.5 trillion parameters that underpin the current Grok 4.3 version, itself based on the earlier V8 architecture. Early internal evaluations indicate that the new model has delivered strong results even prior to the next stages of development. The training incorporated enhanced data curation techniques, an updated training recipe, and optimizations tailored for Nvidia Blackwell GPUs.
Its plan is to conduct supplemental training that will integrate additional data from Cursor, the AI code editor. This step will precede supervised fine-tuning and reinforcement learning processes.
A public release of Grok V9 is expected to happen soon, and during this interval, the existing 0.5 trillion parameter model continues to receive incremental improvements on a frequent basis.
Statements from xAI leadership have described the upcoming version as a substantial advancement in scale and overall performance relative to its predecessor.
The version numbers are a little confusing and deserve some explanation.
Internally, we are working on version 9 of our new foundation model, which is 1.5T params. This is substantially better in every way than v8: data curation, training recipe, size, etc. It is also optimized…— Elon Musk (@elonmusk) May 15, 2026
The update arrives amid active development of even larger architectures within xAI.
Multiple models are currently in training on the Colossus infrastructure, including variants aimed at 6 trillion and 10 trillion parameters. These efforts reflect a sustained focus on expanding model capacity while refining training methodologies across several concurrent projects.
The partnership with Cursor also provides mutual access to computing resources, with xAI supplying GPU infrastructure to support Cursor model training and exploring potential deeper collaboration that could include data sharing for Grok tuning.
We are improving the 0.5T Grok foundation model V8 (public version 4.3) every few days.
The 1.5T V9 just finished training (incorrectly called pre-training) and is a major upgrade. Next, we are adding the Cursor data in supplemental training (others call this mid-training),…— Elon Musk (@elonmusk) May 17, 2026
At the same time xAI has rolled out Grok Build, an early beta coding agent available exclusively to subscribers on the SuperGrok Heavy tier.
The tool functions as a command line interface that supports planning, review, and execution of complex software engineering tasks.
It incorporates features such as parallel sub-agents for research and evaluation, along with an automated mode that ranks competing solutions before human review.
Early users have noted its responsive terminal interface and the speed with which the team has addressed initial feedback through rapid updates. Feedback has emphasized its utility for professional coding workflows, though availability remains restricted to the highest subscription level at present.
Elon on Grok v9.
Foundation model v9, 1.5T params, training run just finished. Cursor data still pending as supplemental.
Versus v8: 3x larger, rebuilt data curation, new training recipe, Blackwell native, v8 was Hopper.
Public Grok 4.2 is still v8 underneath. Elon says the v8… pic.twitter.com/0jlfhKqsYr— tetsuo (@tetsuoai) May 16, 2026
While the public is eager for this Grok V9 and what it may introduce, particularly in coding and general reasoning tasks, developments have coincided with ongoing reports of usage constraints affecting many SuperGrok subscribers.
Individuals on tiers starting at $30 per month and extending upward have described reaching daily or session limits after generating only a small number of images or video clips, in some cases as few as 10 to 20 images or three videos. Voice interaction features have similarly been reported to cap out after 20 to 30 minutes of use for certain accounts.
These limits apply across both standard and higher paid plans and have prompted discussions about the gap between expected availability and actual experience.
Users have pointed to the lack of clear quota visibility in the interface as an additional source of inconvenience during active projects.
xAI has attributed the restrictions to rapid growth in overall demand and the resources required to support model training alongside user services. The company has stated that capacity expansions are in progress to ease these pressures, noting that similar periods of adjustment have occurred during earlier phases of expansion.
Some observers have characterized the situation as temporary growing pains associated with scaling large scale AI systems.
Others have expressed disappointment that paid subscribers continue to encounter such constraints on creative and productivity tools that were highlighted in earlier communications.
The current phase illustrates the challenges involved in advancing frontier AI models while managing the expectations and experiences of a growing user base.