xAI finishes training Grok V9

Elon Musk said on May 17 that xAI had completed training Grok V9, which he described as a 1.5 trillion-parameter foundation model, and outlined a short post-training schedule before release. Musk wrote on X that the next steps were supplemental training on Cursor data, followed by supervised fine-tuning and reinforcement learning, with a public launch in about three to four weeks. The post did not include benchmark results, safety documentation or a product page. xAI’s public model documentation still lists Grok 4.3 as its flagship generally available text model, with a May 12 update date. ### What exactly did Musk say was finished? Musk said in an X post that xAI had “finished training Grok V9,” calling the run “a banger.” The same post said the model was a 1.5 trillion-parameter foundation model and that more work remained before public release. The X post itself did not provide details on architecture, training tokens, compute used, or evaluation scores. (docs.x.ai) The 1.5 trillion-parameter figure matters because it is the clearest technical number Musk attached to the announcement. Musk also framed the release path as staged rather than immediate, saying the model would first receive additional data training and then standard post-training steps. ### How does that compare with what xAI sells today? (x.com) xAI’s developer documentation says Grok 4.3 is the company’s main current chat and coding model and describes it as “the most intelligent and fastest model we’ve built.” The docs list a 1 million-token context window and show Grok 4.3 as the default recommendation for most text use cases. The page does not disclose parameter count for Grok 4.3. (x.com) Grok’s public release notes say Grok 4.3 beta began rolling out on April 17, 2026 to SuperGrok and Premium+ subscribers. The release notes describe Grok 4.3 as a new pre-trained model matching the scale of Grok 4.20 with an improved architecture and a December 2025 knowledge cutoff. ### What is Cursor data doing in the plan? Musk said the next training step would use Cursor data before supervised fine-tuning and reinforcement learning. (docs.x.ai) The post did not specify whether that meant code-assistant interaction data, user-approved traces, synthetic data, or another dataset category, and xAI has not published a separate technical note explaining the source or scope in the materials reviewed here. (grok.com) The sequence Musk described follows a familiar large-model pipeline: pretraining first, then narrower post-training to shape behavior and task performance. In this case, Musk explicitly separated the completed phase — foundation-model training — from the remaining release work. ### Why is the release not immediate if training is done? Musk said xAI still planned supplemental training, supervised fine-tuning and reinforcement learning before making Grok V9 public. (x.com) Those steps typically happen after the base model is trained and can change how a model answers, follows instructions and handles product-specific tasks, though Musk did not provide a date for each stage individually. xAI’s current product pages show the company continuing to ship features around the existing Grok line while newer models are prepared. Grok’s release notes list Connectors on May 6 and Grok Build in beta for SuperGrok Heavy users on May 19, while the developer docs continue to route most users to Grok 4.3. ### What is still missing from the public record? (x.com) No xAI technical report was linked in Musk’s post, and no public benchmark table accompanied the V9 announcement in the sources reviewed here. That leaves open questions about training compute, inference costs, context length, modality support and how Grok V9 would be positioned against Grok 4.3 in the API or consumer apps. (grok.com) The next concrete milestone is the post-training sequence Musk named on May 17: supplemental training on Cursor data, then SFT and RL, followed by a public release in about three to four weeks. Until xAI publishes model cards, benchmarks or pricing, Grok 4.3 remains the company’s documented flagship public model. (x.com)