Musk’s Grok V9-Medium Trained on Cursor Data, Targets High-Difficulty Coding

Elon Musk announced that xAI’s 1.5-trillion-parameter Grok V9-Medium base model has completed training and will be released to the public in two to three weeks. The model is now in supervised fine-tuning and will begin reinforcement learning training in a few days. The research team incorporated a large amount of Cursor code data during supplementary training, and will continue to feed it more samples.

Compared to the current v8-small model that handles all of Grok’s production traffic, the 1.5T parameter Grok V9-Medium will bring a significant upgrade in handling complex programming tasks. Musk previously acknowledged that the 0.5T parameter v8-small (the external test version Grok 4.2) had serious deficiencies in training data quality, comprehensiveness, and proportions. The finished Grok V9-Medium not only triples the parameter count of its predecessor but is also specifically optimized for Nvidia’s Blackwell architecture GPUs.