Elon Musk's Grok V9-Medium Completes Fine-Tuning Training, Introduces Cursor Data for Challenging Programming Scenarios

According to Watchful AI monitoring, Musk has announced that the xAI subsidiary's 1.5T-parameter Grok V9-Medium base model has completed training and is expected to be publicly released in 2 to 3 weeks. The model is currently undergoing supervised fine-tuning and will commence reinforcement learning training in the next few days. The research team has incorporated a large amount of Cursor's code data in the fine-tuning process and plans to inject even more samples in the future.

Compared to the current v8-small model that supports all of Grok's production traffic, the Grok V9-Medium with 1.5T parameters will see a significant improvement in handling complex programming tasks. Musk previously acknowledged that the v8-small with only 0.5T parameters (i.e., the external testing version Grok 4.2) had major flaws in training data quality, comprehensiveness, and scale. The trained Grok V9-Medium has not only tripled its parameter size compared to its predecessor but has also been specifically optimized for the Blackwell architecture GPU.