xAI Completes Training of Grok V9-Medium 1.5T Model
xAI announced that its Grok base model V9-Medium (1.5T parameters) has completed training.
The model evaluation results are good, with a supplementary training phase incorporating a large amount of Cursor data, which will continue to be added. Fine-tuning is underway, and reinforcement learning will start in a few days, with a public release expected in 2 to 3 weeks.
This model will significantly surpass the current 0.5T V8-Small version serving all Grok production traffic, especially excelling in complex coding tasks.
The new model release will further enhance Grok's competitiveness in the AI coding field, as xAI continues to strengthen its leading advantage through increased computing power and data investment, while competitors like OpenAI and Anthropic will face greater performance pressure.
Source: Public Information
ABAB AI Insight
Elon Musk has previously iterated Grok quickly through xAI, starting from early small models and continuously increasing parameter scale and data quality. This V9-Medium jumps directly from 0.5T to 1.5T, continuing its "train first, align later" approach.
On the capital front, xAI is concentrating resources on high-quality code data (such as Cursor) and the reinforcement learning phase, aiming to attract a developer user base through stronger coding capabilities, forming a data flywheel, while paving the way for potential commercialization subscriptions and enterprise services.
Similar to OpenAI's leap from GPT-3.5 to GPT-4, xAI is currently in a rapid expansion phase, attempting to overtake competitors through dual investments in parameters and data.
Essentially, this is about technological substitution and the transfer of pricing power: larger-scale high-quality training allows xAI to form performance barriers in complex tasks, driven by the scarcity of code data and the compounding effect of training efficiency, which will long-term push AI capabilities from general to specialized scenarios.
ABAB News · Cognitive Law
Model scale is just the starting point; high-quality data is the compounding engine.
Leaders widen the gap through iteration, while followers often lose direction in parameters.
The essence of technological leaps is to continuously invest scarce resources into the correct feedback loop.