Musk Responds to Slow Progress of Grok: V9 Training Completed, Release in 3-4 Weeks
Musk responded to community concerns about the slow progress of the Grok model, confirming that the internal V8 (0.5T parameters, publicly known as Grok 4.3) is being continuously optimized every few days.
The internal V9 (1.5T parameters) has completed training and achieved significant upgrades. The next step will involve adding Cursor data for supplementary training, followed by supervised fine-tuning (SFT) and reinforcement learning (RL). The entire process is expected to be released to the public in about 3 to 4 weeks.
This timeline indicates that xAI is accelerating the generational leap from 0.5T to 1.5T while improving model quality through the closed loop of Cursor data, in response to community expectations for iteration pace.
Source: Public Information
ABAB AI Insight
Musk has previously publicly acknowledged the correspondence between xAI's internal and external versions. This direct response regarding the progress of V8/V9 continues his style of transparent communication. Earlier, he admitted shortcomings due to defects in Grok 4.2 and promised rapid iterations. The 3-4 week timeline and integration of Cursor data mark the first substantial implementation following the SpaceX ecosystem acquisition.
On the capital path, xAI is rapidly shifting resources from the Hopper cluster to optimize training with Blackwell, injecting high-quality code and product data through the Cursor acquisition. The motivation is to shorten the scale gap with OpenAI and Anthropic while building a full-stack data closed loop, accelerating vertical integration from foundational models to application layer capabilities.
Similar to the rapid iteration cycles from Anthropic Claude 3 to Claude 3.5, and the Meta Llama series achieving capability leaps through continuous data supplementation, xAI is currently at a critical window of transitioning from "small-scale rapid releases" to "large-scale high-quality closed loops." The training cycle of large models and data assets have become core variables.
Essentially, this is a technological substitution: the capabilities of AI foundational models highly depend on parameter scale, data quality, and training architecture. V9 achieves generational advancement over V8 through 1.5T scale + Cursor data + Blackwell optimization, mechanism-wise allowing hardware generational replacement and acquisition of data assets to become decisive levers for surpassing competitors, pushing the industry from "parameter arms race" to "full-stack data-computing-model closed loop control."
ABAB News · Cognitive Law
Publicly acknowledging slowness is the beginning of accelerating the next leap.
Data acquisition, rather than parameter stacking, can more decisively determine generational gaps in models.
Community skepticism is fuel, and the 3-4 week iteration is the answer.