Musk predicts Grok 4.6/4.7: 1.5T parameters will be launched in August, 2.1T parameters will follow closely

In the model circle, Musk is probably the only one who dares to publish the next generation and next-generation roadmap before release. On July 28, the head of xAI publicly announced the release plans of Grok 4.6 and Grok 4.7: the former is expected to be launched around August 7, with a parameter scale of 1.5 trillion; the latter will take over a few weeks later, and the parameter scale will be further enlarged to 2.1 trillion.

Technology increment of two generations of models

According to disclosed information, the focus of the Grok 4.6 upgrade is not on scale, but on the training process - significant improvements have been made in two key aspects: supervised fine-tuning (SFT) and reinforcement learning (RL). SFT determines the instruction compliance and alignment quality of the model, and RL determines the reasoning ability and behavioral stability. The enhancement of these two means that the "smartness" of Grok 4.6 comes more from the improvement of the training method rather than simply heaping parameters.

Grok 4.7, which followed closely, returned to the parameter route: 2.1 trillion parameter volume, aiming directly at breaking through the upper limit of capabilities. From 4.5 to 4.6 and then to 4.7, xAI used two generations of models to show a clear iteration curve - first using the training process to polish the quality, and then using scale to hit the limit.

"Continuous" competitive strategy

The most informative thing about this trailer is the rhythm itself. From the full launch of Grok 4.5 in mid-July, to Grok 4.6 in early August, and then to 4.7 a few weeks later, xAI is implementing an almost "monthly update" high-frequency shipping strategy. This is in sharp contrast to the half-year-old pace of Anthropic behind Claude and OpenAI behind ChatGPT - xAI chooses to use shorter intervals and faster iterations to continue to suppress its opponents in terms of momentum.

From an industry perspective, the costs and benefits of this strategy are equally obvious: high-frequency updates can quickly absorb community feedback and keep the topic hot, but it also places higher requirements on the stability of the model and the continuity of the API ecosystem. For developers, whether they are willing to follow the pace of xAI and frequently upgrade depends on whether the API opening method and pricing of Grok 4.6/4.7 can be given enough sincerity when it is officially released. At present, these are just "previews", and the official has not yet disclosed the specific performance and access costs.

Several directions worth tracking in the future:

1. Whether it will be on schedule on August 7: Whether the ticket will be delayed or not will directly test the project execution rhythm of xAI.
2. Perceptibility of RL improvements in Grok 4.6: Can the improvements in alignment and reasoning capabilities be realized in benchmarks and actual measurements?
3. Computing power account of 2.1T parameters: The training and inference costs of Grok 4.7 will test xAI’s resource reserves.
4. The pace of API and pricing opening: Under continuous updates, can ecosystem partners keep up with the pace of migration?