July 18th, Elon MaskElon MuskTHIS POST IS PART OF OUR SPECIAL COVERAGE GLOBAL VOICES 2011xA The next generation in training Grok 4.6 The size of the model parameters will reach $2 trillion, with initial training expected to be completed next week。
Mask states that the 2 trillion-billion-parameter model is better in every respect than the current 1.5 trillion-billion-parameter version, while the speed of reasoning and the efficiency of Token will be close to the existing 1.5-T model (i.e. Grok 4.5). Mask speculates that it might exceed the dark side of the moon's latest Kimi 3。
Currently, xAI has not published the official release time, training details and complete technical indicators of Grok 4.6. Previously, xAI had launched several versions of the Grok series。
In terms of the size of the parameters, Mask indicated in May 2026 that the Grok Basic Model V9-Medium (1.5T parameters) had completed its training and that a fine-tuning and intensive learning phase was under way. The model is considered a basic version of Grok 4.5. If this 2T parameter model is finalized and published, the size of Grok parameters will be further increased。
The number of parameters in the large language model is usually used to measure the size of the model, but the actual capacity is also influenced by factors such as the quality of training data, training methods, reasoning structures, computing resources and post-training optimization. What Mask wanted to say was that XAI, while expanding the size of the model, wanted to maintain operational efficiency close to the previous generation。
From a development perspective, XAI is continuously expanding the input of computing resources. The company had previously announced that Grok 4 had used the large AI computing cluster called Colossus for training, which had approximately 200,000 GPUs for intensive learning and model training。
At the same time, competition in the AI large model area is taking place around model size, reasoning ability, response speed and cost efficiency. Businesses like OpenAI, Google and Anthropic are also continuing to introduce a new generation of models, and industry is generally concerned with how to increase capacity while reducing the cost of reasoning。