xAI Teases 2 Trillion-Parameter Model Ahead of Training Milestone

Elon Musk says xAI is close to completing the first training stage of a new 2 trillion-parameter language model, with that milestone expected next week. According to Musk, the model already outperforms Grok 4.5 on all major metrics.

He also suggested the new model could surpass Kimi while keeping the speed and efficiency of the older Grok. That claim lands in a familiar AI race, where companies have long used model size as a proxy for leadership. But the argument is shifting: performance alone is no longer enough. Developers and buyers are paying closer attention to inference speed, cost per response, and how much compute a model needs to deliver results.

The latest round of discussion was prompted by researcher Min Choi, who wrote that Grok 4.5 has roughly 1.5 trillion parameters, while Kimi K3 has about 2.8 trillion. By his estimate, tasks run on Kimi cost roughly three times more.

That makes xAI’s pitch easier to read. The company is not just leaning on raw scale, but on getting more output from the same compute budget. For the market, that marks a real split in strategy. Some companies, including OpenAI, Google, and Chinese labs, have pushed maximum size. Others are trying to show that architecture, training, and inference efficiency matter more.

Recommended reading

Ex-Red Bull engineer lands $55m for robot training data

xAI is hardly new to that contest. Musk has repeatedly used the Grok line to show that the company can quickly close the gap with major players in generative AI. This time, though, the emphasis appears more practical: trillions of parameters, response speed, and task cost.

China’s AI sector is moving quickly too. Moonshot AI positioned Kimi K3 as the largest open-source model, helping it enter the leadership conversation fast. The pattern is familiar: one company scales up, another responds with openness or efficiency, and the market decides which matters more for businesses and developers.

A 2 trillion-parameter label proves little on its own, but it points to where the competition is heading in 2025 and 2026. The focus is moving from sheer model size to the balance between quality, cost, and speed. If xAI does finish the first training phase next week, the next comparisons are likely to center not just on answer quality, but on what it costs to run the model in real-world deployments.