At the Alibaba Cloud Summit, Tongyi Qianwen's technical roadmap left industry professionals in awe.

1M context length is just the starting point; 100M is the goal. Trillion parameters are no longer impressive—the battlefield is now at the 10-trillion level. Testing computational power is set to expand from 64k to 1 million times, with training data aiming for 100 trillion tokens. They even proposed 'unlimited generation of synthetic data.'

On the screen behind the speaker in the photo, three key phrases stood out: Qwen-VL (visual language), Qwen-Omni (all-purpose model), and World Model. Engineers in the audience recorded the event on their phones, some frowning as they noted parameter comparisons.

The comment section was divided. Some cheered, 'Finally, someone dares to propose 100M context,' while others questioned, 'Hybrid thinking models were just proven effective, and now we're pivoting to an all-purpose approach?' More practical concerns came from code engineers: 'Even 200k context can't hold a complete project now.'

The most intriguing observation came from a buried reply: While everyone debates the parameter race, Baidu might be preparing a disruptive solution in the lab. China's AI battlefield never follows a script—three years ago, no one expected large models to iterate monthly.

An easily overlooked detail in the screen's bottom-right corner: Test-time compute was listed as a standalone technical metric. What does this mean? When parameters balloon to 10 trillion, inference costs will become a more pressing bottleneck than technological breakthroughs.

The speaker never once mentioned 'AGI.'