A new look at costs
The growth of generative AI faces a major barrier: computing power and cost. Companies see the potential, but they are hitting the limits of what is financially and environmentally feasible. Researchers from Tencent AI and Tsinghua University now present an alternative model.

AI with less computation
Their design, Continuous Autoregressive Language Models (CALM), no longer predicts individual words but compresses multiple tokens into a single continuous vector. This drastically reduces the number of computation steps. Instead of generating a sentence like “it is raining hard” word by word, CALM compresses it into a single computation. The result: up to 44% less training cost and 34% lower inference cost while maintaining comparable performance. This allows enterprises to achieve more with fewer resources.
A new toolkit for continuous AI
Because CALM does not rely on probability distributions, the researchers had to design new training and evaluation methods. They introduced a “likelihood-free” approach using an Energy Transformer and a new evaluation metric: BrierLM, which provides reliable scoring without traditional probabilities. This breakthrough also restores controlled text generation — essential for business applications.
Smarter scaling, not bigger models
CALM shows that scaling is no longer the only path to better performance. By increasing the semantic bandwidth instead of the number of parameters, it opens a new path toward powerful and affordable AI. This could help businesses integrate generative AI sustainably — from data centers to edge applications — with lower energy use and reduced costs.