AI lab claims 48-billion-parameter model that runs on 2.7 billion — if the math holds

Researchers released an AI language model that claims to use only 5.6% of its stated parameters for each task, rather than activating everything at once. If accurate, this means AI inference becomes dramatically cheaper to run, which would lower the cost barrier for companies deploying large models in production.

The constraint on AI deployment right now is not capability — it's cost. Running a large language model costs money in proportion to how much computation you do. This paper claims a structural solution to that problem: activate fewer parameters per task, keep the same performance. If true, it removes a major friction point for adoption. The catch is that sparse models are notoriously hard to validate outside controlled benchmarks, and this is a preprint with no independent verification yet.