Minor adjustments to the way artificial intelligence models are built and used could lead to major energy savings, according to a new report from UNESCO. The 35-page report, “Smarter, Smaller, Stronger: Resource-Efficient Generative AI & the Future of Digital Transformation,” outlines practical ways developers and users can significantly reduce AI’s environmental footprint.

Three Strategies to Cut AI’s Energy Use

1. Use Smaller, Task-Specific Models

One of the report’s central findings is that smaller models can perform just as well as massive ones—especially when they’re designed for specific tasks.

“Small, tailored models can cut energy use by up to 90%,” the report says.

Rather than using a single, general-purpose model for everything from translation to summarization, it’s more efficient to deploy leaner models focused on one job. This not only saves energy but also improves performance in low-resource environments, where connectivity and hardware may be limited. Plus, smaller models respond faster and are less expensive to run.

2. Keep Prompts and Responses Short

The length of a prompt matters. Trimming input queries and keeping responses brief can cut energy use by over 50%.

“Every extra word the model has to process consumes compute power,” the report explains.

Shorter interactions not only save power but also reduce costs for organizations deploying LLMs at scale.

3. Compress the Models

Techniques like quantization and pruning—known collectively as model compression—can cut energy use by up to 44%. These methods reduce the computational demands of a model without necessarily sacrificing its performance.

Why Smaller AI Models Use Less Energy

The logic is simple: fewer parameters = less processing.

“Smaller models require fewer parameters, less memory, and significantly less GPU throughput,” explained Jim Olsen, CTO of ModelOp.

That translates to lower energy consumption during both training and use. In contrast, large models with billions of parameters must crunch vast amounts of data for every response—like a gas-guzzling V8 engine running at full throttle.

Wyatt Mayham of Northwest AI Consulting compares it to choosing a compact car over a high-powered sports vehicle:

“A smaller, more specialized model simply has less computational overhead for each task.”

And smaller models can be fine-tuned on private or proprietary data, making them highly effective in niche or sensitive applications, according to Virtualitics Chief Scientist Sagar Indurkhya.

The Hidden Cost of Long-Winded Prompts

Despite the “chatbot” nickname, AI doesn’t need to be treated like a conversationalist.

“The model doesn’t benefit from pleasantries,” said Mel Morris, CEO of Corpora.ai. “Extra words mean extra processing.”

Experts agree: long, overly complex prompts increase computational demand. Keeping prompts concise makes AI interactions more energy-efficient.

Still, brevity has its limits. Some prompts need context to produce accurate results. The goal isn’t to oversimplify—but to eliminate redundancy without losing meaning.

“Smarter prompts can save more energy than just shorter ones,” noted CloudX’s Axel Abulafia.

Compression: Powerful, but Tricky

Shrinking a model sounds great—until performance starts to suffer.

“Compress a model too much, and you lose accuracy or reasoning ability,” warned Mayham.

Compression requires deep technical knowledge, and the right approach depends on the model’s architecture and intended use. It’s not a universal solution.