Right-Sized AI Selection
Match model size to task complexity and measure energy cost before defaulting to the biggest available LLM.
Right-Sized AI Selection rejects the bigger-is-better default and treats model choice as an engineering decision with cost, energy, and accuracy tradeoffs. The premise: most AI queries are narrow tasks (a fact lookup, a classification, a summary) and a small, task-specific model handles them with comparable quality at a fraction of the energy and dollar cost.
The framework begins with task scoping. Before reaching for a general-purpose LLM, define what the model actually needs to do. Luccioni's research found general LLMs use up to 30x more energy than task-specific models for simple questions, and a small model can answer with 1/150th the energy of a large reasoning model on equivalent prompts.
Next comes capability matching. Map the task to the smallest viable model class: a 135M-parameter small LM for short factual queries, a domain-tuned model for vertical work (weather, bioacoustics, satellite imagery), and a frontier LLM only when generality is genuinely required. Run small models locally on phone or browser when privacy and latency matter.
Finally, instrument and disclose. Track energy or cost per call, prefer providers that publish efficiency scores, and build the habit of treating compute like any other consumable input rather than a free resource.
- Bigger is not better by default; model size should be justified by task complexity, not assumed.
- Task-specific models routinely match or beat general LLMs on narrow workloads while using a fraction of the energy.
- Every AI call has a measurable energy and carbon cost, and that cost should be visible to the person making the choice.
- Local execution on small models delivers privacy, sovereignty, and latency benefits that cloud LLMs cannot match.
- Diversity of model approaches (vision, time-series, acoustic, physics-based) often outperforms a single LLM forced into every domain.
Developed by Sasha Luccioni, AI researcher at Hugging Face, who led the AI Energy Score project benchmarking 100+ open-source models and published peer-reviewed studies comparing energy use across LLM and task-specific architectures.