Intelligence Density

Intelligence density expresses how much useful capability a model delivers for each gigabyte of memory it occupies. It divides a benchmark-based capability score by the model's deployment footprint, letting practitioners compare models by how much intelligence fits into constrained devices rather than by raw size alone.

The numerator combines normalized results from representative benchmarks, while the denominator measures the memory needed to load and run the model. Models with different architectures, training methods, or numerical precision can therefore be compared using the same ratio. Quantization, pruning, and distillation may improve the result, but intelligence density measures their outcome rather than prescribing a compression technique.

The metric can help device makers choose models for phones, laptops, vehicles, and other memory-constrained systems. However, a single ratio hides latency, energy use, runtime memory, and differences between benchmark suites. It can also reward models optimized for narrow tests unless evaluators disclose the tasks, weighting, and memory assumptions behind the score.

Open questions include which benchmarks should define useful capability, whether the denominator should include caches and temporary activations, and how scores should compare across text, vision, audio, and multimodal models. It is also unclear whether intelligence density will become a standardized reporting metric or remain an informal shorthand for model efficiency.