AI Model Radar publishes a methodology for calculating whether 46 tracked open-weights models fit on specific hardware, rejecting common folklore rules that can be off by factors of four or forty. The site determines fit by summing measured GGUF weight sizes, context cache arithmetic from config files, and runtime overhead, while accounting for hybrid attention mechanisms like sliding windows and multi-head latent attention.

  • Weights are taken from published byte sizes rather than parameter counts to avoid shard-splitting errors.
  • Context cache is computed using layer types (full vs. sliding-window) and head dimensions explicitly stated in config.json.
  • A 12% safety margin is kept for standard GPUs, while Apple Silicon uses 70% of unified memory due to macOS constraints.
  • For an RTX 3060 12 GB, the analysis shows 14 models fit at 8K context and 10 at 32K context; Qwen3 14B requires offloading even at 8K.

The authors refuse to publish speed or quality scores, focusing solely on reproducible memory fit data to help users determine hardware compatibility without guessing.