Calculators
Browser-based calculators for sizing machine learning models. Every number is computed from a primary source, and nothing you enter leaves your machine.
Price your own token shape per request, per day and per month with the cache hit rate included, find the break even between any two models, and see which published prices disagree. 73 first party models, every price read from two independent records.
Rank, alpha, dropout and target modules with values from the QLoRA paper, then the exact trainable parameter count and the training VRAM it implies for 19 open models, with a fit verdict on a real card.
Weights, KV cache, gradients, optimizer states and activations for 19 open models, with exact fetched parameter counts, inference and training modes, and a custom-model path.
Embedding and vector-database cost across 7 embedding models and 3 vector stores: chunk-count math with overlap, storage, reads and writes per month, and the cheaper model-store pairing at your traffic.
Paste scores and labels for the rank-based Mann Whitney U AUC with a bootstrap confidence interval, a tie count, and the precision-recall AUC on the same rows.
Per-token KV cache for 19 open models, read from each model's config.json, with a custom-model path and a naive over-report column.