🧮 MoE Parameter Calculator

Total / active-expert / regular parameter counts for HuggingFace MoE safetensors models — computed from metadata only (config.json + safetensors index). No model weights are ever downloaded.

Read safetensors headers (tensor names + shapes, a few MB of metadata) for exact counts — never any weight data.

🔒 Metadata only: config.json, safetensors index, and (optionally) small safetensors headers — typically a few MB total. Model weights are never downloaded.

Works with any browser (direct fetch, no server needed). Definitions: Total = all parameters in the safetensors files; Active experts (per token) = parameters of the top-k routed experts actually computed for one token across all MoE layers (shared experts, if any, are always active and counted in Regular); Regular = everything that is not a routed expert (embeddings, attention, routers, shared experts, dense layers, norms, output head). Totals are cross-checked against HuggingFace's own safetensors.total.